commit d0944da0109accbabc2c396534dcb76f0a43a71e Author: nyh <175484793@qq.com> Date: Mon Dec 2 15:11:24 2024 +0800 v1.0 diff --git a/README b/README new file mode 100644 index 0000000..0dba805 --- /dev/null +++ b/README @@ -0,0 +1,35 @@ +STDC语义分割模型 +1.通过data里面的json文件，修改任务, 配置文件如下： + "dspth":"../../data/RoadLane/", #数据文件夹 + "cropsize":"1280,720", #模型的宽、高 + "labelJson":"./data/RoadLane_info.json",#标签的信息 + "n_classes":3,#语义分割的类别叔叔 + "ignore_idx":255 #忽略的类别数 +2. 数据组织 + ├── train + │   ├── images [] + │   ├── labels [] + │   └── t.txt + └── val + ├── images [] + └── labels [] + 图像放在images下面，标签放在labels下面。 +3. 标签格式 及数据说明文件 + labels里的数据都是png格式，里面放的是RGB彩色标签。如：道路-（256,0,0）表示 + 数据说明文件：./data/RoadLane_info.json， 一般放在./data 文件下面 + { + "hasInstances": false, + "category": "void", + "catid": 0, + "name": "speedRoad", + "ignoreInEval": true, + "id":1, + "color": [ + 128, + 0, + 0 + ], + 主要是"id"和"color"要对应上，"id"是从0开始编号 +4. 模型训练 + python train.py --parJson ./data/RoadLane.json --respath ./checkpooints/0430pm --gpuId 0 + # ./checkpooints/0430pm --为之前保存的训练路径 diff --git a/__pycache__/cityscapes.cpython-37.pyc b/__pycache__/cityscapes.cpython-37.pyc new file mode 100644 index 0000000..e5aa247 Binary files /dev/null and b/__pycache__/cityscapes.cpython-37.pyc differ diff --git a/__pycache__/cityscapes.cpython-38.pyc b/__pycache__/cityscapes.cpython-38.pyc new file mode 100644 index 0000000..05a3aec Binary files /dev/null and b/__pycache__/cityscapes.cpython-38.pyc differ diff --git a/__pycache__/complexIllegalParkingUtilsNewest.cpython-38.pyc b/__pycache__/complexIllegalParkingUtilsNewest.cpython-38.pyc new file mode 100644 index 0000000..2a2ecd0 Binary files /dev/null and b/__pycache__/complexIllegalParkingUtilsNewest.cpython-38.pyc differ diff --git a/__pycache__/evaluation.cpython-37.pyc b/__pycache__/evaluation.cpython-37.pyc new file mode 100644 index 0000000..e84fa80 Binary files /dev/null and b/__pycache__/evaluation.cpython-37.pyc differ diff --git a/__pycache__/evaluation.cpython-38.pyc b/__pycache__/evaluation.cpython-38.pyc new file mode 100644 index 0000000..10595d3 Binary files /dev/null and b/__pycache__/evaluation.cpython-38.pyc differ diff --git a/__pycache__/evaluation_process.cpython-37.pyc b/__pycache__/evaluation_process.cpython-37.pyc new file mode 100644 index 0000000..3f48e19 Binary files /dev/null and b/__pycache__/evaluation_process.cpython-37.pyc differ diff --git a/__pycache__/evaluation_process.cpython-38.pyc b/__pycache__/evaluation_process.cpython-38.pyc new file mode 100644 index 0000000..ba3faa7 Binary files /dev/null and b/__pycache__/evaluation_process.cpython-38.pyc differ diff --git a/__pycache__/heliushuju.cpython-37.pyc b/__pycache__/heliushuju.cpython-37.pyc new file mode 100644 index 0000000..48361dd Binary files /dev/null and b/__pycache__/heliushuju.cpython-37.pyc differ diff --git a/__pycache__/heliushuju.cpython-38.pyc b/__pycache__/heliushuju.cpython-38.pyc new file mode 100644 index 0000000..06abb8d Binary files /dev/null and b/__pycache__/heliushuju.cpython-38.pyc differ diff --git a/__pycache__/heliushuju_process.cpython-37.pyc b/__pycache__/heliushuju_process.cpython-37.pyc new file mode 100644 index 0000000..7588f86 Binary files /dev/null and b/__pycache__/heliushuju_process.cpython-37.pyc differ diff --git a/__pycache__/heliushuju_process.cpython-38.pyc b/__pycache__/heliushuju_process.cpython-38.pyc new file mode 100644 index 0000000..0f3cac6 Binary files /dev/null and b/__pycache__/heliushuju_process.cpython-38.pyc differ diff --git a/__pycache__/logger.cpython-37.pyc b/__pycache__/logger.cpython-37.pyc new file mode 100644 index 0000000..9b6c0ac Binary files /dev/null and b/__pycache__/logger.cpython-37.pyc differ diff --git a/__pycache__/logger.cpython-38.pyc b/__pycache__/logger.cpython-38.pyc new file mode 100644 index 0000000..98fe4cf Binary files /dev/null and b/__pycache__/logger.cpython-38.pyc differ diff --git a/__pycache__/logger.cpython-39.pyc b/__pycache__/logger.cpython-39.pyc new file mode 100644 index 0000000..ae97d78 Binary files /dev/null and b/__pycache__/logger.cpython-39.pyc differ diff --git a/__pycache__/optimizer_loss.cpython-37.pyc b/__pycache__/optimizer_loss.cpython-37.pyc new file mode 100644 index 0000000..4f71662 Binary files /dev/null and b/__pycache__/optimizer_loss.cpython-37.pyc differ diff --git a/__pycache__/optimizer_loss.cpython-38.pyc b/__pycache__/optimizer_loss.cpython-38.pyc new file mode 100644 index 0000000..9e8cf36 Binary files /dev/null and b/__pycache__/optimizer_loss.cpython-38.pyc differ diff --git a/__pycache__/trafficDetectionUtils.cpython-38.pyc b/__pycache__/trafficDetectionUtils.cpython-38.pyc new file mode 100644 index 0000000..5906f47 Binary files /dev/null and b/__pycache__/trafficDetectionUtils.cpython-38.pyc differ diff --git a/__pycache__/transform.cpython-37.pyc b/__pycache__/transform.cpython-37.pyc new file mode 100644 index 0000000..02fb02b Binary files /dev/null and b/__pycache__/transform.cpython-37.pyc differ diff --git a/__pycache__/transform.cpython-38.pyc b/__pycache__/transform.cpython-38.pyc new file mode 100644 index 0000000..8a029d9 Binary files /dev/null and b/__pycache__/transform.cpython-38.pyc differ diff --git a/checkpoints2/STDCNet1446_76.47.tar b/checkpoints2/STDCNet1446_76.47.tar new file mode 100644 index 0000000..3e0cd45 Binary files /dev/null and b/checkpoints2/STDCNet1446_76.47.tar differ diff --git a/checkpoints2/STDCNet813M_73.91.tar b/checkpoints2/STDCNet813M_73.91.tar new file mode 100644 index 0000000..41e250f Binary files /dev/null and b/checkpoints2/STDCNet813M_73.91.tar differ diff --git a/checkpooints/0430pm/BiSeNet-2024-12-02-14-39-26.log b/checkpooints/0430pm/BiSeNet-2024-12-02-14-39-26.log new file mode 100644 index 0000000..e69de29 diff --git a/cityscapes.py b/cityscapes.py new file mode 100644 index 0000000..1814242 --- /dev/null +++ b/cityscapes.py @@ -0,0 +1,123 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- + + +import torch +from torch.utils.data import Dataset +import torchvision.transforms as transforms + +import os.path as osp +import os +from PIL import Image +import numpy as np +import json + +from transform import * + + + +class CityScapes(Dataset): + def __init__(self, rootpth, cropsize=(640, 480), mode='train', + randomscale=(0.125, 0.25, 0.375, 0.5, 0.675, 0.75, 0.875, 1.0, 1.25, 1.5), *args, **kwargs): + super(CityScapes, self).__init__(*args, **kwargs) + assert mode in ('train', 'val', 'test', 'trainval') + self.mode = mode + print('self.mode', self.mode) + self.ignore_lb = 255 + + with open('./cityscapes_info.json', 'r') as fr: + labels_info = json.load(fr) + self.lb_map = {el['id']: el['trainId'] for el in labels_info} + + + ## parse img directory + self.imgs = {} + imgnames = [] + impth = osp.join(rootpth, 'leftImg8bit', mode) + folders = os.listdir(impth) + for fd in folders: + fdpth = osp.join(impth, fd) + im_names = os.listdir(fdpth) + names = [el.replace('_leftImg8bit.png', '') for el in im_names] + impths = [osp.join(fdpth, el) for el in im_names] + imgnames.extend(names) + self.imgs.update(dict(zip(names, impths))) + + ## parse gt directory + self.labels = {} + gtnames = [] + gtpth = osp.join(rootpth, 'gtFine', mode) + folders = os.listdir(gtpth) + for fd in folders: + fdpth = osp.join(gtpth, fd) + lbnames = os.listdir(fdpth) + lbnames = [el for el in lbnames if 'labelIds' in el] + names = [el.replace('_gtFine_labelIds.png', '') for el in lbnames] + lbpths = [osp.join(fdpth, el) for el in lbnames] + gtnames.extend(names) + self.labels.update(dict(zip(names, lbpths))) + + self.imnames = imgnames + self.len = len(self.imnames) + print('self.len', self.mode, self.len) + assert set(imgnames) == set(gtnames) + assert set(self.imnames) == set(self.imgs.keys()) + assert set(self.imnames) == set(self.labels.keys()) + + ## pre-processing + self.to_tensor = transforms.Compose([ + transforms.ToTensor(), + transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), + ]) + self.trans_train = Compose([ + ColorJitter( + brightness = 0.5, + contrast = 0.5, + saturation = 0.5), + HorizontalFlip(), + # RandomScale((0.25, 0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0)), + RandomScale(randomscale), + # RandomScale((0.125, 1)), + # RandomScale((0.125, 0.25, 0.375, 0.5, 0.675, 0.75, 0.875, 1.0)), + # RandomScale((0.125, 0.25, 0.375, 0.5, 0.675, 0.75, 0.875, 1.0, 1.125, 1.25, 1.375, 1.5)), + RandomCrop(cropsize) + ]) + + + def __getitem__(self, idx): + fn = self.imnames[idx] + impth = self.imgs[fn] + lbpth = self.labels[fn] + img = Image.open(impth).convert('RGB') + label = Image.open(lbpth) + if self.mode == 'train' or self.mode == 'trainval': + im_lb = dict(im = img, lb = label) + im_lb = self.trans_train(im_lb) + img, label = im_lb['im'], im_lb['lb'] + img = self.to_tensor(img) + label = np.array(label).astype(np.int64)[np.newaxis, :] + label = self.convert_labels(label) + return img, label + + + def __len__(self): + return self.len + + + def convert_labels(self, label): + for k, v in self.lb_map.items(): + label[label == k] = v + return label + + + +if __name__ == "__main__": + from tqdm import tqdm + ds = CityScapes('./data/', n_classes=19, mode='val') + uni = [] + for im, lb in tqdm(ds): + lb_uni = np.unique(lb).tolist() + uni.extend(lb_uni) + print(uni) + print(set(uni)) + diff --git a/data/RoadLane.json b/data/RoadLane.json new file mode 100644 index 0000000..3482297 --- /dev/null +++ b/data/RoadLane.json @@ -0,0 +1,8 @@ +{ + "dspth":"../../data/RoadLane/", + "cropsize":"1280,720", + "labelJson":"./data/RoadLane_info.json", + "n_classes":3, + "ignore_idx":255 + +} diff --git a/data/RoadLane_info.json b/data/RoadLane_info.json new file mode 100644 index 0000000..6bec499 --- /dev/null +++ b/data/RoadLane_info.json @@ -0,0 +1,44 @@ +[ + { + "hasInstances": false, + "category": "void", + "catid": 0, + "name": "black", + "ignoreInEval": true, + "id":0, + "color": [ + 0, + 0, + 0 + ], + "trainId": 0 + }, + { + "hasInstances": false, + "category": "void", + "catid": 0, + "name": "speedRoad", + "ignoreInEval": true, + "id":1, + "color": [ + 128, + 0, + 0 + ], + "trainId": 1 + }, + { + "hasInstances": false, + "category": "void", + "catid": 0, + "name": "lane", + "ignoreInEval": true, + "id":2, + "color": [ + 128, + 128, + 0 + ], + "trainId": 3 + } +] diff --git a/data/carRoadLane.json b/data/carRoadLane.json new file mode 100644 index 0000000..0dc0999 --- /dev/null +++ b/data/carRoadLane.json @@ -0,0 +1,8 @@ +{ + "dspth":"../../data/CarRoadLane/", + "cropsize":"1280,720", + "labelJson":"./data/heliushuju_info.json", + "n_classes":4, + "ignore_idx":255 + +} diff --git a/data/heliushuju_info.json b/data/heliushuju_info.json new file mode 100644 index 0000000..4454806 --- /dev/null +++ b/data/heliushuju_info.json @@ -0,0 +1,58 @@ +[ + { + "hasInstances": false, + "category": "void", + "catid": 0, + "name": "black", + "ignoreInEval": true, + "id": 0, + "color": [ + 0, + 0, + 0 + ], + "trainId": 0 + }, + { + "hasInstances": false, + "category": "void", + "catid": 0, + "name": "speedRoad", + "ignoreInEval": true, + "id": 1, + "color": [ + 128, + 0, + 0 + ], + "trainId": 1 + }, + { + "hasInstances": false, + "category": "void", + "catid": 0, + "name": "vehicle", + "ignoreInEval": true, + "id": 2, + "color": [ + 0, + 128, + 0 + ], + "trainId": 2 + }, + { + "hasInstances": false, + "category": "void", + "catid": 0, + "name": "lane", + "ignoreInEval": true, + "id": 3, + "color": [ + 128, + 128, + 0 + ], + "trainId": 3 + } +] diff --git a/evaluation_process.py b/evaluation_process.py new file mode 100644 index 0000000..d928329 --- /dev/null +++ b/evaluation_process.py @@ -0,0 +1,324 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- + +from logger import setup_logger +from models.model_stages import BiSeNet +from cityscapes import CityScapes + +import torch +import torch.nn as nn +from torch.utils.data import DataLoader +import torch.nn.functional as F +import torch.distributed as dist + +import os +import os.path as osp +import logging +import time +import numpy as np +from tqdm import tqdm +import math +from PIL import Image +from heliushuju_process import Heliushuju +import json +from utils.metrics import Evaluator +class MscEvalV0(object): + + def __init__(self, scale=0.5,ignore_label=255): + self.ignore_label = ignore_label + self.scale = scale + + def __call__(self, net, dl, n_classes): + # evaluate + hist = torch.zeros(n_classes, n_classes).cuda().detach() + self.evaluator = Evaluator(n_classes)#创建实例化对象 + self.evaluator.reset() + + + if dist.is_initialized() and dist.get_rank() != 0: + diter = enumerate(dl) + else: + diter = enumerate(tqdm(dl)) + for i, (imgs, label) in diter: + N, _, H, W = label.shape # 原始 + label = label.squeeze(1).cuda() # 原始 + size = label.size()[-2:] + imgs = imgs.cuda() + N, C, H, W = imgs.size() + new_hw = [int(H*self.scale), int(W*self.scale)] + imgs = F.interpolate(imgs, new_hw, mode='bilinear', align_corners=True) + logits = net(imgs)[0] + logits = F.interpolate(logits, size=size, mode='bilinear', align_corners=True) + probs = torch.softmax(logits, dim=1) + preds = torch.argmax(probs, dim=1) + keep = label != self.ignore_label + #print( torch.max( label[keep]), torch.min( label[keep]), torch.max( preds[keep]), torch.min( preds[keep]), ) + hist += torch.bincount(label[keep] * n_classes + preds[keep], minlength=n_classes ** 2).view(n_classes, n_classes).float() # 原始 + + self.evaluator.add_batch(label.cpu().numpy(), preds.cpu().numpy())#更新混淆矩阵 + Acc = self.evaluator.Pixel_Accuracy() + Acc_class = self.evaluator.Pixel_Accuracy_Class() + class_IoU,mIoU= self.evaluator.Mean_Intersection_over_Union() + FWIoU = self.evaluator.Frequency_Weighted_Intersection_over_Union() + recall,precision,f1=self.evaluator.Recall_Precision() + + + print("val Acc:{}, Acc_class:{}, mIoU:{}, fwIoU: {}".format(Acc, Acc_class, mIoU, FWIoU)) + for i,iou in enumerate(class_IoU): + print(' class:%d ,Iou:%.4f '%(i,iou),end='') + print() + + + + + if dist.is_initialized(): + dist.all_reduce(hist, dist.ReduceOp.SUM) + ious = hist.diag() / (hist.sum(dim=0) + hist.sum(dim=1) - hist.diag()) + miou = ious.mean() + return miou.item() + + +def evaluatev0(respth='./pretrained', dspth='./data', backbone='CatNetSmall', scale=0.75, use_boundary_2=False, use_boundary_4=False, use_boundary_8=False, use_boundary_16=False, use_conv_last=False,n_classes=4,modelSize=(640,360),mode='test',outpath='outputs/test/',labelJson='data/heliushuju_info.json'): + print('scale', scale) + print('use_boundary_2', use_boundary_2) + print('use_boundary_4', use_boundary_4) + print('use_boundary_8', use_boundary_8) + print('use_boundary_16', use_boundary_16) + ## dataset + batchsize = 5 + + n_workers = 2 + #dsval = CityScapes(dspth, mode='val') + + dsval = Heliushuju(dspth, mode=mode,cropsize=modelSize,labelJson=labelJson) + with open(labelJson,'r') as fr: + labels_info = json.load(fr) + + lb_map = {el['id']: el['color'] for el in labels_info} + + #print('---line89 lb_map:',lb_map, ' labels_info:',labels_info) + lb_colors = np.array( [lb_map[k] for k in lb_map.keys()]) + + + + + + dl = DataLoader(dsval, + batch_size = batchsize, + shuffle = False, + num_workers = n_workers, + drop_last = False) + + + print("backbone:", backbone) + net = BiSeNet(backbone=backbone, n_classes=n_classes, + use_boundary_2=use_boundary_2, use_boundary_4=use_boundary_4, + use_boundary_8=use_boundary_8, use_boundary_16=use_boundary_16, + use_conv_last=use_conv_last) + net.load_state_dict(torch.load(respth)) + net.cuda() + net.eval() + + if mode=='val': + with torch.no_grad(): + single_scale = MscEvalV0(scale=scale,ignore_label=255) + mIOU = single_scale(net, dl, n_classes) + logger = logging.getLogger() + logger.info('mIOU is: %s\n', mIOU) + else: + diter = enumerate(tqdm(dl)) + with torch.no_grad(): + for i, (imgs, filenames) in diter: + N, _, H, W = imgs.shape # 原始 + + imgs = imgs.cuda() + N, C, H, W = imgs.size() + new_hw = [int(H*scale), int(W*scale)] + imgs = F.interpolate(imgs, new_hw, mode='bilinear', align_corners=True) + logits = net(imgs)[0] + logits = F.interpolate(logits, size=(H,W), mode='bilinear', align_corners=True) + probs = torch.softmax(logits, dim=1) + preds = torch.argmax(probs, dim=1).cpu().numpy() + print(preds.shape,logits.shape) + for jj, ff in enumerate(filenames): + pred = preds[jj] + pred_color = lb_colors[ pred] + #print(jj,pred.shape,pred_color.shape ,type(pred_color ),lb_colors ) + t1=Image.fromarray(np.uint8(pred_color)) + t1.save(os.path.join(outpath,ff+'.png') ) + #cv2.imwrite( os.path.join(outpath,ff+'.png'), imwrite.astype(np.uint8) ) + + + + +class MscEval(object): + def __init__(self, + model, + dataloader, + scales = [0.5, 0.75, 1, 1.25, 1.5, 1.75], + n_classes = 19, + lb_ignore = 255, + cropsize = 1024, + flip = True, + *args, **kwargs): + self.scales = scales + self.n_classes = n_classes + self.lb_ignore = lb_ignore + self.flip = flip + self.cropsize = cropsize + ## dataloader + self.dl = dataloader + self.net = model + + + def pad_tensor(self, inten, size): + N, C, H, W = inten.size() + outten = torch.zeros(N, C, size[0], size[1]).cuda() + outten.requires_grad = False + margin_h, margin_w = size[0]-H, size[1]-W + hst, hed = margin_h//2, margin_h//2+H + wst, wed = margin_w//2, margin_w//2+W + outten[:, :, hst:hed, wst:wed] = inten + return outten, [hst, hed, wst, wed] + + + def eval_chip(self, crop): + with torch.no_grad(): + out = self.net(crop)[0] + prob = F.softmax(out, 1) + if self.flip: + crop = torch.flip(crop, dims=(3,)) + out = self.net(crop)[0] + out = torch.flip(out, dims=(3,)) + prob += F.softmax(out, 1) + prob = torch.exp(prob) + return prob + + + def crop_eval(self, im): + cropsize = self.cropsize + stride_rate = 5/6. + N, C, H, W = im.size() + long_size, short_size = (H,W) if H>W else (W,H) + if long_size < cropsize: + im, indices = self.pad_tensor(im, (cropsize, cropsize)) + prob = self.eval_chip(im) + prob = prob[:, :, indices[0]:indices[1], indices[2]:indices[3]] + else: + stride = math.ceil(cropsize*stride_rate) + if short_size < cropsize: + if H < W: + im, indices = self.pad_tensor(im, (cropsize, W)) + else: + im, indices = self.pad_tensor(im, (H, cropsize)) + N, C, H, W = im.size() + n_x = math.ceil((W-cropsize)/stride)+1 + n_y = math.ceil((H-cropsize)/stride)+1 + prob = torch.zeros(N, self.n_classes, H, W).cuda() + prob.requires_grad = False + for iy in range(n_y): + for ix in range(n_x): + hed, wed = min(H, stride*iy+cropsize), min(W, stride*ix+cropsize) + hst, wst = hed-cropsize, wed-cropsize + chip = im[:, :, hst:hed, wst:wed] + prob_chip = self.eval_chip(chip) + prob[:, :, hst:hed, wst:wed] += prob_chip + if short_size < cropsize: + prob = prob[:, :, indices[0]:indices[1], indices[2]:indices[3]] + return prob + + + def scale_crop_eval(self, im, scale): + N, C, H, W = im.size() + new_hw = [int(H*scale), int(W*scale)] + im = F.interpolate(im, new_hw, mode='bilinear', align_corners=True) + prob = self.crop_eval(im) + prob = F.interpolate(prob, (H, W), mode='bilinear', align_corners=True) + return prob + + + def compute_hist(self, pred, lb): + n_classes = self.n_classes + ignore_idx = self.lb_ignore + keep = np.logical_not(lb==ignore_idx) + merge = pred[keep] * n_classes + lb[keep] + hist = np.bincount(merge, minlength=n_classes**2) + hist = hist.reshape((n_classes, n_classes)) + return hist + + + def evaluate(self): + ## evaluate + n_classes = self.n_classes + hist = np.zeros((n_classes, n_classes), dtype=np.float32) + dloader = tqdm(self.dl) + if dist.is_initialized() and not dist.get_rank()==0: + dloader = self.dl + for i, (imgs, label) in enumerate(dloader): + N, _, H, W = label.shape + probs = torch.zeros((N, self.n_classes, H, W)) + probs.requires_grad = False + imgs = imgs.cuda() + for sc in self.scales: + # prob = self.scale_crop_eval(imgs, sc) + prob = self.eval_chip(imgs) + probs += prob.detach().cpu() + probs = probs.data.numpy() + preds = np.argmax(probs, axis=1) + + hist_once = self.compute_hist(preds, label.data.numpy().squeeze(1)) + hist = hist + hist_once + IOUs = np.diag(hist) / (np.sum(hist, axis=0)+np.sum(hist, axis=1)-np.diag(hist)) + mIOU = np.mean(IOUs) + return mIOU + + +def evaluate(respth='./resv1_catnet/pths/', dspth='./data'): + ## logger + logger = logging.getLogger() + + ## model + logger.info('\n') + logger.info('===='*20) + logger.info('evaluating the model ...\n') + logger.info('setup and restore model') + n_classes = 19 + net = BiSeNet(n_classes=n_classes) + + net.load_state_dict(torch.load(respth)) + net.cuda() + net.eval() + + ## dataset + batchsize = 5 + n_workers = 2 + dsval = CityScapes(dspth, mode='val') + dl = DataLoader(dsval, + batch_size = batchsize, + shuffle = False, + num_workers = n_workers, + drop_last = False) + + ## evaluator + logger.info('compute the mIOU') + evaluator = MscEval(net, dl, scales=[1], flip = False) + + ## eval + mIOU = evaluator.evaluate() + logger.info('mIOU is: {:.6f}'.format(mIOU)) + + + +if __name__ == "__main__": + log_dir = 'evaluation_logs/' + if not os.path.exists(log_dir): + os.makedirs(log_dir) + setup_logger(log_dir) + + #modelpath='./checkpooints/0430/pths/model_final.pth';n_classes=4:labelJson='data/heliushuju_info.json'i;dspth='../../data/carRoadLane/';mode='val' + modelpath='./checkpooints/0430pm/pths/model_final.pth';labelJson='data/RoadLane_info.json';n_classes=3;dspth='../../data/RoadLane/';mode='val' + evaluatev0(modelpath, + dspth=dspth, backbone='STDCNet813', scale=1.0, + use_boundary_2=False, use_boundary_4=False, use_boundary_8=True, use_boundary_16=False,n_classes=n_classes,modelSize=(1920,1080),mode=mode,outpath='outputs/test2/',labelJson=labelJson) + + diff --git a/heliushuju_process.py b/heliushuju_process.py new file mode 100644 index 0000000..d81f5dd --- /dev/null +++ b/heliushuju_process.py @@ -0,0 +1,295 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- + +import torch +from matplotlib import pyplot as plt +from torch.utils.data import Dataset +import torchvision.transforms as transforms + +import os.path as osp +import os +from PIL import Image +import numpy as np +import json +import cv2 +import time +from transform import * + + +class Heliushuju(Dataset): + def __init__(self, rootpth, cropsize=(640, 480), mode='train',labelJson='./heliushuju_info.json', + randomscale=(0.125, 0.25, 0.375, 0.5, 0.675, 0.75, 0.875, 1.0, 1.25, 1.5), *args, **kwargs): + super(Heliushuju, self).__init__(*args, **kwargs) + assert mode in ('train', 'val', 'test', 'trainval') + self.mode = mode + self.modeSize=cropsize + + self.ignore_lb = 255 + + #with open('./heliushuju_info.json', 'r') as fr: + with open(labelJson,'r') as fr: + print('labelJson:',labelJson) + labels_info = json.load(fr) + + self.lb_map = {el['id']: el['color'] for el in labels_info} + + self.imgs = {} + imgnames = [] + impth = osp.join(rootpth, mode, 'images') # 图片所在目录的路径 + folders = os.listdir(impth) # 图片名列表 + names = [el.replace(el[-4:], '') for el in folders] # el是整个图片名,names是图片名前缀 + impths = [osp.join(impth, el) for el in folders] # 图片路径 + imgnames.extend(names) # 存放图片名前缀的列表 + self.imgs.update(dict(zip(names, impths))) + + + if self.mode !='test': + + self.labels = {} + gtnames = [] + gtpth = osp.join(rootpth, mode, 'labels') + folders = os.listdir(gtpth) + names = [el.replace(el[-4:], '') for el in folders] + lbpths = [osp.join(gtpth, el) for el in folders] + gtnames.extend(names) + self.labels.update(dict(zip(names, lbpths))) + + self.imnames = imgnames + self.len = len(self.imnames) + print('self.len', self.mode, self.len) + if self.mode !='test': + assert set(imgnames) == set(gtnames) + assert set(self.imnames) == set(self.imgs.keys()) + assert set(self.imnames) == set(self.labels.keys()) + + # pre-processing + self.to_tensor = transforms.Compose([ + transforms.ToTensor(), + transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), + ]) + self.trans_train = Compose([ + ColorJitter( + brightness = 0.5, + contrast = 0.5, + saturation = 0.5), + HorizontalFlip(), + RandomScale(randomscale), + RandomCrop(cropsize) + ]) + self.mean = (0.485, 0.456, 0.406) + self.std = (0.229, 0.224, 0.225) + + def __getitem__(self, idx): + fn = self.imnames[idx] + impth = self.imgs[fn] + img = Image.open(impth).convert('RGB') + + if self.mode !='test': + lbpth = self.labels[fn] + label = cv2.imread(lbpth) # 原始 + label = cv2.cvtColor(label, cv2.COLOR_BGR2RGB) # 添加（训练交通事故数据，添加了这行代码使标签颜色正确） + + + if self.mode == 'train' or self.mode == 'trainval' or self.mode == 'val': + label = Image.fromarray(label) + im_lb = dict(im = img, lb = label) + im_lb = self.trans_train(im_lb) + img, label = im_lb['im'], im_lb['lb'] + + + img = np.array(img); + img = self.preprocess_image(img) + if self.mode !='test': + label = cv2.resize(np.array(label), self.modeSize) + label = label.astype(np.int64)[np.newaxis, :] # 给行上增加维度 + label = self.convert_labels(label) + return img, label.astype(np.int64) + else: + return img,fn + + def __len__(self): + return self.len + + def convert_labels(self, label): + b, h, w, c = label.shape + + label_index = np.zeros((b, h, w)) + for k, v in self.lb_map.items(): + t_0 = (label[..., 0] == v[0]) + t_1 = (label[..., 1] == v[1]) + t_2 = (label[..., 2] == v[2]) + t_loc = (t_0 & t_1 & t_2) + label_index[t_loc] = k + + # label[label == k] = v + # print(label) + # print("6666666666666666") + return label_index + + def preprocess_image(self, image): + time0 = time.time() + + image = cv2.resize(image, self.modeSize) + + time1 = time.time() + image = image.astype(np.float32) + image /= 255.0 + + time2 = time.time() + # image = image * 3.2 - 1.6 + image[:, :, 0] -= self.mean[0] + image[:, :, 1] -= self.mean[1] + image[:, :, 2] -= self.mean[2] + + time3 = time.time() + image[:, :, 0] /= self.std[0] + image[:, :, 1] /= self.std[1] + image[:, :, 2] /= self.std[2] + + time4 = time.time() + image = np.transpose(image, (2, 0, 1)) + time5 = time.time() + image = torch.from_numpy(image).float() + + + return image + +class Heliushuju_test(Dataset): + def __init__(self, rootpth, cropsize=(640, 480), mode='test',labelJson='./heliushuju_info.json', + randomscale=(0.125, 0.25, 0.375, 0.5, 0.675, 0.75, 0.875, 1.0, 1.25, 1.5), *args, **kwargs): + super(Heliushuju_test, self).__init__(*args, **kwargs) + assert mode in ('train', 'val', 'test', 'trainval') + self.mode = mode + self.modeSize=cropsize + + + + #with open('./heliushuju_info.json', 'r') as fr: + with open(labelJson,'r') as fr: + labels_info = json.load(fr) + + self.lb_map = {el['id']: el['color'] for el in labels_info} + + self.imgs = {} + imgnames = [] + impth = osp.join(rootpth, mode, 'images') # 图片所在目录的路径 + folders = os.listdir(impth) # 图片名列表 + names = [el.replace(el[-4:], '') for el in folders] # el是整个图片名,names是图片名前缀 + impths = [osp.join(impth, el) for el in folders] # 图片路径 + imgnames.extend(names) # 存放图片名前缀的列表 + self.imgs.update(dict(zip(names, impths))) + + + + self.imnames = imgnames + self.len = len(self.imnames) + print('self.len', self.mode, self.len) + assert set(imgnames) == set(gtnames) + assert set(self.imnames) == set(self.imgs.keys()) + assert set(self.imnames) == set(self.labels.keys()) + + # pre-processing + self.to_tensor = transforms.Compose([ + transforms.ToTensor(), + transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), + ]) + self.trans_train = Compose([ + ColorJitter( + brightness = 0.5, + contrast = 0.5, + saturation = 0.5), + HorizontalFlip(), + RandomScale(randomscale), + RandomCrop(cropsize) + ]) + self.mean = (0.485, 0.456, 0.406) + self.std = (0.229, 0.224, 0.225) + + def __getitem__(self, idx): + fn = self.imnames[idx] + impth = self.imgs[fn] + lbpth = self.labels[fn] + + img = Image.open(impth).convert('RGB') + + label = cv2.imread(lbpth) # 原始 + label = cv2.cvtColor(label, cv2.COLOR_BGR2RGB) # 添加（训练交通事故数据，添加了这行代码使标签颜色正确） + + + if self.mode == 'train' or self.mode == 'trainval' or self.mode == 'val': + label = Image.fromarray(label) + im_lb = dict(im = img, lb = label) + im_lb = self.trans_train(im_lb) + img, label = im_lb['im'], im_lb['lb'] + + + img = np.array(img); + img_bak = img.copy() + + img = self.preprocess_image(img) + label = cv2.resize(np.array(label), self.modeSize) + label = label.astype(np.int64)[np.newaxis, :] # 给行上增加维度 + label = self.convert_labels(label) + + return img, label.astype(np.int64) + + def __len__(self): + return self.len + + def convert_labels(self, label): + b, h, w, c = label.shape + + label_index = np.zeros((b, h, w)) + for k, v in self.lb_map.items(): + t_0 = (label[..., 0] == v[0]) + t_1 = (label[..., 1] == v[1]) + t_2 = (label[..., 2] == v[2]) + t_loc = (t_0 & t_1 & t_2) + label_index[t_loc] = k + + # label[label == k] = v + # print(label) + # print("6666666666666666") + return label_index + + def preprocess_image(self, image): + time0 = time.time() + + image = cv2.resize(image, self.modeSize) + + time1 = time.time() + image = image.astype(np.float32) + image /= 255.0 + + time2 = time.time() + # image = image * 3.2 - 1.6 + image[:, :, 0] -= self.mean[0] + image[:, :, 1] -= self.mean[1] + image[:, :, 2] -= self.mean[2] + + time3 = time.time() + image[:, :, 0] /= self.std[0] + image[:, :, 1] /= self.std[1] + image[:, :, 2] /= self.std[2] + + time4 = time.time() + image = np.transpose(image, (2, 0, 1)) + time5 = time.time() + image = torch.from_numpy(image).float() + + + return image + +if __name__ == "__main__": + from tqdm import tqdm + + # ds = Heliushuju('./data/', n_classes=2, mode='val') # 原始 + ds = Heliushuju('./data/', n_classes=3, mode='val') # 改动 + + uni = [] + for im, lb in tqdm(ds): + lb_uni = np.unique(lb).tolist() + uni.extend(lb_uni) + print(uni) + print(set(uni)) + diff --git a/latency/__init__.py b/latency/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/latency/__pycache__/__init__.cpython-37.pyc b/latency/__pycache__/__init__.cpython-37.pyc new file mode 100644 index 0000000..03ab3db Binary files /dev/null and b/latency/__pycache__/__init__.cpython-37.pyc differ diff --git a/latency/model.onnx b/latency/model.onnx new file mode 100644 index 0000000..c72dc89 Binary files /dev/null and b/latency/model.onnx differ diff --git a/latency/run_latency_stages.py b/latency/run_latency_stages.py new file mode 100644 index 0000000..9923725 --- /dev/null +++ b/latency/run_latency_stages.py @@ -0,0 +1,100 @@ +from __future__ import division + +import os +import sys +import logging +import torch +import numpy as np + +from thop import profile +sys.path.append("../") + +#from utils.darts_utils import create_exp_dir, plot_op, plot_path_width, objective_acc_lat +try: + from utils.darts_utils import compute_latency_ms_tensorrt as compute_latency + print("use TensorRT for latency test") +except: + from utils.darts_utils import compute_latency_ms_pytorch as compute_latency + print("use PyTorch for latency test") +from utils.darts_utils import compute_latency_ms_pytorch as compute_latency +print("use PyTorch for latency test") + +from models.model_stages_trt import BiSeNet + +def main(): + + print("begin") + # preparation ################ + torch.backends.cudnn.enabled = True + torch.backends.cudnn.benchmark = True + seed = 12345 + np.random.seed(seed) + torch.manual_seed(seed) + if torch.cuda.is_available(): + torch.cuda.manual_seed(seed) + + # Configuration ############## + use_boundary_2 = False + use_boundary_4 = False + use_boundary_8 = True + use_boundary_16 = False + use_conv_last = False + n_classes = 2 + + # STDC1Seg-50 250.4FPS on NVIDIA GTX 1080Ti + backbone = 'STDCNet813' + # methodName = 'STDC1-Seg' + methodName = 'wurenji_train_STDC1-Seg/pths' + inputSize = 512 + inputScale = 50 + inputDimension = (1, 3, 512, 1024) + + # # STDC1Seg-75 126.7FPS on NVIDIA GTX 1080Ti + # backbone = 'STDCNet813' + # methodName = 'STDC1-Seg' + # inputSize = 768 + # inputScale = 75 + # inputDimension = (1, 3, 768, 1536) + + # # STDC2Seg-50 188.6FPS on NVIDIA GTX 1080Ti + # backbone = 'STDCNet1446' + # methodName = 'STDC2-Seg' + # inputSize = 512 + # inputScale = 50 + # inputDimension = (1, 3, 512, 1024) + + # # STDC2Seg-75 97.0FPS on NVIDIA GTX 1080Ti + # backbone = 'STDCNet1446' + # methodName = 'STDC2-Seg' + # inputSize = 768 + # inputScale = 75 + # inputDimension = (1, 3, 768, 1536) + + model = BiSeNet(backbone=backbone, n_classes=n_classes, + use_boundary_2=use_boundary_2, use_boundary_4=use_boundary_4, + use_boundary_8=use_boundary_8, use_boundary_16=use_boundary_16, + input_size=inputSize, use_conv_last=use_conv_last) + + + print('loading parameters...') + respth = '../checkpoints/{}/'.format(methodName) + save_pth = os.path.join(respth, 'model_maxmIOU{}.pth'.format(inputScale)) + model.load_state_dict(torch.load(save_pth)) + model = model.cuda() + ##################################################### + + latency = compute_latency(model, inputDimension) + print("{}{} FPS:".format(methodName, inputScale) + str(1000./latency)) + logging.info("{}{} FPS:".format(methodName, inputScale) + str(1000./latency)) + + # calculate FLOPS and params + ''' + model = model.cpu() + flops, params = profile(model, inputs=(torch.randn(inputDimension),), verbose=False) + print("params = {}MB, FLOPs = {}GB".format(params / 1e6, flops / 1e9)) + logging.info("params = {}MB, FLOPs = {}GB".format(params / 1e6, flops / 1e9)) + ''' + + +if __name__ == '__main__': + main() diff --git a/latency/utils/__init__.py b/latency/utils/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/latency/utils/__pycache__/__init__.cpython-37.pyc b/latency/utils/__pycache__/__init__.cpython-37.pyc new file mode 100644 index 0000000..c472f57 Binary files /dev/null and b/latency/utils/__pycache__/__init__.cpython-37.pyc differ diff --git a/latency/utils/__pycache__/darts_utils.cpython-37.pyc b/latency/utils/__pycache__/darts_utils.cpython-37.pyc new file mode 100644 index 0000000..7cca14b Binary files /dev/null and b/latency/utils/__pycache__/darts_utils.cpython-37.pyc differ diff --git a/latency/utils/__pycache__/genotypes.cpython-37.pyc b/latency/utils/__pycache__/genotypes.cpython-37.pyc new file mode 100644 index 0000000..c6d7ee1 Binary files /dev/null and b/latency/utils/__pycache__/genotypes.cpython-37.pyc differ diff --git a/latency/utils/darts_utils.py b/latency/utils/darts_utils.py new file mode 100644 index 0000000..6a3ce27 --- /dev/null +++ b/latency/utils/darts_utils.py @@ -0,0 +1,353 @@ +import os +import math +import numpy as np +import torch +import shutil +from torch.autograd import Variable +import time +from tqdm import tqdm +from latency.utils.genotypes import PRIMITIVES +import matplotlib +matplotlib.use('Agg') +from matplotlib import pyplot as plt +from pdb import set_trace as bp +import warnings + + +class AvgrageMeter(object): + + def __init__(self): + self.reset() + + def reset(self): + self.avg = 0 + self.sum = 0 + self.cnt = 0 + + def update(self, val, n=1): + self.sum += val * n + self.cnt += n + self.avg = self.sum / self.cnt + + +class Cutout(object): + def __init__(self, length): + self.length = length + + def __call__(self, img): + h, w = img.size(1), img.size(2) + mask = np.ones((h, w), np.float32) + y = np.random.randint(h) + x = np.random.randint(w) + + y1 = np.clip(y - self.length // 2, 0, h) + y2 = np.clip(y + self.length // 2, 0, h) + x1 = np.clip(x - self.length // 2, 0, w) + x2 = np.clip(x + self.length // 2, 0, w) + + mask[y1: y2, x1: x2] = 0. + mask = torch.from_numpy(mask) + mask = mask.expand_as(img) + img *= mask + return img + + +def count_parameters_in_MB(model): + return np.sum(np.prod(v.size()) for name, v in model.named_parameters() if "auxiliary" not in name)/1e6 + + +def save_checkpoint(state, is_best, save): + filename = os.path.join(save, 'checkpoint.pth.tar') + torch.save(state, filename) + if is_best: + best_filename = os.path.join(save, 'model_best.pth.tar') + shutil.copyfile(filename, best_filename) + + +def save(model, model_path): + torch.save(model.state_dict(), model_path) + + +def load(model, model_path): + model.load_state_dict(torch.load(model_path)) + + +def drop_path(x, drop_prob): + if drop_prob > 0.: + keep_prob = 1.-drop_prob + mask = Variable(torch.cuda.FloatTensor(x.size(0), 1, 1, 1).bernoulli_(keep_prob)) + x.div_(keep_prob) + x.mul_(mask) + return x + + +def create_exp_dir(path, scripts_to_save=None): + if not os.path.exists(path): + os.mkdir(path) + print('Experiment dir : {}'.format(path)) + + if scripts_to_save is not None: + os.mkdir(os.path.join(path, 'scripts')) + for script in scripts_to_save: + dst_file = os.path.join(path, 'scripts', os.path.basename(script)) + shutil.copyfile(script, dst_file) + +########################## TensorRT speed_test ################################# +# try: +import tensorrt as trt +# import pycuda.driver as cuda +# import pycuda.autoinit + +MAX_BATCH_SIZE = 1 +MAX_WORKSPACE_SIZE = 1 << 30 +TRT_LOGGER = trt.Logger(trt.Logger.WARNING) +DTYPE = trt.float32 + +# Model +INPUT_NAME = 'input' +OUTPUT_NAME = 'output' + +def allocate_buffers(engine): + h_input = cuda.pagelocked_empty(trt.volume(engine.get_binding_shape(0))* engine.max_batch_size, dtype=trt.nptype(DTYPE)) + h_output = cuda.pagelocked_empty(trt.volume(engine.get_binding_shape(1))* engine.max_batch_size, dtype=trt.nptype(DTYPE)) + d_input = cuda.mem_alloc(h_input.nbytes) + d_output = cuda.mem_alloc(h_output.nbytes) + return h_input, d_input, h_output, d_output + + +def build_engine(model_file): + + with trt.Builder(TRT_LOGGER) as builder, builder.create_network() as network, trt.OnnxParser(network, TRT_LOGGER) as parser: + builder.max_workspace_size = MAX_WORKSPACE_SIZE + builder.max_batch_size = MAX_BATCH_SIZE + with open(model_file, 'rb') as model: + parser.parse(model.read()) + engine = builder.build_cuda_engine(network) + return engine + + + + +def load_input(input_size, host_buffer): + assert len(input_size) == 4 + b, c, h, w = input_size + dtype = trt.nptype(DTYPE) + img_array = np.random.randn(MAX_BATCH_SIZE, c, h, w).astype(dtype).ravel() + np.copyto(host_buffer, img_array) + +def do_inference(context, h_input, d_input, h_output, d_output, iterations=None): + # Transfer input data to the GPU. + cuda.memcpy_htod(d_input, h_input) + # warm-up + for _ in range(10): + context.execute(batch_size=MAX_BATCH_SIZE, bindings=[int(d_input), int(d_output)]) + # test proper iterations + if iterations is None: + elapsed_time = 0 + iterations = 100 + while elapsed_time < 1: + t_start = time.time() + for _ in range(iterations): + context.execute(batch_size=MAX_BATCH_SIZE, bindings=[int(d_input), int(d_output)]) + elapsed_time = time.time() - t_start + iterations *= 2 + FPS = iterations / elapsed_time + iterations = int(FPS * 3) + # Run inference. + t_start = time.time() + for _ in tqdm(range(iterations)): + context.execute(batch_size=MAX_BATCH_SIZE, bindings=[int(d_input), int(d_output)]) + elapsed_time = time.time() - t_start + latency = elapsed_time / iterations * 1000 + return latency + + +def compute_latency_ms_tensorrt(model, input_size, iterations=None): + # print('input_size: ', input_size) + model = model.cuda() + model.eval() + _, c, h, w = input_size + dummy_input = torch.randn(MAX_BATCH_SIZE, c, h, w, device='cuda') + torch.onnx.export(model, dummy_input, "model.onnx", verbose=True, input_names=["input"], output_names=["output"], export_params=True,) + + with build_engine("model.onnx") as engine: + print('engine', engine) + h_input, d_input, h_output, d_output = allocate_buffers(engine) + load_input(input_size, h_input) + with engine.create_execution_context() as context: + latency = do_inference(context, h_input, d_input, h_output, d_output, iterations=iterations) + # FPS = 1000 / latency (in ms) + print('MAX_BATCH_SIZE: ', MAX_BATCH_SIZE) + return latency/ MAX_BATCH_SIZE +# except: +# warnings.warn("TensorRT (or pycuda) is not installed. compute_latency_ms_tensorrt() cannot be used.") +######################################################################### + +def compute_latency_ms_pytorch(model, input_size, iterations=None, device=None): + torch.backends.cudnn.enabled = True + torch.backends.cudnn.benchmark = True + + model.eval() + # model = model.cpu() + # input = torch.randn(*input_size) + model = model.cuda() + input = torch.randn(*input_size).cuda() + + with torch.no_grad(): + for _ in range(10): + model(input) + + if iterations is None: + elapsed_time = 0 + iterations = 100 + while elapsed_time < 1: + torch.cuda.synchronize() + torch.cuda.synchronize() + t_start = time.time() + for _ in range(iterations): + model(input) + torch.cuda.synchronize() + torch.cuda.synchronize() + elapsed_time = time.time() - t_start + iterations *= 2 + FPS = iterations / elapsed_time + iterations = int(FPS * 6) + + print('=========Speed Testing=========') + torch.cuda.synchronize() + torch.cuda.synchronize() + t_start = time.time() + for _ in tqdm(range(iterations)): + model(input) + torch.cuda.synchronize() + torch.cuda.synchronize() + elapsed_time = time.time() - t_start + latency = elapsed_time / iterations * 1000 + torch.cuda.empty_cache() + # FPS = 1000 / latency (in ms) + return latency + + +def plot_path(lasts, paths=[]): + ''' + paths: list of path0~path2 + ''' + assert len(paths) > 0 + path0 = paths[0] + path1 = paths[1] if len(paths) > 1 else [] + path2 = paths[2] if len(paths) > 2 else [] + + if path0[-1] != lasts[0]: path0.append(lasts[0]) + if len(path1) != 0 and path1[-1] != lasts[1]: path1.append(lasts[1]) + if len(path2) != 0 and path2[-1] != lasts[2]: path2.append(lasts[2]) + x_len = max(len(path0), len(path1), len(path2)) + f, ax = plt.subplots(figsize=(x_len, 3)) + ax.plot(np.arange(len(path0)), 2 - np.array(path0), label='1/32', lw=2.5, color='#000000', linestyle='-')#, marker='o', markeredgecolor='r', markerfacecolor='r') + ax.plot(np.arange(len(path1)), 2 - np.array(path1) - 0.08, lw=1.8, label='1/16', color='#313131', linestyle='--')#, marker='^', markeredgecolor='b', markerfacecolor='b') + ax.plot(np.arange(len(path2)), 2 - np.array(path2) - 0.16, lw=1.2, label='1/8', color='#5a5858', linestyle='-.')#, marker='s', markeredgecolor='m', markerfacecolor='m') + plt.xticks(np.arange(x_len), list(range(1, x_len+1))) + plt.yticks(np.array([0, 1, 2]), ["1/32", "1/16", "1/8"]) + plt.ylabel("Scale", fontsize=17) + plt.xlabel("Layer", fontsize=17) + for tick in ax.xaxis.get_major_ticks(): + tick.label.set_fontsize(14) + for tick in ax.yaxis.get_major_ticks(): + tick.label.set_fontsize(14) + f.tight_layout() + plt.legend(prop={'size': 14}, loc=3) + return f + + +def plot_path_width(lasts, paths=[], widths=[]): + ''' + paths: list of path0~path2 + ''' + assert len(paths) > 0 and len(widths) > 0 + path0 = paths[0] + path1 = paths[1] if len(paths) > 1 else [] + path2 = paths[2] if len(paths) > 2 else [] + width0 = widths[0] + width1 = widths[1] if len(widths) > 1 else [] + width2 = widths[2] if len(widths) > 2 else [] + + # just for visualization purpose + if path0[-1] != lasts[0]: path0.append(lasts[0]) + if len(path1) != 0 and path1[-1] != lasts[1]: path1.append(lasts[1]) + if len(path2) != 0 and path2[-1] != lasts[2]: path2.append(lasts[2]) + line_updown = -0.07 + annotation_updown = 0.05; annotation_down_scale = 1.7 + x_len = max(len(path0), len(path1), len(path2)) + f, ax = plt.subplots(figsize=(x_len, 3)) + + assert len(path0) == len(width0) + 1 or len(path0) + len(width0) == 0, "path0 %d, width0 %d"%(len(path0), len(width0)) + assert len(path1) == len(width1) + 1 or len(path1) + len(width1) == 0, "path1 %d, width1 %d"%(len(path1), len(width1)) + assert len(path2) == len(width2) + 1 or len(path2) + len(width2) == 0, "path2 %d, width2 %d"%(len(path2), len(width2)) + + ax.plot(np.arange(len(path0)), 2 - np.array(path0), label='1/32', lw=2.5, color='#000000', linestyle='-') + ax.plot(np.arange(len(path1)), 2 - np.array(path1) + line_updown, lw=1.8, label='1/16', color='#313131', linestyle='--') + ax.plot(np.arange(len(path2)), 2 - np.array(path2) + line_updown*2, lw=1.2, label='1/8', color='#5a5858', linestyle='-.') + + annotations = {} # (idx, scale, width, down): ((x, y), width) + for idx, width in enumerate(width2): + annotations[(idx, path2[idx], width, path2[idx+1]-path2[idx])] = ((0.35 + idx, 2 - path2[idx] + line_updown*2 + annotation_updown - (path2[idx+1]-path2[idx])/annotation_down_scale), width) + for idx, width in enumerate(width1): + annotations[(idx, path1[idx], width, path1[idx+1]-path1[idx])] = ((0.35 + idx, 2 - path1[idx] + line_updown + annotation_updown - (path1[idx+1]-path1[idx])/annotation_down_scale), width) + for idx, width in enumerate(width0): + annotations[(idx, path0[idx], width, path0[idx+1]-path0[idx])] = ((0.35 + idx, 2 - path0[idx] + annotation_updown - (path0[idx+1]-path0[idx])/annotation_down_scale), width) + for k, v in annotations.items(): + plt.annotate("%.2f"%v[1], v[0], fontsize=12, color='red') + + plt.xticks(np.arange(x_len), list(range(1, x_len+1))) + plt.yticks(np.array([0, 1, 2]), ["1/32", "1/16", "1/8"]) + plt.ylim([-0.4, 2.5]) + plt.ylabel("Scale", fontsize=17) + plt.xlabel("Layer", fontsize=17) + for tick in ax.xaxis.get_major_ticks(): + tick.label.set_fontsize(14) + for tick in ax.yaxis.get_major_ticks(): + tick.label.set_fontsize(14) + f.tight_layout() + plt.legend(prop={'size': 14}, loc=3) + return f + +def plot_op(ops, path, width=[], head_width=None, F_base=16): + assert len(width) == 0 or len(width) == len(ops) - 1 + table_vals = [] + scales = {0: "1/8", 1: "1/16", 2: "1/32"}; base_scale = 3 + for idx, op in enumerate(ops): + scale = path[idx] + if len(width) > 0: + if idx < len(width): + ch = int(F_base*2**(scale+base_scale)*width[idx]) + else: + ch = int(F_base*2**(scale+base_scale)*head_width) + else: + ch = F_base*2**(scale+base_scale) + row = [idx+1, PRIMITIVES[op], scales[scale], ch] + table_vals.append(row) + + # Based on http://stackoverflow.com/a/8531491/190597 (Andrey Sobolev) + col_labels = ['Stage', 'Operator', 'Scale', '#Channel_out'] + plt.tight_layout() + fig = plt.figure(figsize=(3,3)) + ax = fig.add_subplot(111, frame_on=False) + ax.xaxis.set_visible(False) # hide the x axis + ax.yaxis.set_visible(False) # hide the y axis + + table = plt.table(cellText=table_vals, + colWidths=[0.22, 0.6, 0.25, 0.5], + colLabels=col_labels, + cellLoc='center', + loc='center') + table.auto_set_font_size(False) + table.set_fontsize(20) + table.scale(2, 2) + + return fig + +def objective_acc_lat(acc, lat, lat_target=8.3, alpha=-0.07, beta=-0.07): + if lat <= lat_target: + w = alpha + else: + w = beta + return acc * math.pow(lat / lat_target, w) \ No newline at end of file diff --git a/latency/utils/genotypes.py b/latency/utils/genotypes.py new file mode 100644 index 0000000..eed8eff --- /dev/null +++ b/latency/utils/genotypes.py @@ -0,0 +1,75 @@ +from collections import namedtuple + +Genotype = namedtuple('Genotype', 'normal normal_concat reduce reduce_concat') + +PRIMITIVES = [ + 'skip', + 'conv', + 'conv_di', + 'conv_2x', + 'conv_2x_di', +] + +NASNet = Genotype( + normal = [ + ('sep_conv_5x5', 1), + ('sep_conv_3x3', 0), + ('sep_conv_5x5', 0), + ('sep_conv_3x3', 0), + ('avg_pool_3x3', 1), + ('skip_connect', 0), + ('avg_pool_3x3', 0), + ('avg_pool_3x3', 0), + ('sep_conv_3x3', 1), + ('skip_connect', 1), + ], + normal_concat = [2, 3, 4, 5, 6], + reduce = [ + ('sep_conv_5x5', 1), + ('sep_conv_7x7', 0), + ('max_pool_3x3', 1), + ('sep_conv_7x7', 0), + ('avg_pool_3x3', 1), + ('sep_conv_5x5', 0), + ('skip_connect', 3), + ('avg_pool_3x3', 2), + ('sep_conv_3x3', 2), + ('max_pool_3x3', 1), + ], + reduce_concat = [4, 5, 6], +) + +AmoebaNet = Genotype( + normal = [ + ('avg_pool_3x3', 0), + ('max_pool_3x3', 1), + ('sep_conv_3x3', 0), + ('sep_conv_5x5', 2), + ('sep_conv_3x3', 0), + ('avg_pool_3x3', 3), + ('sep_conv_3x3', 1), + ('skip_connect', 1), + ('skip_connect', 0), + ('avg_pool_3x3', 1), + ], + normal_concat = [4, 5, 6], + reduce = [ + ('avg_pool_3x3', 0), + ('sep_conv_3x3', 1), + ('max_pool_3x3', 0), + ('sep_conv_7x7', 2), + ('sep_conv_7x7', 0), + ('avg_pool_3x3', 1), + ('max_pool_3x3', 0), + ('max_pool_3x3', 1), + ('conv_7x1_1x7', 0), + ('sep_conv_3x3', 5), + ], + reduce_concat = [3, 4, 6] +) + +DARTS_V1 = Genotype(normal=[('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('skip_connect', 0), ('sep_conv_3x3', 1), ('skip_connect', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('skip_connect', 2)], normal_concat=[2, 3, 4, 5], reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2), ('max_pool_3x3', 0), ('max_pool_3x3', 0), ('skip_connect', 2), ('skip_connect', 2), ('avg_pool_3x3', 0)], reduce_concat=[2, 3, 4, 5]) +DARTS_V2 = Genotype(normal=[('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 1), ('skip_connect', 0), ('skip_connect', 0), ('dil_conv_3x3', 2)], normal_concat=[2, 3, 4, 5], reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2), ('max_pool_3x3', 1), ('max_pool_3x3', 0), ('skip_connect', 2), ('skip_connect', 2), ('max_pool_3x3', 1)], reduce_concat=[2, 3, 4, 5]) + +DARTS = DARTS_V2 + diff --git a/logger.py b/logger.py new file mode 100644 index 0000000..d3f9ddc --- /dev/null +++ b/logger.py @@ -0,0 +1,23 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- + + +import os.path as osp +import time +import sys +import logging + +import torch.distributed as dist + + +def setup_logger(logpth): + logfile = 'BiSeNet-{}.log'.format(time.strftime('%Y-%m-%d-%H-%M-%S')) + logfile = osp.join(logpth, logfile) + FORMAT = '%(levelname)s %(filename)s(%(lineno)d): %(message)s' + log_level = logging.INFO + if dist.is_initialized() and not dist.get_rank()==0: + log_level = logging.ERROR + logging.basicConfig(level=log_level, format=FORMAT, filename=logfile) + logging.root.addHandler(logging.StreamHandler()) + + diff --git a/loss/__pycache__/detail_loss.cpython-37.pyc b/loss/__pycache__/detail_loss.cpython-37.pyc new file mode 100644 index 0000000..2c12e98 Binary files /dev/null and b/loss/__pycache__/detail_loss.cpython-37.pyc differ diff --git a/loss/__pycache__/detail_loss.cpython-38.pyc b/loss/__pycache__/detail_loss.cpython-38.pyc new file mode 100644 index 0000000..3900f3c Binary files /dev/null and b/loss/__pycache__/detail_loss.cpython-38.pyc differ diff --git a/loss/__pycache__/loss.cpython-37.pyc b/loss/__pycache__/loss.cpython-37.pyc new file mode 100644 index 0000000..3831b43 Binary files /dev/null and b/loss/__pycache__/loss.cpython-37.pyc differ diff --git a/loss/__pycache__/loss.cpython-38.pyc b/loss/__pycache__/loss.cpython-38.pyc new file mode 100644 index 0000000..e1fc8b2 Binary files /dev/null and b/loss/__pycache__/loss.cpython-38.pyc differ diff --git a/loss/__pycache__/util.cpython-37.pyc b/loss/__pycache__/util.cpython-37.pyc new file mode 100644 index 0000000..6a33c04 Binary files /dev/null and b/loss/__pycache__/util.cpython-37.pyc differ diff --git a/loss/__pycache__/util.cpython-38.pyc b/loss/__pycache__/util.cpython-38.pyc new file mode 100644 index 0000000..906a0fc Binary files /dev/null and b/loss/__pycache__/util.cpython-38.pyc differ diff --git a/loss/detail_loss.py b/loss/detail_loss.py new file mode 100644 index 0000000..ca18b08 --- /dev/null +++ b/loss/detail_loss.py @@ -0,0 +1,128 @@ + +import torch +from torch import nn +from torch.nn import functional as F +import cv2 +import numpy as np +import json + +def dice_loss_func(input, target): + smooth = 1. + n = input.size(0) + iflat = input.view(n, -1) + tflat = target.view(n, -1) + intersection = (iflat * tflat).sum(1) + loss = 1 - ((2. * intersection + smooth) / + (iflat.sum(1) + tflat.sum(1) + smooth)) + return loss.mean() + +def get_one_hot(label, N): + size = list(label.size()) + label = label.view(-1) # reshape 为向量 + ones = torch.sparse.torch.eye(N).cuda() + ones = ones.index_select(0, label.long()) # 用上面的办法转为换one hot + size.append(N) # 把类别输目添到size的尾后，准备reshape回原来的尺寸 + return ones.view(*size) + +def get_boundary(gtmasks): + + laplacian_kernel = torch.tensor( + [-1, -1, -1, -1, 8, -1, -1, -1, -1], + dtype=torch.float32, device=gtmasks.device).reshape(1, 1, 3, 3).requires_grad_(False) + # boundary_logits = boundary_logits.unsqueeze(1) + boundary_targets = F.conv2d(gtmasks.unsqueeze(1), laplacian_kernel, padding=1) + boundary_targets = boundary_targets.clamp(min=0) + boundary_targets[boundary_targets > 0.1] = 1 + boundary_targets[boundary_targets <= 0.1] = 0 + return boundary_targets + + +class DetailAggregateLoss(nn.Module): + def __init__(self, *args, **kwargs): + super(DetailAggregateLoss, self).__init__() + + self.laplacian_kernel = torch.tensor( + [-1, -1, -1, -1, 8, -1, -1, -1, -1], + dtype=torch.float32).reshape(1, 1, 3, 3).requires_grad_(False).type(torch.cuda.FloatTensor) + + self.fuse_kernel = torch.nn.Parameter(torch.tensor([[6./10], [3./10], [1./10]], + dtype=torch.float32).reshape(1, 3, 1, 1).type(torch.cuda.FloatTensor)) + + def forward(self, boundary_logits, gtmasks): + + # boundary_logits = boundary_logits.unsqueeze(1) + boundary_targets = F.conv2d(gtmasks.unsqueeze(1).type(torch.cuda.FloatTensor), self.laplacian_kernel, padding=1) + boundary_targets = boundary_targets.clamp(min=0) + boundary_targets[boundary_targets > 0.1] = 1 + boundary_targets[boundary_targets <= 0.1] = 0 + + boundary_targets_x2 = F.conv2d(gtmasks.unsqueeze(1).type(torch.cuda.FloatTensor), self.laplacian_kernel, stride=2, padding=1) + boundary_targets_x2 = boundary_targets_x2.clamp(min=0) + + boundary_targets_x4 = F.conv2d(gtmasks.unsqueeze(1).type(torch.cuda.FloatTensor), self.laplacian_kernel, stride=4, padding=1) + boundary_targets_x4 = boundary_targets_x4.clamp(min=0) + + boundary_targets_x8 = F.conv2d(gtmasks.unsqueeze(1).type(torch.cuda.FloatTensor), self.laplacian_kernel, stride=8, padding=1) + boundary_targets_x8 = boundary_targets_x8.clamp(min=0) + + boundary_targets_x8_up = F.interpolate(boundary_targets_x8, boundary_targets.shape[2:], mode='nearest') + boundary_targets_x4_up = F.interpolate(boundary_targets_x4, boundary_targets.shape[2:], mode='nearest') + boundary_targets_x2_up = F.interpolate(boundary_targets_x2, boundary_targets.shape[2:], mode='nearest') + + boundary_targets_x2_up[boundary_targets_x2_up > 0.1] = 1 + boundary_targets_x2_up[boundary_targets_x2_up <= 0.1] = 0 + + + boundary_targets_x4_up[boundary_targets_x4_up > 0.1] = 1 + boundary_targets_x4_up[boundary_targets_x4_up <= 0.1] = 0 + + + boundary_targets_x8_up[boundary_targets_x8_up > 0.1] = 1 + boundary_targets_x8_up[boundary_targets_x8_up <= 0.1] = 0 + + boudary_targets_pyramids = torch.stack((boundary_targets, boundary_targets_x2_up, boundary_targets_x4_up), dim=1) + + boudary_targets_pyramids = boudary_targets_pyramids.squeeze(2) + boudary_targets_pyramid = F.conv2d(boudary_targets_pyramids, self.fuse_kernel) + + boudary_targets_pyramid[boudary_targets_pyramid > 0.1] = 1 + boudary_targets_pyramid[boudary_targets_pyramid <= 0.1] = 0 + + + if boundary_logits.shape[-1] != boundary_targets.shape[-1]: + boundary_logits = F.interpolate( + boundary_logits, boundary_targets.shape[2:], mode='bilinear', align_corners=True) + + bce_loss = F.binary_cross_entropy_with_logits(boundary_logits, boudary_targets_pyramid) + dice_loss = dice_loss_func(torch.sigmoid(boundary_logits), boudary_targets_pyramid) + return bce_loss, dice_loss + + def get_params(self): + wd_params, nowd_params = [], [] + for name, module in self.named_modules(): + nowd_params += list(module.parameters()) + return nowd_params + +if __name__ == '__main__': + torch.manual_seed(15) + with open('../cityscapes_info.json', 'r') as fr: + labels_info = json.load(fr) + lb_map = {el['id']: el['trainId'] for el in labels_info} + + img_path = 'data/gtFine/val/frankfurt/frankfurt_000001_037705_gtFine_labelIds.png' + img = cv2.imread(img_path, 0) + + label = np.zeros(img.shape, np.uint8) + for k, v in lb_map.items(): + label[img == k] = v + + img_tensor = torch.from_numpy(label).cuda() + img_tensor = torch.unsqueeze(img_tensor, 0).type(torch.cuda.FloatTensor) + + + detailAggregateLoss = DetailAggregateLoss() + for param in detailAggregateLoss.parameters(): + print(param) + + bce_loss, dice_loss = detailAggregateLoss(torch.unsqueeze(img_tensor, 0), img_tensor) + print(bce_loss, dice_loss) \ No newline at end of file diff --git a/loss/loss.py b/loss/loss.py new file mode 100644 index 0000000..4264bfa --- /dev/null +++ b/loss/loss.py @@ -0,0 +1,95 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- + + +import torch +import torch.nn as nn +import torch.nn.functional as F +from loss.util import enet_weighing +import numpy as np + + +class OhemCELoss(nn.Module): + def __init__(self, thresh, n_min, ignore_lb=255, *args, **kwargs): + super(OhemCELoss, self).__init__() + self.thresh = -torch.log(torch.tensor(thresh, dtype=torch.float)).cuda() + self.n_min = n_min + self.ignore_lb = ignore_lb + + self.criteria = nn.CrossEntropyLoss(ignore_index=ignore_lb, reduction='none') + + def forward(self, logits, labels): + N, C, H, W = logits.size() + loss = self.criteria(logits, labels).view(-1) + loss, _ = torch.sort(loss, descending=True) + if loss[self.n_min] > self.thresh: + loss = loss[loss>self.thresh] + else: + loss = loss[:self.n_min] + return torch.mean(loss) + +class WeightedOhemCELoss(nn.Module): + def __init__(self, thresh, n_min, num_classes, ignore_lb=255, *args, **kwargs): + super(WeightedOhemCELoss, self).__init__() + self.thresh = -torch.log(torch.tensor(thresh, dtype=torch.float)).cuda() + self.n_min = n_min + self.ignore_lb = ignore_lb + self.num_classes = num_classes + # self.criteria = nn.CrossEntropyLoss(ignore_index=ignore_lb, reduction='none') + + def forward(self, logits, labels): + N, C, H, W = logits.size() + criteria = nn.CrossEntropyLoss(weight=enet_weighing(labels, self.num_classes).cuda(), ignore_index=self.ignore_lb, reduction='none') + loss = criteria(logits, labels).view(-1) + loss, _ = torch.sort(loss, descending=True) + if loss[self.n_min] > self.thresh: + loss = loss[loss>self.thresh] + else: + loss = loss[:self.n_min] + return torch.mean(loss) + +class SoftmaxFocalLoss(nn.Module): + def __init__(self, gamma, ignore_lb=255, *args, **kwargs): + super(FocalLoss, self).__init__() + self.gamma = gamma + self.nll = nn.NLLLoss(ignore_index=ignore_lb) + + def forward(self, logits, labels): + scores = F.softmax(logits, dim=1) + factor = torch.pow(1.-scores, self.gamma) + log_score = F.log_softmax(logits, dim=1) + log_score = factor * log_score + loss = self.nll(log_score, labels) + return loss + + +if __name__ == '__main__': + torch.manual_seed(15) + criteria1 = OhemCELoss(thresh=0.7, n_min=16*20*20//16).cuda() + criteria2 = OhemCELoss(thresh=0.7, n_min=16*20*20//16).cuda() + net1 = nn.Sequential( + nn.Conv2d(3, 19, kernel_size=3, stride=2, padding=1), + ) + net1.cuda() + net1.train() + net2 = nn.Sequential( + nn.Conv2d(3, 19, kernel_size=3, stride=2, padding=1), + ) + net2.cuda() + net2.train() + + with torch.no_grad(): + inten = torch.randn(16, 3, 20, 20).cuda() + lbs = torch.randint(0, 19, [16, 20, 20]).cuda() + lbs[1, :, :] = 255 + + logits1 = net1(inten) + logits1 = F.interpolate(logits1, inten.size()[2:], mode='bilinear') + logits2 = net2(inten) + logits2 = F.interpolate(logits2, inten.size()[2:], mode='bilinear') + + loss1 = criteria1(logits1, lbs) + loss2 = criteria2(logits2, lbs) + loss = loss1 + loss2 + print(loss.detach().cpu()) + loss.backward() diff --git a/loss/util.py b/loss/util.py new file mode 100644 index 0000000..9ac9b27 --- /dev/null +++ b/loss/util.py @@ -0,0 +1,43 @@ +import numpy as np +import torch + +def enet_weighing(label, num_classes, c=1.02): + """Computes class weights as described in the ENet paper: + w_class = 1 / (ln(c + p_class)), + where c is usually 1.02 and p_class is the propensity score of that + class: + propensity_score = freq_class / total_pixels. + References: https://arxiv.org/abs/1606.02147 + Keyword arguments: + - dataloader (``data.Dataloader``): A data loader to iterate over the + dataset. + - num_classes (``int``): The number of classes. + - c (``int``, optional): AN additional hyper-parameter which restricts + the interval of values for the weights. Default: 1.02. + """ + class_count = 0 + total = 0 + + label = label.cpu().numpy() + + # Flatten label + flat_label = label.flatten() + + # Sum up the number of pixels of each class and the total pixel + # counts for each label + class_count += np.bincount(flat_label, minlength=num_classes) + total += flat_label.size + + # Compute propensity score and then the weights for each class + propensity_score = class_count / total + class_weights = 1 / (np.log(c + propensity_score)) + + class_weights = torch.from_numpy(class_weights).float() + # print(class_weights) + return class_weights + +def minmax_scale(input_arr): + min_val = np.min(input_arr) + max_val = np.max(input_arr) + output_arr = (input_arr - min_val) * 255.0 / (max_val - min_val) + return output_arr \ No newline at end of file diff --git a/models/__init__.py b/models/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/models/__pycache__/__init__.cpython-37.pyc b/models/__pycache__/__init__.cpython-37.pyc new file mode 100644 index 0000000..6009c59 Binary files /dev/null and b/models/__pycache__/__init__.cpython-37.pyc differ diff --git a/models/__pycache__/__init__.cpython-38.pyc b/models/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000..7af6b21 Binary files /dev/null and b/models/__pycache__/__init__.cpython-38.pyc differ diff --git a/models/__pycache__/__init__.cpython-39.pyc b/models/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000..dc71787 Binary files /dev/null and b/models/__pycache__/__init__.cpython-39.pyc differ diff --git a/models/__pycache__/common.cpython-38.pyc b/models/__pycache__/common.cpython-38.pyc new file mode 100644 index 0000000..9607d61 Binary files /dev/null and b/models/__pycache__/common.cpython-38.pyc differ diff --git a/models/__pycache__/experimental.cpython-38.pyc b/models/__pycache__/experimental.cpython-38.pyc new file mode 100644 index 0000000..fbb33c0 Binary files /dev/null and b/models/__pycache__/experimental.cpython-38.pyc differ diff --git a/models/__pycache__/model_stages.cpython-37.pyc b/models/__pycache__/model_stages.cpython-37.pyc new file mode 100644 index 0000000..5da7b57 Binary files /dev/null and b/models/__pycache__/model_stages.cpython-37.pyc differ diff --git a/models/__pycache__/model_stages.cpython-38.pyc b/models/__pycache__/model_stages.cpython-38.pyc new file mode 100644 index 0000000..a5742a1 Binary files /dev/null and b/models/__pycache__/model_stages.cpython-38.pyc differ diff --git a/models/__pycache__/model_stages.cpython-39.pyc b/models/__pycache__/model_stages.cpython-39.pyc new file mode 100644 index 0000000..7a01916 Binary files /dev/null and b/models/__pycache__/model_stages.cpython-39.pyc differ diff --git a/models/__pycache__/model_stages_trt.cpython-37.pyc b/models/__pycache__/model_stages_trt.cpython-37.pyc new file mode 100644 index 0000000..3f3b9f6 Binary files /dev/null and b/models/__pycache__/model_stages_trt.cpython-37.pyc differ diff --git a/models/__pycache__/yolo.cpython-38.pyc b/models/__pycache__/yolo.cpython-38.pyc new file mode 100644 index 0000000..e450879 Binary files /dev/null and b/models/__pycache__/yolo.cpython-38.pyc differ diff --git a/models/bisenet.py b/models/bisenet.py new file mode 100644 index 0000000..d3e269e --- /dev/null +++ b/models/bisenet.py @@ -0,0 +1,323 @@ +"""Bilateral Segmentation Network""" +import torch +import torch.nn as nn +import torch.nn.functional as F +import numpy as np +# from core.models.base_models.resnet import resnet18,resnet50 +from torchvision import models +# from core.nn import _ConvBNReLU + +# __all__ = ['BiSeNet', 'get_bisenet', 'get_bisenet_resnet18_citys'] + +class _ConvBNReLU(nn.Module): + def __init__(self,in_channels,out_channels, k, s, p, norm_layer=None): + super(_ConvBNReLU, self).__init__() + self.conv =nn.Conv2d(in_channels, out_channels, kernel_size=k, stride=s, padding=p) + self.bn = nn.BatchNorm2d(out_channels) + self.relu = nn.ReLU(inplace = True) + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + + return x +class BiSeNet(nn.Module): + def __init__(self, nclass, backbone='resnet18', aux=False, jpu=False, pretrained_base=True, **kwargs): + super(BiSeNet, self).__init__() + self.aux = aux + self.spatial_path = SpatialPath(3, 128, **kwargs) + self.context_path = ContextPath(backbone, pretrained_base, **kwargs) + self.ffm = FeatureFusion(256, 256, 4, **kwargs) + self.head = _BiSeHead(256, 64, nclass, **kwargs) + if aux: + self.auxlayer1 = _BiSeHead(128, 256, nclass, **kwargs) + self.auxlayer2 = _BiSeHead(128, 256, nclass, **kwargs) + + self.__setattr__('exclusive', + ['spatial_path', 'context_path', 'ffm', 'head', 'auxlayer1', 'auxlayer2'] if aux else [ + 'spatial_path', 'context_path', 'ffm', 'head']) + + def forward(self, x,outsize=None,test_flag=False): + size = x.size()[2:] + spatial_out = self.spatial_path(x) + context_out = self.context_path(x) + fusion_out = self.ffm(spatial_out, context_out[-1]) + outputs = [] + x = self.head(fusion_out) + x = F.interpolate(x, size, mode='bilinear', align_corners=True) + + if outsize: + print('######using torch resize#######',outsize) + x = F.interpolate(x, outsize, mode='bilinear', align_corners=True) + outputs.append(x) + + if self.aux: + auxout1 = self.auxlayer1(context_out[0]) + auxout1 = F.interpolate(auxout1, size, mode='bilinear', align_corners=True) + outputs.append(auxout1) + auxout2 = self.auxlayer2(context_out[1]) + auxout2 = F.interpolate(auxout2, size, mode='bilinear', align_corners=True) + outputs.append(auxout2) + if test_flag: + outputs = [torch.argmax(outputx, axis=1) for outputx in outputs] + #return tuple(outputs) + return outputs[0] + +class BiSeNet_MultiOutput(nn.Module): + def __init__(self, nclass, backbone='resnet18', aux=False, jpu=False, pretrained_base=True, **kwargs): + super(BiSeNet_MultiOutput, self).__init__() + self.aux = aux + self.spatial_path = SpatialPath(3, 128, **kwargs) + self.context_path = ContextPath(backbone, pretrained_base, **kwargs) + self.ffm = FeatureFusion(256, 256, 4, **kwargs) + assert isinstance(nclass, list) + self.outCnt = len(nclass) + for ii, nclassii in enumerate(nclass): + setattr(self, 'head%d'%(ii), _BiSeHead(256, 64, nclassii, **kwargs)) + + if aux: + self.auxlayer1 = _BiSeHead(128, 256, nclass, **kwargs) + self.auxlayer2 = _BiSeHead(128, 256, nclass, **kwargs) + + self.__setattr__('exclusive', + ['spatial_path', 'context_path', 'ffm', 'head', 'auxlayer1', 'auxlayer2'] if aux else [ + 'spatial_path', 'context_path', 'ffm', 'head']) + + def forward(self, x, outsize=None, test_flag=False, smooth_kernel=0): + size = x.size()[2:] + spatial_out = self.spatial_path(x) + context_out = self.context_path(x) + fusion_out = self.ffm(spatial_out, context_out[-1]) + outputs = [] + for ii in range(self.outCnt): + x = getattr(self, 'head%d'%(ii))(fusion_out) + x = F.interpolate(x, size, mode='bilinear', align_corners=True) + outputs.append(x) + + if self.aux: + auxout1 = self.auxlayer1(context_out[0]) + auxout1 = F.interpolate(auxout1, size, mode='bilinear', align_corners=True) + outputs.append(auxout1) + auxout2 = self.auxlayer2(context_out[1]) + auxout2 = F.interpolate(auxout2, size, mode='bilinear', align_corners=True) + outputs.append(auxout2) + if test_flag: + outputs = [torch.argmax(outputx ,axis=1) for outputx in outputs] + if smooth_kernel>0: + gaussian_kernel = torch.from_numpy(np.ones((1,1,smooth_kernel,smooth_kernel)) ) + + pad = int((smooth_kernel - 1)/2) + if not gaussian_kernel.is_cuda: + gaussian_kernel = gaussian_kernel.to(x.device) + #print(gaussian_kernel.dtype,gaussian_kernel,outputs[0].dtype) + outputs = [x.unsqueeze(1).double() for x in outputs] + outputs = [torch.conv2d(x, gaussian_kernel, padding=pad) for x in outputs] + outputs = [x.squeeze(1).long() for x in outputs] + #return tuple(outputs) + return outputs + +class _BiSeHead(nn.Module): + def __init__(self, in_channels, inter_channels, nclass, norm_layer=nn.BatchNorm2d, **kwargs): + super(_BiSeHead, self).__init__() + self.block = nn.Sequential( + _ConvBNReLU(in_channels, inter_channels, 3, 1, 1, norm_layer=norm_layer), + nn.Dropout(0.1), + nn.Conv2d(inter_channels, nclass, 1) + ) + + def forward(self, x): + x = self.block(x) + return x + + +class SpatialPath(nn.Module): + """Spatial path""" + + def __init__(self, in_channels, out_channels, norm_layer=nn.BatchNorm2d, **kwargs): + super(SpatialPath, self).__init__() + inter_channels = 64 + self.conv7x7 = _ConvBNReLU(in_channels, inter_channels, 7, 2, 3, norm_layer=norm_layer) + self.conv3x3_1 = _ConvBNReLU(inter_channels, inter_channels, 3, 2, 1, norm_layer=norm_layer) + self.conv3x3_2 = _ConvBNReLU(inter_channels, inter_channels, 3, 2, 1, norm_layer=norm_layer) + self.conv1x1 = _ConvBNReLU(inter_channels, out_channels, 1, 1, 0, norm_layer=norm_layer) + + def forward(self, x): + x = self.conv7x7(x) + x = self.conv3x3_1(x) + x = self.conv3x3_2(x) + x = self.conv1x1(x) + + return x + + +class _GlobalAvgPooling(nn.Module): + def __init__(self, in_channels, out_channels, norm_layer, **kwargs): + super(_GlobalAvgPooling, self).__init__() + self.gap = nn.Sequential( + nn.AdaptiveAvgPool2d(1), + nn.Conv2d(in_channels, out_channels, 1, bias=False), + norm_layer(out_channels), + nn.ReLU(True) + ) + + def forward(self, x): + size = x.size()[2:] + pool = self.gap(x) + out = F.interpolate(pool, size, mode='bilinear', align_corners=True) + return out + + +class AttentionRefinmentModule(nn.Module): + def __init__(self, in_channels, out_channels, norm_layer=nn.BatchNorm2d, **kwargs): + super(AttentionRefinmentModule, self).__init__() + self.conv3x3 = _ConvBNReLU(in_channels, out_channels, 3, 1, 1, norm_layer=norm_layer) + self.channel_attention = nn.Sequential( + nn.AdaptiveAvgPool2d(1), + _ConvBNReLU(out_channels, out_channels, 1, 1, 0, norm_layer=norm_layer), + nn.Sigmoid() + ) + + def forward(self, x): + x = self.conv3x3(x) + attention = self.channel_attention(x) + x = x * attention + return x + + +class ContextPath(nn.Module): + def __init__(self, backbone='resnet18', pretrained_base=True, norm_layer=nn.BatchNorm2d, **kwargs): + super(ContextPath, self).__init__() + if backbone == 'resnet18': + pretrained = models.resnet18(pretrained=pretrained_base, **kwargs) + elif backbone=='resnet50': + pretrained = models.resnet50(pretrained=pretrained_base, **kwargs) + else: + raise RuntimeError('unknown backbone: {}'.format(backbone)) + self.conv1 = pretrained.conv1 + self.bn1 = pretrained.bn1 + self.relu = pretrained.relu + self.maxpool = pretrained.maxpool + self.layer1 = pretrained.layer1 + self.layer2 = pretrained.layer2 + self.layer3 = pretrained.layer3 + self.layer4 = pretrained.layer4 + + inter_channels = 128 + self.global_context = _GlobalAvgPooling(512, inter_channels, norm_layer) + + self.arms = nn.ModuleList( + [AttentionRefinmentModule(512, inter_channels, norm_layer, **kwargs), + AttentionRefinmentModule(256, inter_channels, norm_layer, **kwargs)] + ) + self.refines = nn.ModuleList( + [_ConvBNReLU(inter_channels, inter_channels, 3, 1, 1, norm_layer=norm_layer), + _ConvBNReLU(inter_channels, inter_channels, 3, 1, 1, norm_layer=norm_layer)] + ) + + def forward(self, x): + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + x = self.layer1(x) + + context_blocks = [] + context_blocks.append(x) + x = self.layer2(x) + context_blocks.append(x) + c3 = self.layer3(x) + context_blocks.append(c3) + c4 = self.layer4(c3) + context_blocks.append(c4) + context_blocks.reverse() + + global_context = self.global_context(c4) + last_feature = global_context + context_outputs = [] + for i, (feature, arm, refine) in enumerate(zip(context_blocks[:2], self.arms, self.refines)): + feature = arm(feature) + feature += last_feature + last_feature = F.interpolate(feature, size=context_blocks[i + 1].size()[2:], + mode='bilinear', align_corners=True) + last_feature = refine(last_feature) + context_outputs.append(last_feature) + + return context_outputs + + +class FeatureFusion(nn.Module): + def __init__(self, in_channels, out_channels, reduction=1, norm_layer=nn.BatchNorm2d, **kwargs): + super(FeatureFusion, self).__init__() + self.conv1x1 = _ConvBNReLU(in_channels, out_channels, 1, 1, 0, norm_layer=norm_layer, **kwargs) + self.channel_attention = nn.Sequential( + nn.AdaptiveAvgPool2d(1), + _ConvBNReLU(out_channels, out_channels // reduction, 1, 1, 0, norm_layer=norm_layer), + _ConvBNReLU(out_channels // reduction, out_channels, 1, 1, 0, norm_layer=norm_layer), + nn.Sigmoid() + ) + + def forward(self, x1, x2): + fusion = torch.cat([x1, x2], dim=1) + out = self.conv1x1(fusion) + attention = self.channel_attention(out) + out = out + out * attention + return out + + +# def get_bisenet(dataset='citys', backbone='resnet18', pretrained=False, root='~/.torch/models', +# pretrained_base=True, **kwargs): +# acronyms = { +# 'pascal_voc': 'pascal_voc', +# 'pascal_aug': 'pascal_aug', +# 'ade20k': 'ade', +# 'coco': 'coco', +# 'citys': 'citys', +# } +# from ..data.dataloader import datasets +# model = BiSeNet(datasets[dataset].NUM_CLASS, backbone=backbone, pretrained_base=pretrained_base, **kwargs) +# if pretrained: +# from .model_store import get_model_file +# device = torch.device(kwargs['local_rank']) +# model.load_state_dict(torch.load(get_model_file('bisenet_%s_%s' % (backbone, acronyms[dataset]), root=root), +# map_location=device)) +# return model +# +# +# def get_bisenet_resnet18_citys(**kwargs): +# return get_bisenet('citys', 'resnet18', **kwargs) + + +# if __name__ == '__main__': +# # img = torch.randn(2, 3, 224, 224) +# # model = BiSeNet(19, backbone='resnet18') +# # print(model.exclusive) +# input = torch.rand(2, 3, 224, 224) +# model = BiSeNet(4, pretrained_base=True) +# # target = torch.zeros(4, 512, 512).cuda() +# # model.eval() +# # print(model) +# loss = model(input) +# print(loss, loss.shape) +# +# # from torchsummary import summary +# # +# # summary(model, (3, 224, 224)) # 打印表格，按顺序输出每层的输出形状和参数 +# import torch +# from thop import profile +# from torchsummary import summary +# +# flop, params = profile(model, input_size=(1, 3, 512, 512)) +# print('flops:{:.3f}G\nparams:{:.3f}M'.format(flop / 1e9, params / 1e6)) + +if __name__ == '__main__': + x = torch.rand(2, 3, 256, 256) + + # model = BiSeNet_MultiOutput(nclass=[2, 2]) # 原始 + # model = BiSeNet_MultiOutput(nclass=[3, 3]) # 改动 + model = BiSeNet_MultiOutput(nclass=[3, 3]) # 改动 + + # print(model) + out = model(x) + print(out[0].size()) + # print() \ No newline at end of file diff --git a/models/common.py b/models/common.py new file mode 100644 index 0000000..f1cc0f8 --- /dev/null +++ b/models/common.py @@ -0,0 +1,404 @@ +# YOLOv5 common modules + +import math +import warnings +from copy import copy +from pathlib import Path + +import numpy as np +import pandas as pd +import requests +import torch +import torch.nn as nn +from PIL import Image +from torch.cuda import amp + +from utils.datasets import letterbox +from utils.general import non_max_suppression, make_divisible, scale_coords, increment_path, xyxy2xywh +from utils.plots import color_list, plot_one_box +from utils.torch_utils import time_synchronized + + +def autopad(k, p=None): # kernel, padding + # Pad to 'same' + if p is None: + p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad + return p + + +def DWConv(c1, c2, k=1, s=1, act=True): + # Depthwise convolution + return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act) + + +class Conv(nn.Module): + # Standard convolution + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups + super(Conv, self).__init__() + self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) + self.bn = nn.BatchNorm2d(c2) + self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) + + def forward(self, x): + return self.act(self.bn(self.conv(x))) + + def fuseforward(self, x): + return self.act(self.conv(x)) + + +class TransformerLayer(nn.Module): + # Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance) + def __init__(self, c, num_heads): + super().__init__() + self.q = nn.Linear(c, c, bias=False) + self.k = nn.Linear(c, c, bias=False) + self.v = nn.Linear(c, c, bias=False) + self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads) + self.fc1 = nn.Linear(c, c, bias=False) + self.fc2 = nn.Linear(c, c, bias=False) + + def forward(self, x): + x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x + x = self.fc2(self.fc1(x)) + x + return x + + +class TransformerBlock(nn.Module): + # Vision Transformer https://arxiv.org/abs/2010.11929 + def __init__(self, c1, c2, num_heads, num_layers): + super().__init__() + self.conv = None + if c1 != c2: + self.conv = Conv(c1, c2) + self.linear = nn.Linear(c2, c2) # learnable position embedding + self.tr = nn.Sequential(*[TransformerLayer(c2, num_heads) for _ in range(num_layers)]) + self.c2 = c2 + + def forward(self, x): + if self.conv is not None: + x = self.conv(x) + b, _, w, h = x.shape + p = x.flatten(2) + p = p.unsqueeze(0) + p = p.transpose(0, 3) + p = p.squeeze(3) + e = self.linear(p) + x = p + e + + x = self.tr(x) + x = x.unsqueeze(3) + x = x.transpose(0, 3) + x = x.reshape(b, self.c2, w, h) + return x + + +class Bottleneck(nn.Module): + # Standard bottleneck + def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion + super(Bottleneck, self).__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_, c2, 3, 1, g=g) + self.add = shortcut and c1 == c2 + + def forward(self, x): + return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) + + +class BottleneckCSP(nn.Module): + # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion + super(BottleneckCSP, self).__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) + self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) + self.cv4 = Conv(2 * c_, c2, 1, 1) + self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) + self.act = nn.LeakyReLU(0.1, inplace=True) + self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) + + def forward(self, x): + y1 = self.cv3(self.m(self.cv1(x))) + y2 = self.cv2(x) + return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1)))) + + +class C3(nn.Module): + # CSP Bottleneck with 3 convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion + super(C3, self).__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c1, c_, 1, 1) + self.cv3 = Conv(2 * c_, c2, 1) # act=FReLU(c2) + self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) + # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)]) + + def forward(self, x): + return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1)) + + +class C3TR(C3): + # C3 module with TransformerBlock() + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) + self.m = TransformerBlock(c_, c_, 4, n) + + +class SPPF(nn.Module): # 添加的 + def __init__(self, c1, c2, k=5): + super().__init__() + c_ = c1 // 2 + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_ * 4, c2, 1, 1) + self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2) + + def forward(self, x): + x = self.cv1(x) + with warnings.catch_warnings(): + warnings.simplefilter('ignore') + y1 = self.m(x) + y2 = self.m(y1) + return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1)) + + +class SPP(nn.Module): + # Spatial pyramid pooling layer used in YOLOv3-SPP + def __init__(self, c1, c2, k=(5, 9, 13)): + super(SPP, self).__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) + self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k]) + + def forward(self, x): + x = self.cv1(x) + return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1)) + + +class Focus(nn.Module): + # Focus wh information into c-space + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups + super(Focus, self).__init__() + self.conv = Conv(c1 * 4, c2, k, s, p, g, act) + # self.contract = Contract(gain=2) + + def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) + return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)) + # return self.conv(self.contract(x)) + + +class Contract(nn.Module): + # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40) + def __init__(self, gain=2): + super().__init__() + self.gain = gain + + def forward(self, x): + N, C, H, W = x.size() # assert (H / s == 0) and (W / s == 0), 'Indivisible gain' + s = self.gain + x = x.view(N, C, H // s, s, W // s, s) # x(1,64,40,2,40,2) + x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # x(1,2,2,64,40,40) + return x.view(N, C * s * s, H // s, W // s) # x(1,256,40,40) + + +class Expand(nn.Module): + # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160) + def __init__(self, gain=2): + super().__init__() + self.gain = gain + + def forward(self, x): + N, C, H, W = x.size() # assert C / s ** 2 == 0, 'Indivisible gain' + s = self.gain + x = x.view(N, s, s, C // s ** 2, H, W) # x(1,2,2,16,80,80) + x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # x(1,16,80,2,80,2) + return x.view(N, C // s ** 2, H * s, W * s) # x(1,16,160,160) + + +class Concat(nn.Module): + # Concatenate a list of tensors along dimension + def __init__(self, dimension=1): + super(Concat, self).__init__() + self.d = dimension + + def forward(self, x): + return torch.cat(x, self.d) + + +class NMS(nn.Module): + # Non-Maximum Suppression (NMS) module + conf = 0.25 # confidence threshold + iou = 0.45 # IoU threshold + classes = None # (optional list) filter by class + + def __init__(self): + super(NMS, self).__init__() + + def forward(self, x): + return non_max_suppression(x[0], conf_thres=self.conf, iou_thres=self.iou, classes=self.classes) + + +class autoShape(nn.Module): + # input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS + conf = 0.25 # NMS confidence threshold + iou = 0.45 # NMS IoU threshold + classes = None # (optional list) filter by class + + def __init__(self, model): + super(autoShape, self).__init__() + self.model = model.eval() + + def autoshape(self): + print('autoShape already enabled, skipping... ') # model already converted to model.autoshape() + return self + + @torch.no_grad() + def forward(self, imgs, size=640, augment=False, profile=False): + # Inference from various sources. For height=640, width=1280, RGB images example inputs are: + # filename: imgs = 'data/samples/zidane.jpg' + # URI: = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/zidane.jpg' + # OpenCV: = cv2.imread('image.jpg')[:,:,::-1] # HWC BGR to RGB x(640,1280,3) + # PIL: = Image.open('image.jpg') # HWC x(640,1280,3) + # numpy: = np.zeros((640,1280,3)) # HWC + # torch: = torch.zeros(16,3,320,640) # BCHW (scaled to size=640, 0-1 values) + # multiple: = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...] # list of images + + t = [time_synchronized()] + p = next(self.model.parameters()) # for device and type + if isinstance(imgs, torch.Tensor): # torch + with amp.autocast(enabled=p.device.type != 'cpu'): + return self.model(imgs.to(p.device).type_as(p), augment, profile) # inference + + # Pre-process + n, imgs = (len(imgs), imgs) if isinstance(imgs, list) else (1, [imgs]) # number of images, list of images + shape0, shape1, files = [], [], [] # image and inference shapes, filenames + for i, im in enumerate(imgs): + f = f'image{i}' # filename + if isinstance(im, str): # filename or uri + im, f = np.asarray(Image.open(requests.get(im, stream=True).raw if im.startswith('http') else im)), im + elif isinstance(im, Image.Image): # PIL Image + im, f = np.asarray(im), getattr(im, 'filename', f) or f + files.append(Path(f).with_suffix('.jpg').name) + if im.shape[0] < 5: # image in CHW + im = im.transpose((1, 2, 0)) # reverse dataloader .transpose(2, 0, 1) + im = im[:, :, :3] if im.ndim == 3 else np.tile(im[:, :, None], 3) # enforce 3ch input + s = im.shape[:2] # HWC + shape0.append(s) # image shape + g = (size / max(s)) # gain + shape1.append([y * g for y in s]) + imgs[i] = im # update + shape1 = [make_divisible(x, int(self.stride.max())) for x in np.stack(shape1, 0).max(0)] # inference shape + x = [letterbox(im, new_shape=shape1, auto=False)[0] for im in imgs] # pad + x = np.stack(x, 0) if n > 1 else x[0][None] # stack + x = np.ascontiguousarray(x.transpose((0, 3, 1, 2))) # BHWC to BCHW + x = torch.from_numpy(x).to(p.device).type_as(p) / 255. # uint8 to fp16/32 + t.append(time_synchronized()) + + with amp.autocast(enabled=p.device.type != 'cpu'): + # Inference + y = self.model(x, augment, profile)[0] # forward + t.append(time_synchronized()) + + # Post-process + y = non_max_suppression(y, conf_thres=self.conf, iou_thres=self.iou, classes=self.classes) # NMS + for i in range(n): + scale_coords(shape1, y[i][:, :4], shape0[i]) + + t.append(time_synchronized()) + return Detections(imgs, y, files, t, self.names, x.shape) + + +class Detections: + # detections class for YOLOv5 inference results + def __init__(self, imgs, pred, files, times=None, names=None, shape=None): + super(Detections, self).__init__() + d = pred[0].device # device + gn = [torch.tensor([*[im.shape[i] for i in [1, 0, 1, 0]], 1., 1.], device=d) for im in imgs] # normalizations + self.imgs = imgs # list of images as numpy arrays + self.pred = pred # list of tensors pred[0] = (xyxy, conf, cls) + self.names = names # class names + self.files = files # image filenames + self.xyxy = pred # xyxy pixels + self.xywh = [xyxy2xywh(x) for x in pred] # xywh pixels + self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)] # xyxy normalized + self.xywhn = [x / g for x, g in zip(self.xywh, gn)] # xywh normalized + self.n = len(self.pred) # number of images (batch size) + self.t = tuple((times[i + 1] - times[i]) * 1000 / self.n for i in range(3)) # timestamps (ms) + self.s = shape # inference BCHW shape + + def display(self, pprint=False, show=False, save=False, render=False, save_dir=''): + colors = color_list() + for i, (img, pred) in enumerate(zip(self.imgs, self.pred)): + str = f'image {i + 1}/{len(self.pred)}: {img.shape[0]}x{img.shape[1]} ' + if pred is not None: + for c in pred[:, -1].unique(): + n = (pred[:, -1] == c).sum() # detections per class + str += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, " # add to string + if show or save or render: + for *box, conf, cls in pred: # xyxy, confidence, class + label = f'{self.names[int(cls)]} {conf:.2f}' + plot_one_box(box, img, label=label, color=colors[int(cls) % 10]) + img = Image.fromarray(img.astype(np.uint8)) if isinstance(img, np.ndarray) else img # from np + if pprint: + print(str.rstrip(', ')) + if show: + img.show(self.files[i]) # show + if save: + f = self.files[i] + img.save(Path(save_dir) / f) # save + print(f"{'Saved' * (i == 0)} {f}", end=',' if i < self.n - 1 else f' to {save_dir}\n') + if render: + self.imgs[i] = np.asarray(img) + + def print(self): + self.display(pprint=True) # print results + print(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {tuple(self.s)}' % self.t) + + def show(self): + self.display(show=True) # show results + + def save(self, save_dir='runs/hub/exp'): + save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/hub/exp') # increment save_dir + Path(save_dir).mkdir(parents=True, exist_ok=True) + self.display(save=True, save_dir=save_dir) # save results + + def render(self): + self.display(render=True) # render results + return self.imgs + + def pandas(self): + # return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0]) + new = copy(self) # return copy + ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name' # xyxy columns + cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name' # xywh columns + for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]): + a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)] # update + setattr(new, k, [pd.DataFrame(x, columns=c) for x in a]) + return new + + def tolist(self): + # return a list of Detections objects, i.e. 'for result in results.tolist():' + x = [Detections([self.imgs[i]], [self.pred[i]], self.names, self.s) for i in range(self.n)] + for d in x: + for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']: + setattr(d, k, getattr(d, k)[0]) # pop out of list + return x + + def __len__(self): + return self.n + + +class Classify(nn.Module): + # Classification head, i.e. x(b,c1,20,20) to x(b,c2) + def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups + super(Classify, self).__init__() + self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1) + self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g) # to x(b,c2,1,1) + self.flat = nn.Flatten() + + def forward(self, x): + z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1) # cat if list + return self.flat(self.conv(z)) # flatten to x(b,c2) diff --git a/models/experimental.py b/models/experimental.py new file mode 100644 index 0000000..548353c --- /dev/null +++ b/models/experimental.py @@ -0,0 +1,134 @@ +# YOLOv5 experimental modules + +import numpy as np +import torch +import torch.nn as nn + +from models.common import Conv, DWConv +from utils.google_utils import attempt_download + + +class CrossConv(nn.Module): + # Cross Convolution Downsample + def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False): + # ch_in, ch_out, kernel, stride, groups, expansion, shortcut + super(CrossConv, self).__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, (1, k), (1, s)) + self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g) + self.add = shortcut and c1 == c2 + + def forward(self, x): + return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) + + +class Sum(nn.Module): + # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070 + def __init__(self, n, weight=False): # n: number of inputs + super(Sum, self).__init__() + self.weight = weight # apply weights boolean + self.iter = range(n - 1) # iter object + if weight: + self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True) # layer weights + + def forward(self, x): + y = x[0] # no weight + if self.weight: + w = torch.sigmoid(self.w) * 2 + for i in self.iter: + y = y + x[i + 1] * w[i] + else: + for i in self.iter: + y = y + x[i + 1] + return y + + +class GhostConv(nn.Module): + # Ghost Convolution https://github.com/huawei-noah/ghostnet + def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups + super(GhostConv, self).__init__() + c_ = c2 // 2 # hidden channels + self.cv1 = Conv(c1, c_, k, s, None, g, act) + self.cv2 = Conv(c_, c_, 5, 1, None, c_, act) + + def forward(self, x): + y = self.cv1(x) + return torch.cat([y, self.cv2(y)], 1) + + +class GhostBottleneck(nn.Module): + # Ghost Bottleneck https://github.com/huawei-noah/ghostnet + def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride + super(GhostBottleneck, self).__init__() + c_ = c2 // 2 + self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1), # pw + DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw + GhostConv(c_, c2, 1, 1, act=False)) # pw-linear + self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False), + Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity() + + def forward(self, x): + return self.conv(x) + self.shortcut(x) + + +class MixConv2d(nn.Module): + # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595 + def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): + super(MixConv2d, self).__init__() + groups = len(k) + if equal_ch: # equal c_ per group + i = torch.linspace(0, groups - 1E-6, c2).floor() # c2 indices + c_ = [(i == g).sum() for g in range(groups)] # intermediate channels + else: # equal weight.numel() per group + b = [c2] + [0] * groups + a = np.eye(groups + 1, groups, k=-1) + a -= np.roll(a, 1, axis=1) + a *= np.array(k) ** 2 + a[0] = 1 + c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b + + self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)]) + self.bn = nn.BatchNorm2d(c2) + self.act = nn.LeakyReLU(0.1, inplace=True) + + def forward(self, x): + return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1))) + + +class Ensemble(nn.ModuleList): + # Ensemble of models + def __init__(self): + super(Ensemble, self).__init__() + + def forward(self, x, augment=False): + y = [] + for module in self: + y.append(module(x, augment)[0]) + # y = torch.stack(y).max(0)[0] # max ensemble + # y = torch.stack(y).mean(0) # mean ensemble + y = torch.cat(y, 1) # nms ensemble + return y, None # inference, train output + + +def attempt_load(weights, map_location=None): + # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a + model = Ensemble() + for w in weights if isinstance(weights, list) else [weights]: + attempt_download(w) + ckpt = torch.load(w, map_location=map_location) # load + model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval()) # FP32 model + + # Compatibility updates + for m in model.modules(): + if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]: + m.inplace = True # pytorch 1.7.0 compatibility + elif type(m) is Conv: + m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility + + if len(model) == 1: + return model[-1] # return model + else: + print('Ensemble created with %s\n' % weights) + for k in ['names', 'stride']: + setattr(model, k, getattr(model[-1], k)) + return model # return ensemble diff --git a/models/export.py b/models/export.py new file mode 100644 index 0000000..11e60c7 --- /dev/null +++ b/models/export.py @@ -0,0 +1,104 @@ +"""Exports a YOLOv5 *.pt model to ONNX and TorchScript formats + +Usage: + $ export PYTHONPATH="$PWD" && python models/export.py --weights ./weights/yolov5s.pt --img 640 --batch 1 +""" + +import argparse +import sys +import time + +sys.path.append('./') # to run '$ python *.py' files in subdirectories + +import torch +import torch.nn as nn + +import models +from models.experimental import attempt_load +from utils.activations import Hardswish, SiLU +from utils.general import set_logging, check_img_size +from utils.torch_utils import select_device + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default='./yolov5s.pt', help='weights path') # from yolov5/models/ + parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size') # height, width + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--dynamic', action='store_true', help='dynamic ONNX axes') + parser.add_argument('--grid', action='store_true', help='export Detect() layer grid') + parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + opt = parser.parse_args() + opt.img_size *= 2 if len(opt.img_size) == 1 else 1 # expand + print(opt) + set_logging() + t = time.time() + + # Load PyTorch model + device = select_device(opt.device) + model = attempt_load(opt.weights, map_location=device) # load FP32 model + labels = model.names + + # Checks + gs = int(max(model.stride)) # grid size (max stride) + opt.img_size = [check_img_size(x, gs) for x in opt.img_size] # verify img_size are gs-multiples + + # Input + img = torch.zeros(opt.batch_size, 3, *opt.img_size).to(device) # image size(1,3,320,192) iDetection + + # Update model + for k, m in model.named_modules(): + m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility + if isinstance(m, models.common.Conv): # assign export-friendly activations + if isinstance(m.act, nn.Hardswish): + m.act = Hardswish() + elif isinstance(m.act, nn.SiLU): + m.act = SiLU() + # elif isinstance(m, models.yolo.Detect): + # m.forward = m.forward_export # assign forward (optional) + model.model[-1].export = not opt.grid # set Detect() layer grid export + y = model(img) # dry run + + # TorchScript export + try: + print('\nStarting TorchScript export with torch %s...' % torch.__version__) + f = opt.weights.replace('.pt', '.torchscript.pt') # filename + ts = torch.jit.trace(model, img) + ts.save(f) + print('TorchScript export success, saved as %s' % f) + except Exception as e: + print('TorchScript export failure: %s' % e) + + # ONNX export + try: + import onnx + + print('\nStarting ONNX export with onnx %s...' % onnx.__version__) + f = opt.weights.replace('.pt', '.onnx') # filename + torch.onnx.export(model, img, f, verbose=False, opset_version=12, input_names=['images'], + output_names=['classes', 'boxes'] if y is None else ['output'], + dynamic_axes={'images': {0: 'batch', 2: 'height', 3: 'width'}, # size(1,3,640,640) + 'output': {0: 'batch', 2: 'y', 3: 'x'}} if opt.dynamic else None) + + # Checks + onnx_model = onnx.load(f) # load onnx model + onnx.checker.check_model(onnx_model) # check onnx model + # print(onnx.helper.printable_graph(onnx_model.graph)) # print a human readable model + print('ONNX export success, saved as %s' % f) + except Exception as e: + print('ONNX export failure: %s' % e) + + # CoreML export + try: + import coremltools as ct + + print('\nStarting CoreML export with coremltools %s...' % ct.__version__) + # convert model from torchscript and apply pixel scaling as per detect.py + model = ct.convert(ts, inputs=[ct.ImageType(name='image', shape=img.shape, scale=1 / 255.0, bias=[0, 0, 0])]) + f = opt.weights.replace('.pt', '.mlmodel') # filename + model.save(f) + print('CoreML export success, saved as %s' % f) + except Exception as e: + print('CoreML export failure: %s' % e) + + # Finish + print('\nExport complete (%.2fs). Visualize with https://github.com/lutzroeder/netron.' % (time.time() - t)) diff --git a/models/hub/anchors.yaml b/models/hub/anchors.yaml new file mode 100644 index 0000000..a07a4dc --- /dev/null +++ b/models/hub/anchors.yaml @@ -0,0 +1,58 @@ +# Default YOLOv5 anchors for COCO data + + +# P5 ------------------------------------------------------------------------------------------------------------------- +# P5-640: +anchors_p5_640: + - [ 10,13, 16,30, 33,23 ] # P3/8 + - [ 30,61, 62,45, 59,119 ] # P4/16 + - [ 116,90, 156,198, 373,326 ] # P5/32 + + +# P6 ------------------------------------------------------------------------------------------------------------------- +# P6-640: thr=0.25: 0.9964 BPR, 5.54 anchors past thr, n=12, img_size=640, metric_all=0.281/0.716-mean/best, past_thr=0.469-mean: 9,11, 21,19, 17,41, 43,32, 39,70, 86,64, 65,131, 134,130, 120,265, 282,180, 247,354, 512,387 +anchors_p6_640: + - [ 9,11, 21,19, 17,41 ] # P3/8 + - [ 43,32, 39,70, 86,64 ] # P4/16 + - [ 65,131, 134,130, 120,265 ] # P5/32 + - [ 282,180, 247,354, 512,387 ] # P6/64 + +# P6-1280: thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1280, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 19,27, 44,40, 38,94, 96,68, 86,152, 180,137, 140,301, 303,264, 238,542, 436,615, 739,380, 925,792 +anchors_p6_1280: + - [ 19,27, 44,40, 38,94 ] # P3/8 + - [ 96,68, 86,152, 180,137 ] # P4/16 + - [ 140,301, 303,264, 238,542 ] # P5/32 + - [ 436,615, 739,380, 925,792 ] # P6/64 + +# P6-1920: thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1920, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 28,41, 67,59, 57,141, 144,103, 129,227, 270,205, 209,452, 455,396, 358,812, 653,922, 1109,570, 1387,1187 +anchors_p6_1920: + - [ 28,41, 67,59, 57,141 ] # P3/8 + - [ 144,103, 129,227, 270,205 ] # P4/16 + - [ 209,452, 455,396, 358,812 ] # P5/32 + - [ 653,922, 1109,570, 1387,1187 ] # P6/64 + + +# P7 ------------------------------------------------------------------------------------------------------------------- +# P7-640: thr=0.25: 0.9962 BPR, 6.76 anchors past thr, n=15, img_size=640, metric_all=0.275/0.733-mean/best, past_thr=0.466-mean: 11,11, 13,30, 29,20, 30,46, 61,38, 39,92, 78,80, 146,66, 79,163, 149,150, 321,143, 157,303, 257,402, 359,290, 524,372 +anchors_p7_640: + - [ 11,11, 13,30, 29,20 ] # P3/8 + - [ 30,46, 61,38, 39,92 ] # P4/16 + - [ 78,80, 146,66, 79,163 ] # P5/32 + - [ 149,150, 321,143, 157,303 ] # P6/64 + - [ 257,402, 359,290, 524,372 ] # P7/128 + +# P7-1280: thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1280, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 19,22, 54,36, 32,77, 70,83, 138,71, 75,173, 165,159, 148,334, 375,151, 334,317, 251,626, 499,474, 750,326, 534,814, 1079,818 +anchors_p7_1280: + - [ 19,22, 54,36, 32,77 ] # P3/8 + - [ 70,83, 138,71, 75,173 ] # P4/16 + - [ 165,159, 148,334, 375,151 ] # P5/32 + - [ 334,317, 251,626, 499,474 ] # P6/64 + - [ 750,326, 534,814, 1079,818 ] # P7/128 + +# P7-1920: thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1920, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 29,34, 81,55, 47,115, 105,124, 207,107, 113,259, 247,238, 222,500, 563,227, 501,476, 376,939, 749,711, 1126,489, 801,1222, 1618,1227 +anchors_p7_1920: + - [ 29,34, 81,55, 47,115 ] # P3/8 + - [ 105,124, 207,107, 113,259 ] # P4/16 + - [ 247,238, 222,500, 563,227 ] # P5/32 + - [ 501,476, 376,939, 749,711 ] # P6/64 + - [ 1126,489, 801,1222, 1618,1227 ] # P7/128 diff --git a/models/hub/yolov3-spp.yaml b/models/hub/yolov3-spp.yaml new file mode 100644 index 0000000..38dcc44 --- /dev/null +++ b/models/hub/yolov3-spp.yaml @@ -0,0 +1,51 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# darknet53 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [32, 3, 1]], # 0 + [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 + [-1, 1, Bottleneck, [64]], + [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 + [-1, 8, Bottleneck, [256]], + [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ] + +# YOLOv3-SPP head +head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, SPP, [512, [5, 9, 13]]], + [-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, Conv, [1024, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 22 (P4/16-medium) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Bottleneck, [256, False]], + [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small) + + [[27, 22, 15], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/models/hub/yolov3-tiny.yaml b/models/hub/yolov3-tiny.yaml new file mode 100644 index 0000000..ff7638c --- /dev/null +++ b/models/hub/yolov3-tiny.yaml @@ -0,0 +1,41 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: + - [10,14, 23,27, 37,58] # P4/16 + - [81,82, 135,169, 344,319] # P5/32 + +# YOLOv3-tiny backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [16, 3, 1]], # 0 + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 1-P1/2 + [-1, 1, Conv, [32, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 3-P2/4 + [-1, 1, Conv, [64, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 5-P3/8 + [-1, 1, Conv, [128, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 7-P4/16 + [-1, 1, Conv, [256, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 9-P5/32 + [-1, 1, Conv, [512, 3, 1]], + [-1, 1, nn.ZeroPad2d, [[0, 1, 0, 1]]], # 11 + [-1, 1, nn.MaxPool2d, [2, 1, 0]], # 12 + ] + +# YOLOv3-tiny head +head: + [[-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Conv, [256, 3, 1]], # 19 (P4/16-medium) + + [[19, 15], 1, Detect, [nc, anchors]], # Detect(P4, P5) + ] diff --git a/models/hub/yolov3.yaml b/models/hub/yolov3.yaml new file mode 100644 index 0000000..f2e7613 --- /dev/null +++ b/models/hub/yolov3.yaml @@ -0,0 +1,51 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# darknet53 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [32, 3, 1]], # 0 + [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 + [-1, 1, Bottleneck, [64]], + [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 + [-1, 8, Bottleneck, [256]], + [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ] + +# YOLOv3 head +head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, Conv, [512, [1, 1]]], + [-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, Conv, [1024, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 22 (P4/16-medium) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Bottleneck, [256, False]], + [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small) + + [[27, 22, 15], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/models/hub/yolov5-fpn.yaml b/models/hub/yolov5-fpn.yaml new file mode 100644 index 0000000..e772bff --- /dev/null +++ b/models/hub/yolov5-fpn.yaml @@ -0,0 +1,42 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Focus, [64, 3]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, Bottleneck, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 9, BottleneckCSP, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, BottleneckCSP, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 1, SPP, [1024, [5, 9, 13]]], + [-1, 6, BottleneckCSP, [1024]], # 9 + ] + +# YOLOv5 FPN head +head: + [[-1, 3, BottleneckCSP, [1024, False]], # 10 (P5/32-large) + + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Conv, [512, 1, 1]], + [-1, 3, BottleneckCSP, [512, False]], # 14 (P4/16-medium) + + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Conv, [256, 1, 1]], + [-1, 3, BottleneckCSP, [256, False]], # 18 (P3/8-small) + + [[18, 14, 10], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/models/hub/yolov5-p2.yaml b/models/hub/yolov5-p2.yaml new file mode 100644 index 0000000..0633a90 --- /dev/null +++ b/models/hub/yolov5-p2.yaml @@ -0,0 +1,54 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: 3 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 9, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 7-P5/32 + [ -1, 1, SPP, [ 1024, [ 5, 9, 13 ] ] ], + [ -1, 3, C3, [ 1024, False ] ], # 9 + ] + +# YOLOv5 head +head: + [ [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 13 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 17 (P3/8-small) + + [ -1, 1, Conv, [ 128, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 2 ], 1, Concat, [ 1 ] ], # cat backbone P2 + [ -1, 1, C3, [ 128, False ] ], # 21 (P2/4-xsmall) + + [ -1, 1, Conv, [ 128, 3, 2 ] ], + [ [ -1, 18 ], 1, Concat, [ 1 ] ], # cat head P3 + [ -1, 3, C3, [ 256, False ] ], # 24 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 14 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 27 (P4/16-medium) + + [ -1, 1, Conv, [ 512, 3, 2 ] ], + [ [ -1, 10 ], 1, Concat, [ 1 ] ], # cat head P5 + [ -1, 3, C3, [ 1024, False ] ], # 30 (P5/32-large) + + [ [ 24, 27, 30 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5) + ] diff --git a/models/hub/yolov5-p6.yaml b/models/hub/yolov5-p6.yaml new file mode 100644 index 0000000..3728a11 --- /dev/null +++ b/models/hub/yolov5-p6.yaml @@ -0,0 +1,56 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: 3 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 9, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 + [ -1, 3, C3, [ 768 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 + [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], + [ -1, 3, C3, [ 1024, False ] ], # 11 + ] + +# YOLOv5 head +head: + [ [ -1, 1, Conv, [ 768, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 + [ -1, 3, C3, [ 768, False ] ], # 15 + + [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 19 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) + + [ -1, 1, Conv, [ 512, 3, 2 ] ], + [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 + [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) + + [ -1, 1, Conv, [ 768, 3, 2 ] ], + [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 + [ -1, 3, C3, [ 1024, False ] ], # 32 (P5/64-xlarge) + + [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) + ] diff --git a/models/hub/yolov5-p7.yaml b/models/hub/yolov5-p7.yaml new file mode 100644 index 0000000..ca8f849 --- /dev/null +++ b/models/hub/yolov5-p7.yaml @@ -0,0 +1,67 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: 3 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 9, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 + [ -1, 3, C3, [ 768 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 + [ -1, 3, C3, [ 1024 ] ], + [ -1, 1, Conv, [ 1280, 3, 2 ] ], # 11-P7/128 + [ -1, 1, SPP, [ 1280, [ 3, 5 ] ] ], + [ -1, 3, C3, [ 1280, False ] ], # 13 + ] + +# YOLOv5 head +head: + [ [ -1, 1, Conv, [ 1024, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 10 ], 1, Concat, [ 1 ] ], # cat backbone P6 + [ -1, 3, C3, [ 1024, False ] ], # 17 + + [ -1, 1, Conv, [ 768, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 + [ -1, 3, C3, [ 768, False ] ], # 21 + + [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 25 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 29 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 26 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 32 (P4/16-medium) + + [ -1, 1, Conv, [ 512, 3, 2 ] ], + [ [ -1, 22 ], 1, Concat, [ 1 ] ], # cat head P5 + [ -1, 3, C3, [ 768, False ] ], # 35 (P5/32-large) + + [ -1, 1, Conv, [ 768, 3, 2 ] ], + [ [ -1, 18 ], 1, Concat, [ 1 ] ], # cat head P6 + [ -1, 3, C3, [ 1024, False ] ], # 38 (P6/64-xlarge) + + [ -1, 1, Conv, [ 1024, 3, 2 ] ], + [ [ -1, 14 ], 1, Concat, [ 1 ] ], # cat head P7 + [ -1, 3, C3, [ 1280, False ] ], # 41 (P7/128-xxlarge) + + [ [ 29, 32, 35, 38, 41 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6, P7) + ] diff --git a/models/hub/yolov5-panet.yaml b/models/hub/yolov5-panet.yaml new file mode 100644 index 0000000..340f95a --- /dev/null +++ b/models/hub/yolov5-panet.yaml @@ -0,0 +1,48 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Focus, [64, 3]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, BottleneckCSP, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 9, BottleneckCSP, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, BottleneckCSP, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 1, SPP, [1024, [5, 9, 13]]], + [-1, 3, BottleneckCSP, [1024, False]], # 9 + ] + +# YOLOv5 PANet head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, BottleneckCSP, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, BottleneckCSP, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, BottleneckCSP, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, BottleneckCSP, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/models/hub/yolov5l6.yaml b/models/hub/yolov5l6.yaml new file mode 100644 index 0000000..11298b0 --- /dev/null +++ b/models/hub/yolov5l6.yaml @@ -0,0 +1,60 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: + - [ 19,27, 44,40, 38,94 ] # P3/8 + - [ 96,68, 86,152, 180,137 ] # P4/16 + - [ 140,301, 303,264, 238,542 ] # P5/32 + - [ 436,615, 739,380, 925,792 ] # P6/64 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 9, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 + [ -1, 3, C3, [ 768 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 + [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], + [ -1, 3, C3, [ 1024, False ] ], # 11 + ] + +# YOLOv5 head +head: + [ [ -1, 1, Conv, [ 768, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 + [ -1, 3, C3, [ 768, False ] ], # 15 + + [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 19 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) + + [ -1, 1, Conv, [ 512, 3, 2 ] ], + [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 + [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) + + [ -1, 1, Conv, [ 768, 3, 2 ] ], + [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 + [ -1, 3, C3, [ 1024, False ] ], # 32 (P6/64-xlarge) + + [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) + ] diff --git a/models/hub/yolov5m6.yaml b/models/hub/yolov5m6.yaml new file mode 100644 index 0000000..48afc86 --- /dev/null +++ b/models/hub/yolov5m6.yaml @@ -0,0 +1,60 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 0.67 # model depth multiple +width_multiple: 0.75 # layer channel multiple + +# anchors +anchors: + - [ 19,27, 44,40, 38,94 ] # P3/8 + - [ 96,68, 86,152, 180,137 ] # P4/16 + - [ 140,301, 303,264, 238,542 ] # P5/32 + - [ 436,615, 739,380, 925,792 ] # P6/64 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 9, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 + [ -1, 3, C3, [ 768 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 + [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], + [ -1, 3, C3, [ 1024, False ] ], # 11 + ] + +# YOLOv5 head +head: + [ [ -1, 1, Conv, [ 768, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 + [ -1, 3, C3, [ 768, False ] ], # 15 + + [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 19 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) + + [ -1, 1, Conv, [ 512, 3, 2 ] ], + [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 + [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) + + [ -1, 1, Conv, [ 768, 3, 2 ] ], + [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 + [ -1, 3, C3, [ 1024, False ] ], # 32 (P6/64-xlarge) + + [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) + ] diff --git a/models/hub/yolov5s-transformer.yaml b/models/hub/yolov5s-transformer.yaml new file mode 100644 index 0000000..f2d6667 --- /dev/null +++ b/models/hub/yolov5s-transformer.yaml @@ -0,0 +1,48 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Focus, [64, 3]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 9, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 1, SPP, [1024, [5, 9, 13]]], + [-1, 3, C3TR, [1024, False]], # 9 <-------- C3TR() Transformer module + ] + +# YOLOv5 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/models/hub/yolov5s6.yaml b/models/hub/yolov5s6.yaml new file mode 100644 index 0000000..1df577a --- /dev/null +++ b/models/hub/yolov5s6.yaml @@ -0,0 +1,60 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple + +# anchors +anchors: + - [ 19,27, 44,40, 38,94 ] # P3/8 + - [ 96,68, 86,152, 180,137 ] # P4/16 + - [ 140,301, 303,264, 238,542 ] # P5/32 + - [ 436,615, 739,380, 925,792 ] # P6/64 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 9, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 + [ -1, 3, C3, [ 768 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 + [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], + [ -1, 3, C3, [ 1024, False ] ], # 11 + ] + +# YOLOv5 head +head: + [ [ -1, 1, Conv, [ 768, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 + [ -1, 3, C3, [ 768, False ] ], # 15 + + [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 19 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) + + [ -1, 1, Conv, [ 512, 3, 2 ] ], + [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 + [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) + + [ -1, 1, Conv, [ 768, 3, 2 ] ], + [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 + [ -1, 3, C3, [ 1024, False ] ], # 32 (P6/64-xlarge) + + [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) + ] diff --git a/models/hub/yolov5x6.yaml b/models/hub/yolov5x6.yaml new file mode 100644 index 0000000..5ebc021 --- /dev/null +++ b/models/hub/yolov5x6.yaml @@ -0,0 +1,60 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 1.33 # model depth multiple +width_multiple: 1.25 # layer channel multiple + +# anchors +anchors: + - [ 19,27, 44,40, 38,94 ] # P3/8 + - [ 96,68, 86,152, 180,137 ] # P4/16 + - [ 140,301, 303,264, 238,542 ] # P5/32 + - [ 436,615, 739,380, 925,792 ] # P6/64 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 9, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 + [ -1, 3, C3, [ 768 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 + [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], + [ -1, 3, C3, [ 1024, False ] ], # 11 + ] + +# YOLOv5 head +head: + [ [ -1, 1, Conv, [ 768, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 + [ -1, 3, C3, [ 768, False ] ], # 15 + + [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 19 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) + + [ -1, 1, Conv, [ 512, 3, 2 ] ], + [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 + [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) + + [ -1, 1, Conv, [ 768, 3, 2 ] ], + [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 + [ -1, 3, C3, [ 1024, False ] ], # 32 (P6/64-xlarge) + + [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) + ] diff --git a/models/model_stages.py b/models/model_stages.py new file mode 100644 index 0000000..1be35f7 --- /dev/null +++ b/models/model_stages.py @@ -0,0 +1,334 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- + + +import torch +import torch.nn as nn +import torch.nn.functional as F +import torchvision + +from nets.stdcnet import STDCNet1446, STDCNet813 +from modules.bn import InPlaceABNSync as BatchNorm2d +# BatchNorm2d = nn.BatchNorm2d + +class ConvBNReLU(nn.Module): + def __init__(self, in_chan, out_chan, ks=3, stride=1, padding=1, *args, **kwargs): + super(ConvBNReLU, self).__init__() + self.conv = nn.Conv2d(in_chan, + out_chan, + kernel_size = ks, + stride = stride, + padding = padding, + bias = False) + # self.bn = BatchNorm2d(out_chan) + # self.bn = BatchNorm2d(out_chan, activation='none') + self.bn = nn.BatchNorm2d(out_chan) + self.relu = nn.ReLU() + self.init_weight() + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + +class BiSeNetOutput(nn.Module): + def __init__(self, in_chan, mid_chan, n_classes, *args, **kwargs): + super(BiSeNetOutput, self).__init__() + self.conv = ConvBNReLU(in_chan, mid_chan, ks=3, stride=1, padding=1) + self.conv_out = nn.Conv2d(mid_chan, n_classes, kernel_size=1, bias=False) + self.init_weight() + + def forward(self, x): + x = self.conv(x) + x = self.conv_out(x) + return x + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params = [], [] + for name, module in self.named_modules(): + if isinstance(module, (nn.Linear, nn.Conv2d)): + wd_params.append(module.weight) + if not module.bias is None: + nowd_params.append(module.bias) + elif isinstance(module, nn.BatchNorm2d):######################1 + nowd_params += list(module.parameters()) + return wd_params, nowd_params + + +class AttentionRefinementModule(nn.Module): + def __init__(self, in_chan, out_chan, *args, **kwargs): + super(AttentionRefinementModule, self).__init__() + self.conv = ConvBNReLU(in_chan, out_chan, ks=3, stride=1, padding=1) + self.conv_atten = nn.Conv2d(out_chan, out_chan, kernel_size= 1, bias=False) + # self.bn_atten = nn.BatchNorm2d(out_chan) + # self.bn_atten = BatchNorm2d(out_chan, activation='none') + self.bn_atten = nn.BatchNorm2d(out_chan)########################2 + + self.sigmoid_atten = nn.Sigmoid() + self.init_weight() + + def forward(self, x): + feat = self.conv(x) + atten = F.avg_pool2d(feat, feat.size()[2:]) + atten = self.conv_atten(atten) + atten = self.bn_atten(atten) + atten = self.sigmoid_atten(atten) + out = torch.mul(feat, atten) + return out + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + +class ContextPath(nn.Module): + def __init__(self, backbone='CatNetSmall', pretrain_model='', use_conv_last=False, *args, **kwargs): + super(ContextPath, self).__init__() + + self.backbone_name = backbone + if backbone == 'STDCNet1446': + self.backbone = STDCNet1446(pretrain_model=pretrain_model, use_conv_last=use_conv_last) + self.arm16 = AttentionRefinementModule(512, 128) + inplanes = 1024 + if use_conv_last: + inplanes = 1024 + self.arm32 = AttentionRefinementModule(inplanes, 128) + self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1) + self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1) + self.conv_avg = ConvBNReLU(inplanes, 128, ks=1, stride=1, padding=0) + + elif backbone == 'STDCNet813': + self.backbone = STDCNet813(pretrain_model=pretrain_model, use_conv_last=use_conv_last) + self.arm16 = AttentionRefinementModule(512, 128) + inplanes = 1024 + if use_conv_last: + inplanes = 1024 + self.arm32 = AttentionRefinementModule(inplanes, 128) + self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1) + self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1) + self.conv_avg = ConvBNReLU(inplanes, 128, ks=1, stride=1, padding=0) + else: + print("backbone is not in backbone lists") + exit(0) + + self.init_weight() + + def forward(self, x): + H0, W0 = x.size()[2:] + + feat2, feat4, feat8, feat16, feat32 = self.backbone(x) + H8, W8 = feat8.size()[2:] + H16, W16 = feat16.size()[2:] + H32, W32 = feat32.size()[2:] + + avg = F.avg_pool2d(feat32, feat32.size()[2:]) + + avg = self.conv_avg(avg) + avg_up = F.interpolate(avg, (H32, W32), mode='nearest') + + feat32_arm = self.arm32(feat32) + feat32_sum = feat32_arm + avg_up + feat32_up = F.interpolate(feat32_sum, (H16, W16), mode='nearest') + feat32_up = self.conv_head32(feat32_up) + + feat16_arm = self.arm16(feat16) + feat16_sum = feat16_arm + feat32_up + feat16_up = F.interpolate(feat16_sum, (H8, W8), mode='nearest') + feat16_up = self.conv_head16(feat16_up) + + return feat2, feat4, feat8, feat16, feat16_up, feat32_up # x8, x16 + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params = [], [] + for name, module in self.named_modules(): + if isinstance(module, (nn.Linear, nn.Conv2d)): + wd_params.append(module.weight) + if not module.bias is None: + nowd_params.append(module.bias) + elif isinstance(module, nn.BatchNorm2d):#################3 + nowd_params += list(module.parameters()) + return wd_params, nowd_params + + +class FeatureFusionModule(nn.Module): + def __init__(self, in_chan, out_chan, *args, **kwargs): + super(FeatureFusionModule, self).__init__() + self.convblk = ConvBNReLU(in_chan, out_chan, ks=1, stride=1, padding=0) + self.conv1 = nn.Conv2d(out_chan, + out_chan//4, + kernel_size = 1, + stride = 1, + padding = 0, + bias = False) + self.conv2 = nn.Conv2d(out_chan//4, + out_chan, + kernel_size = 1, + stride = 1, + padding = 0, + bias = False) + self.relu = nn.ReLU(inplace=True) + self.sigmoid = nn.Sigmoid() + self.init_weight() + + def forward(self, fsp, fcp): + fcat = torch.cat([fsp, fcp], dim=1) + feat = self.convblk(fcat) + atten = F.avg_pool2d(feat, feat.size()[2:]) + atten = self.conv1(atten) + atten = self.relu(atten) + atten = self.conv2(atten) + atten = self.sigmoid(atten) + feat_atten = torch.mul(feat, atten) + feat_out = feat_atten + feat + return feat_out + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params = [], [] + for name, module in self.named_modules(): + if isinstance(module, (nn.Linear, nn.Conv2d)): + wd_params.append(module.weight) + if not module.bias is None: + nowd_params.append(module.bias) + elif isinstance(module, nn.BatchNorm2d):##################4 + nowd_params += list(module.parameters()) + return wd_params, nowd_params + + +class BiSeNet(nn.Module): + def __init__(self, backbone, n_classes, pretrain_model='', use_boundary_2=False, use_boundary_4=False, use_boundary_8=False, use_boundary_16=False, use_conv_last=False, heat_map=False, *args, **kwargs): + super(BiSeNet, self).__init__() + + self.use_boundary_2 = use_boundary_2 + self.use_boundary_4 = use_boundary_4 + self.use_boundary_8 = use_boundary_8 + self.use_boundary_16 = use_boundary_16 + # self.heat_map = heat_map + self.cp = ContextPath(backbone, pretrain_model, use_conv_last=use_conv_last) + + if backbone == 'STDCNet1446': + conv_out_inplanes = 128 + sp2_inplanes = 32 + sp4_inplanes = 64 + sp8_inplanes = 256 + sp16_inplanes = 512 + inplane = sp8_inplanes + conv_out_inplanes + + elif backbone == 'STDCNet813': + conv_out_inplanes = 128 + sp2_inplanes = 32 + sp4_inplanes = 64 + sp8_inplanes = 256 + sp16_inplanes = 512 + inplane = sp8_inplanes + conv_out_inplanes + + else: + print("backbone is not in backbone lists") + exit(0) + + self.ffm = FeatureFusionModule(inplane, 256) + self.conv_out = BiSeNetOutput(256, 256, n_classes) + self.conv_out16 = BiSeNetOutput(conv_out_inplanes, 64, n_classes) + self.conv_out32 = BiSeNetOutput(conv_out_inplanes, 64, n_classes) + + self.conv_out_sp16 = BiSeNetOutput(sp16_inplanes, 64, 1) + + self.conv_out_sp8 = BiSeNetOutput(sp8_inplanes, 64, 1) + self.conv_out_sp4 = BiSeNetOutput(sp4_inplanes, 64, 1) + self.conv_out_sp2 = BiSeNetOutput(sp2_inplanes, 64, 1) + self.init_weight() + + def forward(self, x): + H, W = x.size()[2:] + + feat_res2, feat_res4, feat_res8, feat_res16, feat_cp8, feat_cp16 = self.cp(x) + + feat_out_sp2 = self.conv_out_sp2(feat_res2) + + feat_out_sp4 = self.conv_out_sp4(feat_res4) + + feat_out_sp8 = self.conv_out_sp8(feat_res8) + + feat_out_sp16 = self.conv_out_sp16(feat_res16) + + feat_fuse = self.ffm(feat_res8, feat_cp8) + + feat_out = self.conv_out(feat_fuse) + feat_out16 = self.conv_out16(feat_cp8) + feat_out32 = self.conv_out32(feat_cp16) + + feat_out = F.interpolate(feat_out, (H, W), mode='bilinear', align_corners=True) + feat_out16 = F.interpolate(feat_out16, (H, W), mode='bilinear', align_corners=True) + feat_out32 = F.interpolate(feat_out32, (H, W), mode='bilinear', align_corners=True) + + if self.use_boundary_2 and self.use_boundary_4 and self.use_boundary_8: + return feat_out, feat_out16, feat_out32, feat_out_sp2, feat_out_sp4, feat_out_sp8 + + if (not self.use_boundary_2) and self.use_boundary_4 and self.use_boundary_8: + return feat_out, feat_out16, feat_out32, feat_out_sp4, feat_out_sp8 + + if (not self.use_boundary_2) and (not self.use_boundary_4) and self.use_boundary_8: + return feat_out, feat_out16, feat_out32, feat_out_sp8 + + if (not self.use_boundary_2) and (not self.use_boundary_4) and (not self.use_boundary_8): + return feat_out, feat_out16, feat_out32 + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params = [], [], [], [] + for name, child in self.named_children(): + child_wd_params, child_nowd_params = child.get_params() + if isinstance(child, (FeatureFusionModule, BiSeNetOutput)): + lr_mul_wd_params += child_wd_params + lr_mul_nowd_params += child_nowd_params + else: + wd_params += child_wd_params + nowd_params += child_nowd_params + return wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params + + +if __name__ == "__main__": + + # net = BiSeNet('STDCNet813', 19) # 原始 + net = BiSeNet('STDCNet813', 3) # 改动 + + net.cuda() + net.eval() + in_ten = torch.randn(1, 3, 768, 1536).cuda() + out, out16, out32 = net(in_ten) + print(out.shape) + # torch.save(net.state_dict(), 'STDCNet813.pth')### + + diff --git a/models/model_stages_trt.py b/models/model_stages_trt.py new file mode 100644 index 0000000..4dadc43 --- /dev/null +++ b/models/model_stages_trt.py @@ -0,0 +1,408 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- +import torch +import torch.nn as nn +import torch.nn.functional as F +import torchvision + +from nets.stdcnet import STDCNet1446, STDCNet813 +BatchNorm2d = nn.BatchNorm2d + +class ConvBNReLU(nn.Module): + def __init__(self, in_chan, out_chan, ks=3, stride=1, padding=1, *args, **kwargs): + super(ConvBNReLU, self).__init__() + self.conv = nn.Conv2d(in_chan, + out_chan, + kernel_size = ks, + stride = stride, + padding = padding, + bias = False) + self.bn = BatchNorm2d(out_chan) + # self.bn = BatchNorm2d(out_chan, activation='none') + self.relu = nn.ReLU() + self.init_weight() + + def forward(self, x): + x = self.conv(x) + x = self.bn(x) + x = self.relu(x) + return x + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + +class BiSeNetOutput(nn.Module): + def __init__(self, in_chan, mid_chan, n_classes, *args, **kwargs): + super(BiSeNetOutput, self).__init__() + self.conv = ConvBNReLU(in_chan, mid_chan, ks=3, stride=1, padding=1) + self.conv_out = nn.Conv2d(mid_chan, n_classes, kernel_size=1, bias=False) + self.init_weight() + + def forward(self, x): + x = self.conv(x) + x = self.conv_out(x) + return x + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params = [], [] + for name, module in self.named_modules(): + if isinstance(module, (nn.Linear, nn.Conv2d)): + wd_params.append(module.weight) + if not module.bias is None: + nowd_params.append(module.bias) + elif isinstance(module, BatchNorm2d): + nowd_params += list(module.parameters()) + return wd_params, nowd_params + + +class AttentionRefinementModule(nn.Module): + def __init__(self, in_chan, out_chan, *args, **kwargs): + super(AttentionRefinementModule, self).__init__() + self.conv = ConvBNReLU(in_chan, out_chan, ks=3, stride=1, padding=1) + self.conv_atten = nn.Conv2d(out_chan, out_chan, kernel_size= 1, bias=False) + self.bn_atten = BatchNorm2d(out_chan) + # self.bn_atten = BatchNorm2d(out_chan, activation='none') + self.sigmoid_atten = nn.Sigmoid() + self.init_weight() + + def forward(self, x): + feat = self.conv(x) + # atten = F.avg_pool2d(feat, feat.size()[2:]) + size_array = [int(s) for s in feat.size()[2:]] + atten = torch.nn.functional.avg_pool2d(feat, size_array) + atten = self.conv_atten(atten) + atten = self.bn_atten(atten) + atten = self.sigmoid_atten(atten) + out = torch.mul(feat, atten) + return out + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + +class ContextPath(nn.Module): + def __init__(self, backbone='CatNetSmall', pretrain_model='', use_conv_last=False, input_size=512, *args, **kwargs): + super(ContextPath, self).__init__() + + self.backbone_name = backbone + self.input_size = input_size + print('backbone: ', backbone) + if backbone == 'STDCNet1446': + self.backbone = STDCNet1446(pretrain_model=pretrain_model, use_conv_last=use_conv_last) + self.arm16 = AttentionRefinementModule(512, 128) + inplanes = 1024 + if use_conv_last: + inplanes = 1024 + self.arm32 = AttentionRefinementModule(inplanes, 128) + self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1) + self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1) + self.conv_avg = ConvBNReLU(inplanes, 128, ks=1, stride=1, padding=0) + + elif backbone == 'STDCNet813': + self.backbone = STDCNet813(pretrain_model=pretrain_model, use_conv_last=use_conv_last) + self.arm16 = AttentionRefinementModule(512, 128) + inplanes = 1024 + if use_conv_last: + inplanes = 1024 + self.arm32 = AttentionRefinementModule(inplanes, 128) + self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1) + self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1) + self.conv_avg = ConvBNReLU(inplanes, 128, ks=1, stride=1, padding=0) + else: + print("backbone is not in backbone lists") + exit(0) + + if self.input_size == 512: + self.H8 = torch.tensor(64) + self.W8 = torch.tensor(128) + + self.H16 = torch.tensor(32) + self.W16 = torch.tensor(64) + + self.H32 = torch.tensor(16) + self.W32 = torch.tensor(32) + elif self.input_size == 768: + self.H8 = torch.tensor(96) + self.W8 = torch.tensor(192) + + self.H16 = torch.tensor(48) + self.W16 = torch.tensor(96) + + self.H32 = torch.tensor(24) + self.W32 = torch.tensor(48) + elif self.input_size == 1024: + self.H8 = torch.tensor(128) + self.W8 = torch.tensor(256) + + self.H16 = torch.tensor(64) + self.W16 = torch.tensor(128) + + self.H32 = torch.tensor(32) + self.W32 = torch.tensor(64) + + elif self.input_size == 720: + self.H8 = torch.tensor(90) + self.W8 = torch.tensor(120) + + self.H16 = torch.tensor(45) + self.W16 = torch.tensor(60) + + self.H32 = torch.tensor(23) + self.W32 = torch.tensor(30) + else: + print("input_size is not in input_size lists") + exit(0) + + self.init_weight() + + def forward(self, x): + + feat2, feat4, feat8, feat16, feat32 = self.backbone(x) + size_array = [int(s) for s in feat32.size()[2:]] + avg = torch.nn.functional.avg_pool2d(feat32, size_array) + + avg = self.conv_avg(avg) + avg_up = F.interpolate(avg, (self.H32, self.W32), mode='nearest') + + feat32_arm = self.arm32(feat32) + feat32_sum = feat32_arm + avg_up + feat32_up = F.interpolate(feat32_sum, (self.H16, self.W16), mode='nearest') + feat32_up = self.conv_head32(feat32_up) + + feat16_arm = self.arm16(feat16) + feat16_sum = feat16_arm + feat32_up + feat16_up = F.interpolate(feat16_sum, (self.H8, self.W8), mode='nearest') + feat16_up = self.conv_head16(feat16_up) + + return feat2, feat4, feat8, feat16, feat16_up, feat32_up # x8, x16 + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params = [], [] + for name, module in self.named_modules(): + if isinstance(module, (nn.Linear, nn.Conv2d)): + wd_params.append(module.weight) + if not module.bias is None: + nowd_params.append(module.bias) + elif isinstance(module, BatchNorm2d): + nowd_params += list(module.parameters()) + return wd_params, nowd_params + +class SpatialPath(nn.Module): + def __init__(self, *args, **kwargs): + super(SpatialPath, self).__init__() + self.conv1 = ConvBNReLU(3, 64, ks=7, stride=2, padding=3) + self.conv2 = ConvBNReLU(64, 64, ks=3, stride=2, padding=1) + self.conv3 = ConvBNReLU(64, 64, ks=3, stride=2, padding=1) + self.conv_out = ConvBNReLU(64, 128, ks=1, stride=1, padding=0) + self.init_weight() + + def forward(self, x): + feat = self.conv1(x) + feat = self.conv2(feat) + feat = self.conv3(feat) + feat = self.conv_out(feat) + return feat + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params = [], [] + for name, module in self.named_modules(): + if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d): + wd_params.append(module.weight) + if not module.bias is None: + nowd_params.append(module.bias) + elif isinstance(module, BatchNorm2d): + nowd_params += list(module.parameters()) + return wd_params, nowd_params + + +class FeatureFusionModule(nn.Module): + def __init__(self, in_chan, out_chan, *args, **kwargs): + super(FeatureFusionModule, self).__init__() + self.convblk = ConvBNReLU(in_chan, out_chan, ks=1, stride=1, padding=0) + self.conv1 = nn.Conv2d(out_chan, + out_chan//4, + kernel_size = 1, + stride = 1, + padding = 0, + bias = False) + self.conv2 = nn.Conv2d(out_chan//4, + out_chan, + kernel_size = 1, + stride = 1, + padding = 0, + bias = False) + self.relu = nn.ReLU(inplace=True) + self.sigmoid = nn.Sigmoid() + self.init_weight() + + def forward(self, fsp, fcp): + fcat = torch.cat([fsp, fcp], dim=1) + feat = self.convblk(fcat) + # atten = F.avg_pool2d(feat, feat.size()[2:]) + + size_array = [int(s) for s in feat.size()[2:]] + atten = torch.nn.functional.avg_pool2d(feat, size_array) + atten = self.conv1(atten) + atten = self.relu(atten) + atten = self.conv2(atten) + atten = self.sigmoid(atten) + feat_atten = torch.mul(feat, atten) + feat_out = feat_atten + feat + return feat_out + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params = [], [] + for name, module in self.named_modules(): + if isinstance(module, (nn.Linear, nn.Conv2d)): + wd_params.append(module.weight) + if not module.bias is None: + nowd_params.append(module.bias) + elif isinstance(module, BatchNorm2d): + nowd_params += list(module.parameters()) + return wd_params, nowd_params + + +class BiSeNet(nn.Module): + def __init__(self, backbone, n_classes, pretrain_model='', use_boundary_2=False, use_boundary_4=False, use_boundary_8=False, use_boundary_16=False, input_size=512, use_conv_last=False, heat_map=False, *args, **kwargs): + super(BiSeNet, self).__init__() + + self.use_boundary_2 = use_boundary_2 + self.use_boundary_4 = use_boundary_4 + self.use_boundary_8 = use_boundary_8 + self.use_boundary_16 = use_boundary_16 + self.input_size = input_size + + print('BiSeNet backbone: ', backbone) + self.cp = ContextPath(backbone, pretrain_model, input_size=self.input_size, use_conv_last=use_conv_last) + + if backbone == 'STDCNet1446': + conv_out_inplanes = 128 + sp2_inplanes = 32 + sp4_inplanes = 64 + sp8_inplanes = 256 + sp16_inplanes = 512 + inplane = sp8_inplanes + conv_out_inplanes + + elif backbone == 'STDCNet813': + conv_out_inplanes = 128 + sp2_inplanes = 32 + sp4_inplanes = 64 + sp8_inplanes = 256 + sp16_inplanes = 512 + inplane = sp8_inplanes + conv_out_inplanes + + else: + print("backbone is not in backbone lists") + exit(0) + + self.ffm = FeatureFusionModule(inplane, 256) + self.conv_out = BiSeNetOutput(256, 256, n_classes) + self.conv_out16 = BiSeNetOutput(conv_out_inplanes, 64, n_classes) + self.conv_out32 = BiSeNetOutput(conv_out_inplanes, 64, n_classes) + + self.conv_out_sp16 = BiSeNetOutput(sp16_inplanes, 64, 1) + self.conv_out_sp8 = BiSeNetOutput(sp8_inplanes, 64, 1) + self.conv_out_sp4 = BiSeNetOutput(sp4_inplanes, 64, 1) + self.conv_out_sp2 = BiSeNetOutput(sp2_inplanes, 64, 1) + + if self.input_size == 512: + self.H = torch.tensor(512) + self.W = torch.tensor(1024) + elif self.input_size == 768: + self.H = torch.tensor(768) + self.W = torch.tensor(1536) + elif self.input_size == 1024: + self.H = torch.tensor(1024) + self.W = torch.tensor(2048) + elif self.input_size == 720: + self.H = torch.tensor(720) + self.W = torch.tensor(960) + else: + print("input_size is not in input_size lists") + exit(0) + + self.init_weight() + + def forward(self, x): + # H, W = x.size()[2:] + + feat_res2, feat_res4, feat_res8, feat_res16, feat_cp8, feat_cp16 = self.cp(x) + # 16, 24, 40, 112, + + feat_out_sp8 = self.conv_out_sp8(feat_res8) + + feat_out_sp16 = self.conv_out_sp16(feat_res16) + + feat_fuse = self.ffm(feat_res8, feat_cp8) + + feat_out = self.conv_out(feat_fuse) + feat_out16 = self.conv_out16(feat_cp8) + feat_out32 = self.conv_out32(feat_cp16) + + feat_out = F.interpolate(feat_out, (self.H, self.W), mode='nearest') + feat_out16 = F.interpolate(feat_out16, (self.H, self.W), mode='nearest') + feat_out32 = F.interpolate(feat_out32, (self.H, self.W), mode='nearest') + + return feat_out + + def init_weight(self): + for ly in self.children(): + if isinstance(ly, nn.Conv2d): + nn.init.kaiming_normal_(ly.weight, a=1) + if not ly.bias is None: nn.init.constant_(ly.bias, 0) + + def get_params(self): + wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params = [], [], [], [] + for name, child in self.named_children(): + child_wd_params, child_nowd_params = child.get_params() + if isinstance(child, (FeatureFusionModule, BiSeNetOutput)): + lr_mul_wd_params += child_wd_params + lr_mul_nowd_params += child_nowd_params + else: + wd_params += child_wd_params + nowd_params += child_nowd_params + return wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params + + +if __name__ == "__main__": + + net = BiSeNet('STDCNet813', 19) + net.cuda() + net.eval() + in_ten = torch.randn(1, 3, 768, 1536).cuda() + out, out16, out32 = net(in_ten) + print(out.shape) + torch.save(net.state_dict(), 'STDCNet813.pth') + + diff --git a/models/yolo.py b/models/yolo.py new file mode 100644 index 0000000..f730a1e --- /dev/null +++ b/models/yolo.py @@ -0,0 +1,277 @@ +# YOLOv5 YOLO-specific modules + +import argparse +import logging +import sys +from copy import deepcopy + +sys.path.append('./') # to run '$ python *.py' files in subdirectories +logger = logging.getLogger(__name__) + +from models.common import * +from models.experimental import * +from utils.autoanchor import check_anchor_order +from utils.general import make_divisible, check_file, set_logging +from utils.torch_utils import time_synchronized, fuse_conv_and_bn, model_info, scale_img, initialize_weights, \ + select_device, copy_attr + +try: + import thop # for FLOPS computation +except ImportError: + thop = None + + +class Detect(nn.Module): + stride = None # strides computed during build + export = False # onnx export + + def __init__(self, nc=80, anchors=(), ch=()): # detection layer + super(Detect, self).__init__() + self.nc = nc # number of classes + self.no = nc + 5 # number of outputs per anchor + self.nl = len(anchors) # number of detection layers + self.na = len(anchors[0]) // 2 # number of anchors + self.grid = [torch.zeros(1)] * self.nl # init grid + a = torch.tensor(anchors).float().view(self.nl, -1, 2) + self.register_buffer('anchors', a) # shape(nl,na,2) + self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) + self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv + + def forward(self, x): + # x = x.copy() # for profiling + z = [] # inference output + self.training |= self.export + for i in range(self.nl): + x[i] = self.m[i](x[i]) # conv + bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) + x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() + + if not self.training: # inference + if self.grid[i].shape[2:4] != x[i].shape[2:4]: + self.grid[i] = self._make_grid(nx, ny).to(x[i].device) + + y = x[i].sigmoid() + y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy + y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh + z.append(y.view(bs, -1, self.no)) + + return x if self.training else (torch.cat(z, 1), x) + + @staticmethod + def _make_grid(nx=20, ny=20): + yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) + return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() + + +class Model(nn.Module): + def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None): # model, input channels, number of classes + super(Model, self).__init__() + if isinstance(cfg, dict): + self.yaml = cfg # model dict + else: # is *.yaml + import yaml # for torch hub + self.yaml_file = Path(cfg).name + with open(cfg) as f: + self.yaml = yaml.load(f, Loader=yaml.SafeLoader) # model dict + + # Define model + ch = self.yaml['ch'] = self.yaml.get('ch', ch) # input channels + if nc and nc != self.yaml['nc']: + logger.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}") + self.yaml['nc'] = nc # override yaml value + if anchors: + logger.info(f'Overriding model.yaml anchors with anchors={anchors}') + self.yaml['anchors'] = round(anchors) # override yaml value + self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) # model, savelist + self.names = [str(i) for i in range(self.yaml['nc'])] # default names + # print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))]) + + # Build strides, anchors + m = self.model[-1] # Detect() + if isinstance(m, Detect): + s = 256 # 2x min stride + m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward + m.anchors /= m.stride.view(-1, 1, 1) + check_anchor_order(m) + self.stride = m.stride + self._initialize_biases() # only run once + # print('Strides: %s' % m.stride.tolist()) + + # Init weights, biases + initialize_weights(self) + self.info() + logger.info('') + + def forward(self, x, augment=False, profile=False): + if augment: + img_size = x.shape[-2:] # height, width + s = [1, 0.83, 0.67] # scales + f = [None, 3, None] # flips (2-ud, 3-lr) + y = [] # outputs + for si, fi in zip(s, f): + xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max())) + yi = self.forward_once(xi)[0] # forward + # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1]) # save + yi[..., :4] /= si # de-scale + if fi == 2: + yi[..., 1] = img_size[0] - yi[..., 1] # de-flip ud + elif fi == 3: + yi[..., 0] = img_size[1] - yi[..., 0] # de-flip lr + y.append(yi) + return torch.cat(y, 1), None # augmented inference, train + else: + return self.forward_once(x, profile) # single-scale inference, train + + def forward_once(self, x, profile=False): + y, dt = [], [] # outputs + for m in self.model: + if m.f != -1: # if not from previous layer + x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers + + if profile: + o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 if thop else 0 # FLOPS + t = time_synchronized() + for _ in range(10): + _ = m(x) + dt.append((time_synchronized() - t) * 100) + print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type)) + + x = m(x) # run + y.append(x if m.i in self.save else None) # save output + + if profile: + print('%.1fms total' % sum(dt)) + return x + + def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency + # https://arxiv.org/abs/1708.02002 section 3.3 + # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. + m = self.model[-1] # Detect() module + for mi, s in zip(m.m, m.stride): # from + b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) + b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) + b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls + mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) + + def _print_biases(self): + m = self.model[-1] # Detect() module + for mi in m.m: # from + b = mi.bias.detach().view(m.na, -1).T # conv.bias(255) to (3,85) + print(('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean())) + + # def _print_weights(self): + # for m in self.model.modules(): + # if type(m) is Bottleneck: + # print('%10.3g' % (m.w.detach().sigmoid() * 2)) # shortcut weights + + def fuse(self): # fuse model Conv2d() + BatchNorm2d() layers + print('Fusing layers... ') + for m in self.model.modules(): + if type(m) is Conv and hasattr(m, 'bn'): + m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv + delattr(m, 'bn') # remove batchnorm + m.forward = m.fuseforward # update forward + self.info() + return self + + def nms(self, mode=True): # add or remove NMS module + present = type(self.model[-1]) is NMS # last layer is NMS + if mode and not present: + print('Adding NMS... ') + m = NMS() # module + m.f = -1 # from + m.i = self.model[-1].i + 1 # index + self.model.add_module(name='%s' % m.i, module=m) # add + self.eval() + elif not mode and present: + print('Removing NMS... ') + self.model = self.model[:-1] # remove + return self + + def autoshape(self): # add autoShape module + print('Adding autoShape... ') + m = autoShape(self) # wrap model + copy_attr(m, self, include=('yaml', 'nc', 'hyp', 'names', 'stride'), exclude=()) # copy attributes + return m + + def info(self, verbose=False, img_size=640): # print model information + model_info(self, verbose, img_size) + + +def parse_model(d, ch): # model_dict, input_channels(3) + logger.info('\n%3s%18s%3s%10s %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments')) + anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'] + na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors + no = na * (nc + 5) # number of outputs = anchors * (classes + 5) + + layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out + for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args + m = eval(m) if isinstance(m, str) else m # eval strings + for j, a in enumerate(args): + try: + args[j] = eval(a) if isinstance(a, str) else a # eval strings + except: + pass + + n = max(round(n * gd), 1) if n > 1 else n # depth gain + if m in [Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, DWConv, MixConv2d, Focus, CrossConv, BottleneckCSP, + C3, C3TR]: + c1, c2 = ch[f], args[0] + if c2 != no: # if not output + c2 = make_divisible(c2 * gw, 8) + + args = [c1, c2, *args[1:]] + if m in [BottleneckCSP, C3, C3TR]: + args.insert(2, n) # number of repeats + n = 1 + elif m is nn.BatchNorm2d: + args = [ch[f]] + elif m is Concat: + c2 = sum([ch[x] for x in f]) + elif m is Detect: + args.append([ch[x] for x in f]) + if isinstance(args[1], int): # number of anchors + args[1] = [list(range(args[1] * 2))] * len(f) + elif m is Contract: + c2 = ch[f] * args[0] ** 2 + elif m is Expand: + c2 = ch[f] // args[0] ** 2 + else: + c2 = ch[f] + + m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args) # module + t = str(m)[8:-2].replace('__main__.', '') # module type + np = sum([x.numel() for x in m_.parameters()]) # number params + m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params + logger.info('%3s%18s%3s%10.0f %-40s%-30s' % (i, f, n, np, t, args)) # print + save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist + layers.append(m_) + if i == 0: + ch = [] + ch.append(c2) + return nn.Sequential(*layers), sorted(save) + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + opt = parser.parse_args() + opt.cfg = check_file(opt.cfg) # check file + set_logging() + device = select_device(opt.device) + + # Create model + model = Model(opt.cfg).to(device) + model.train() + + # Profile + # img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device) + # y = model(img, profile=True) + + # Tensorboard + # from torch.utils.tensorboard import SummaryWriter + # tb_writer = SummaryWriter() + # print("Run 'tensorboard --logdir=models/runs' to view tensorboard at http://localhost:6006/") + # tb_writer.add_graph(model.model, img) # add model to tensorboard + # tb_writer.add_image('test', img[0], dataformats='CWH') # add model to tensorboard diff --git a/models/yolov5l.yaml b/models/yolov5l.yaml new file mode 100644 index 0000000..616be92 --- /dev/null +++ b/models/yolov5l.yaml @@ -0,0 +1,48 @@ +# parameters +nc: 3 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Focus, [64, 3]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 9, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 1, SPP, [1024, [5, 9, 13]]], + [-1, 3, C3, [1024, False]], # 9 + ] + +# YOLOv5 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/models/yolov5m.yaml b/models/yolov5m.yaml new file mode 100644 index 0000000..3c749c9 --- /dev/null +++ b/models/yolov5m.yaml @@ -0,0 +1,48 @@ +# parameters +nc: 80 # number of classes +depth_multiple: 0.67 # model depth multiple +width_multiple: 0.75 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Focus, [64, 3]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 9, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 1, SPP, [1024, [5, 9, 13]]], + [-1, 3, C3, [1024, False]], # 9 + ] + +# YOLOv5 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/models/yolov5s.yaml b/models/yolov5s.yaml new file mode 100644 index 0000000..3211b8e --- /dev/null +++ b/models/yolov5s.yaml @@ -0,0 +1,49 @@ +# parameters +#nc: 80 # number of classes +nc: 1 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Focus, [64, 3]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 9, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 1, SPP, [1024, [5, 9, 13]]], + [-1, 3, C3, [1024, False]], # 9 + ] + +# YOLOv5 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/models/yolov5x.yaml b/models/yolov5x.yaml new file mode 100644 index 0000000..006242e --- /dev/null +++ b/models/yolov5x.yaml @@ -0,0 +1,48 @@ +# parameters +nc: 3 # number of classes +depth_multiple: 1.33 # model depth multiple +width_multiple: 1.25 # layer channel multiple + +# anchors +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Focus, [64, 3]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 9, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 1, SPP, [1024, [5, 9, 13]]], + [-1, 3, C3, [1024, False]], # 9 + ] + +# YOLOv5 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/modules/__init__.py b/modules/__init__.py new file mode 100644 index 0000000..8a098de --- /dev/null +++ b/modules/__init__.py @@ -0,0 +1,5 @@ +from .bn import ABN, InPlaceABN, InPlaceABNSync +from .functions import ACT_RELU, ACT_LEAKY_RELU, ACT_ELU, ACT_NONE +from .misc import GlobalAvgPool2d, SingleGPU +from .residual import IdentityResidualBlock +from .dense import DenseModule diff --git a/modules/__pycache__/__init__.cpython-35.pyc b/modules/__pycache__/__init__.cpython-35.pyc new file mode 100644 index 0000000..fed14a7 Binary files /dev/null and b/modules/__pycache__/__init__.cpython-35.pyc differ diff --git a/modules/__pycache__/__init__.cpython-37.pyc b/modules/__pycache__/__init__.cpython-37.pyc new file mode 100644 index 0000000..8bce89b Binary files /dev/null and b/modules/__pycache__/__init__.cpython-37.pyc differ diff --git a/modules/__pycache__/__init__.cpython-38.pyc b/modules/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000..2d9bda7 Binary files /dev/null and b/modules/__pycache__/__init__.cpython-38.pyc differ diff --git a/modules/__pycache__/bn.cpython-35.pyc b/modules/__pycache__/bn.cpython-35.pyc new file mode 100644 index 0000000..60913e8 Binary files /dev/null and b/modules/__pycache__/bn.cpython-35.pyc differ diff --git a/modules/__pycache__/bn.cpython-37.pyc b/modules/__pycache__/bn.cpython-37.pyc new file mode 100644 index 0000000..de5223c Binary files /dev/null and b/modules/__pycache__/bn.cpython-37.pyc differ diff --git a/modules/__pycache__/bn.cpython-38.pyc b/modules/__pycache__/bn.cpython-38.pyc new file mode 100644 index 0000000..a84ed30 Binary files /dev/null and b/modules/__pycache__/bn.cpython-38.pyc differ diff --git a/modules/__pycache__/dense.cpython-35.pyc b/modules/__pycache__/dense.cpython-35.pyc new file mode 100644 index 0000000..40528cc Binary files /dev/null and b/modules/__pycache__/dense.cpython-35.pyc differ diff --git a/modules/__pycache__/dense.cpython-37.pyc b/modules/__pycache__/dense.cpython-37.pyc new file mode 100644 index 0000000..92a3a5d Binary files /dev/null and b/modules/__pycache__/dense.cpython-37.pyc differ diff --git a/modules/__pycache__/dense.cpython-38.pyc b/modules/__pycache__/dense.cpython-38.pyc new file mode 100644 index 0000000..9b10976 Binary files /dev/null and b/modules/__pycache__/dense.cpython-38.pyc differ diff --git a/modules/__pycache__/functions.cpython-35.pyc b/modules/__pycache__/functions.cpython-35.pyc new file mode 100644 index 0000000..73895b9 Binary files /dev/null and b/modules/__pycache__/functions.cpython-35.pyc differ diff --git a/modules/__pycache__/functions.cpython-37.pyc b/modules/__pycache__/functions.cpython-37.pyc new file mode 100644 index 0000000..4df73f1 Binary files /dev/null and b/modules/__pycache__/functions.cpython-37.pyc differ diff --git a/modules/__pycache__/functions.cpython-38.pyc b/modules/__pycache__/functions.cpython-38.pyc new file mode 100644 index 0000000..fd52434 Binary files /dev/null and b/modules/__pycache__/functions.cpython-38.pyc differ diff --git a/modules/__pycache__/misc.cpython-35.pyc b/modules/__pycache__/misc.cpython-35.pyc new file mode 100644 index 0000000..4353bf8 Binary files /dev/null and b/modules/__pycache__/misc.cpython-35.pyc differ diff --git a/modules/__pycache__/misc.cpython-37.pyc b/modules/__pycache__/misc.cpython-37.pyc new file mode 100644 index 0000000..efbbf55 Binary files /dev/null and b/modules/__pycache__/misc.cpython-37.pyc differ diff --git a/modules/__pycache__/misc.cpython-38.pyc b/modules/__pycache__/misc.cpython-38.pyc new file mode 100644 index 0000000..13d8638 Binary files /dev/null and b/modules/__pycache__/misc.cpython-38.pyc differ diff --git a/modules/__pycache__/residual.cpython-35.pyc b/modules/__pycache__/residual.cpython-35.pyc new file mode 100644 index 0000000..28dd895 Binary files /dev/null and b/modules/__pycache__/residual.cpython-35.pyc differ diff --git a/modules/__pycache__/residual.cpython-37.pyc b/modules/__pycache__/residual.cpython-37.pyc new file mode 100644 index 0000000..24a7814 Binary files /dev/null and b/modules/__pycache__/residual.cpython-37.pyc differ diff --git a/modules/__pycache__/residual.cpython-38.pyc b/modules/__pycache__/residual.cpython-38.pyc new file mode 100644 index 0000000..269a39b Binary files /dev/null and b/modules/__pycache__/residual.cpython-38.pyc differ diff --git a/modules/bn.py b/modules/bn.py new file mode 100644 index 0000000..cd3928b --- /dev/null +++ b/modules/bn.py @@ -0,0 +1,130 @@ +import torch +import torch.nn as nn +import torch.nn.functional as functional + +try: + from queue import Queue +except ImportError: + from Queue import Queue + +from .functions import * + + +class ABN(nn.Module): + """Activated Batch Normalization + + This gathers a `BatchNorm2d` and an activation function in a single module + """ + + def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, activation="leaky_relu", slope=0.01): + """Creates an Activated Batch Normalization module + + Parameters + ---------- + num_features : int + Number of feature channels in the input and output. + eps : float + Small constant to prevent numerical issues. + momentum : float + Momentum factor applied to compute running statistics as. + affine : bool + If `True` apply learned scale and shift transformation after normalization. + activation : str + Name of the activation functions, one of: `leaky_relu`, `elu` or `none`. + slope : float + Negative slope for the `leaky_relu` activation. + """ + super(ABN, self).__init__() + self.num_features = num_features + self.affine = affine + self.eps = eps + self.momentum = momentum + self.activation = activation + self.slope = slope + if self.affine: + self.weight = nn.Parameter(torch.ones(num_features)) + self.bias = nn.Parameter(torch.zeros(num_features)) + else: + self.register_parameter('weight', None) + self.register_parameter('bias', None) + self.register_buffer('running_mean', torch.zeros(num_features)) + self.register_buffer('running_var', torch.ones(num_features)) + self.reset_parameters() + + def reset_parameters(self): + nn.init.constant_(self.running_mean, 0) + nn.init.constant_(self.running_var, 1) + if self.affine: + nn.init.constant_(self.weight, 1) + nn.init.constant_(self.bias, 0) + + def forward(self, x): + x = functional.batch_norm(x, self.running_mean, self.running_var, self.weight, self.bias, + self.training, self.momentum, self.eps) + + if self.activation == ACT_RELU: + return functional.relu(x, inplace=True) + elif self.activation == ACT_LEAKY_RELU: + return functional.leaky_relu(x, negative_slope=self.slope, inplace=True) + elif self.activation == ACT_ELU: + return functional.elu(x, inplace=True) + else: + return x + + def __repr__(self): + rep = '{name}({num_features}, eps={eps}, momentum={momentum},' \ + ' affine={affine}, activation={activation}' + if self.activation == "leaky_relu": + rep += ', slope={slope})' + else: + rep += ')' + return rep.format(name=self.__class__.__name__, **self.__dict__) + + +class InPlaceABN(ABN): + """InPlace Activated Batch Normalization""" + + def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, activation="leaky_relu", slope=0.01): + """Creates an InPlace Activated Batch Normalization module + + Parameters + ---------- + num_features : int + Number of feature channels in the input and output. + eps : float + Small constant to prevent numerical issues. + momentum : float + Momentum factor applied to compute running statistics as. + affine : bool + If `True` apply learned scale and shift transformation after normalization. + activation : str + Name of the activation functions, one of: `leaky_relu`, `elu` or `none`. + slope : float + Negative slope for the `leaky_relu` activation. + """ + super(InPlaceABN, self).__init__(num_features, eps, momentum, affine, activation, slope) + + def forward(self, x): + return inplace_abn(x, self.weight, self.bias, self.running_mean, self.running_var, + self.training, self.momentum, self.eps, self.activation, self.slope) + + +class InPlaceABNSync(ABN): + """InPlace Activated Batch Normalization with cross-GPU synchronization + This assumes that it will be replicated across GPUs using the same mechanism as in `nn.DistributedDataParallel`. + """ + + def forward(self, x): + return inplace_abn_sync(x, self.weight, self.bias, self.running_mean, self.running_var, + self.training, self.momentum, self.eps, self.activation, self.slope) + + def __repr__(self): + rep = '{name}({num_features}, eps={eps}, momentum={momentum},' \ + ' affine={affine}, activation={activation}' + if self.activation == "leaky_relu": + rep += ', slope={slope})' + else: + rep += ')' + return rep.format(name=self.__class__.__name__, **self.__dict__) + + diff --git a/modules/deeplab.py b/modules/deeplab.py new file mode 100644 index 0000000..fd25b78 --- /dev/null +++ b/modules/deeplab.py @@ -0,0 +1,84 @@ +import torch +import torch.nn as nn +import torch.nn.functional as functional + +from models._util import try_index +from .bn import ABN + + +class DeeplabV3(nn.Module): + def __init__(self, + in_channels, + out_channels, + hidden_channels=256, + dilations=(12, 24, 36), + norm_act=ABN, + pooling_size=None): + super(DeeplabV3, self).__init__() + self.pooling_size = pooling_size + + self.map_convs = nn.ModuleList([ + nn.Conv2d(in_channels, hidden_channels, 1, bias=False), + nn.Conv2d(in_channels, hidden_channels, 3, bias=False, dilation=dilations[0], padding=dilations[0]), + nn.Conv2d(in_channels, hidden_channels, 3, bias=False, dilation=dilations[1], padding=dilations[1]), + nn.Conv2d(in_channels, hidden_channels, 3, bias=False, dilation=dilations[2], padding=dilations[2]) + ]) + self.map_bn = norm_act(hidden_channels * 4) + + self.global_pooling_conv = nn.Conv2d(in_channels, hidden_channels, 1, bias=False) + self.global_pooling_bn = norm_act(hidden_channels) + + self.red_conv = nn.Conv2d(hidden_channels * 4, out_channels, 1, bias=False) + self.pool_red_conv = nn.Conv2d(hidden_channels, out_channels, 1, bias=False) + self.red_bn = norm_act(out_channels) + + self.reset_parameters(self.map_bn.activation, self.map_bn.slope) + + def reset_parameters(self, activation, slope): + gain = nn.init.calculate_gain(activation, slope) + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.xavier_normal_(m.weight.data, gain) + if hasattr(m, "bias") and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, ABN): + if hasattr(m, "weight") and m.weight is not None: + nn.init.constant_(m.weight, 1) + if hasattr(m, "bias") and m.bias is not None: + nn.init.constant_(m.bias, 0) + + def forward(self, x): + # Map convolutions + out = torch.cat([m(x) for m in self.map_convs], dim=1) + out = self.map_bn(out) + out = self.red_conv(out) + + # Global pooling + pool = self._global_pooling(x) + pool = self.global_pooling_conv(pool) + pool = self.global_pooling_bn(pool) + pool = self.pool_red_conv(pool) + if self.training or self.pooling_size is None: + pool = pool.repeat(1, 1, x.size(2), x.size(3)) + + out += pool + out = self.red_bn(out) + return out + + def _global_pooling(self, x): + if self.training or self.pooling_size is None: + pool = x.view(x.size(0), x.size(1), -1).mean(dim=-1) + pool = pool.view(x.size(0), x.size(1), 1, 1) + else: + pooling_size = (min(try_index(self.pooling_size, 0), x.shape[2]), + min(try_index(self.pooling_size, 1), x.shape[3])) + padding = ( + (pooling_size[1] - 1) // 2, + (pooling_size[1] - 1) // 2 if pooling_size[1] % 2 == 1 else (pooling_size[1] - 1) // 2 + 1, + (pooling_size[0] - 1) // 2, + (pooling_size[0] - 1) // 2 if pooling_size[0] % 2 == 1 else (pooling_size[0] - 1) // 2 + 1 + ) + + pool = functional.avg_pool2d(x, pooling_size, stride=1) + pool = functional.pad(pool, pad=padding, mode="replicate") + return pool diff --git a/modules/dense.py b/modules/dense.py new file mode 100644 index 0000000..9638d6e --- /dev/null +++ b/modules/dense.py @@ -0,0 +1,42 @@ +from collections import OrderedDict + +import torch +import torch.nn as nn + +from .bn import ABN + + +class DenseModule(nn.Module): + def __init__(self, in_channels, growth, layers, bottleneck_factor=4, norm_act=ABN, dilation=1): + super(DenseModule, self).__init__() + self.in_channels = in_channels + self.growth = growth + self.layers = layers + + self.convs1 = nn.ModuleList() + self.convs3 = nn.ModuleList() + for i in range(self.layers): + self.convs1.append(nn.Sequential(OrderedDict([ + ("bn", norm_act(in_channels)), + ("conv", nn.Conv2d(in_channels, self.growth * bottleneck_factor, 1, bias=False)) + ]))) + self.convs3.append(nn.Sequential(OrderedDict([ + ("bn", norm_act(self.growth * bottleneck_factor)), + ("conv", nn.Conv2d(self.growth * bottleneck_factor, self.growth, 3, padding=dilation, bias=False, + dilation=dilation)) + ]))) + in_channels += self.growth + + @property + def out_channels(self): + return self.in_channels + self.growth * self.layers + + def forward(self, x): + inputs = [x] + for i in range(self.layers): + x = torch.cat(inputs, dim=1) + x = self.convs1[i](x) + x = self.convs3[i](x) + inputs += [x] + + return torch.cat(inputs, dim=1) diff --git a/modules/functions.py b/modules/functions.py new file mode 100644 index 0000000..093615f --- /dev/null +++ b/modules/functions.py @@ -0,0 +1,234 @@ +from os import path +import torch +import torch.distributed as dist +import torch.autograd as autograd +import torch.cuda.comm as comm +from torch.autograd.function import once_differentiable +from torch.utils.cpp_extension import load + +_src_path = path.join(path.dirname(path.abspath(__file__)), "src") +_backend = load(name="inplace_abn", + extra_cflags=["-O3"], + sources=[path.join(_src_path, f) for f in [ + "inplace_abn.cpp", + "inplace_abn_cpu.cpp", + "inplace_abn_cuda.cu", + "inplace_abn_cuda_half.cu" + ]], + extra_cuda_cflags=["--expt-extended-lambda"]) + +# Activation names +ACT_RELU = "relu" +ACT_LEAKY_RELU = "leaky_relu" +ACT_ELU = "elu" +ACT_NONE = "none" + + +def _check(fn, *args, **kwargs): + success = fn(*args, **kwargs) + if not success: + raise RuntimeError("CUDA Error encountered in {}".format(fn)) + + +def _broadcast_shape(x): + out_size = [] + for i, s in enumerate(x.size()): + if i != 1: + out_size.append(1) + else: + out_size.append(s) + return out_size + + +def _reduce(x): + if len(x.size()) == 2: + return x.sum(dim=0) + else: + n, c = x.size()[0:2] + return x.contiguous().view((n, c, -1)).sum(2).sum(0) + + +def _count_samples(x): + count = 1 + for i, s in enumerate(x.size()): + if i != 1: + count *= s + return count + + +def _act_forward(ctx, x): + if ctx.activation == ACT_LEAKY_RELU: + _backend.leaky_relu_forward(x, ctx.slope) + elif ctx.activation == ACT_ELU: + _backend.elu_forward(x) + elif ctx.activation == ACT_NONE: + pass + + +def _act_backward(ctx, x, dx): + if ctx.activation == ACT_LEAKY_RELU: + _backend.leaky_relu_backward(x, dx, ctx.slope) + elif ctx.activation == ACT_ELU: + _backend.elu_backward(x, dx) + elif ctx.activation == ACT_NONE: + pass + + +class InPlaceABN(autograd.Function): + @staticmethod + def forward(ctx, x, weight, bias, running_mean, running_var, + training=True, momentum=0.1, eps=1e-05, activation=ACT_LEAKY_RELU, slope=0.01): + # Save context + ctx.training = training + ctx.momentum = momentum + ctx.eps = eps + ctx.activation = activation + ctx.slope = slope + ctx.affine = weight is not None and bias is not None + + # Prepare inputs + count = _count_samples(x) + x = x.contiguous() + weight = weight.contiguous() if ctx.affine else x.new_empty(0) + bias = bias.contiguous() if ctx.affine else x.new_empty(0) + + if ctx.training: + mean, var = _backend.mean_var(x) + + # Update running stats + running_mean.mul_((1 - ctx.momentum)).add_(ctx.momentum * mean) + running_var.mul_((1 - ctx.momentum)).add_(ctx.momentum * var * count / (count - 1)) + + # Mark in-place modified tensors + ctx.mark_dirty(x, running_mean, running_var) + else: + mean, var = running_mean.contiguous(), running_var.contiguous() + ctx.mark_dirty(x) + + # BN forward + activation + _backend.forward(x, mean, var, weight, bias, ctx.affine, ctx.eps) + _act_forward(ctx, x) + + # Output + ctx.var = var + ctx.save_for_backward(x, var, weight, bias) + return x + + @staticmethod + @once_differentiable + def backward(ctx, dz): + z, var, weight, bias = ctx.saved_tensors + dz = dz.contiguous() + + # Undo activation + _act_backward(ctx, z, dz) + + if ctx.training: + edz, eydz = _backend.edz_eydz(z, dz, weight, bias, ctx.affine, ctx.eps) + else: + # TODO: implement simplified CUDA backward for inference mode + edz = dz.new_zeros(dz.size(1)) + eydz = dz.new_zeros(dz.size(1)) + + dx = _backend.backward(z, dz, var, weight, bias, edz, eydz, ctx.affine, ctx.eps) + dweight = eydz * weight.sign() if ctx.affine else None + dbias = edz if ctx.affine else None + + return dx, dweight, dbias, None, None, None, None, None, None, None + +class InPlaceABNSync(autograd.Function): + @classmethod + def forward(cls, ctx, x, weight, bias, running_mean, running_var, + training=True, momentum=0.1, eps=1e-05, activation=ACT_LEAKY_RELU, slope=0.01, equal_batches=True): + # Save context + ctx.training = training + ctx.momentum = momentum + ctx.eps = eps + ctx.activation = activation + ctx.slope = slope + ctx.affine = weight is not None and bias is not None + + # Prepare inputs + ctx.world_size = dist.get_world_size() if dist.is_initialized() else 1 + + #count = _count_samples(x) + batch_size = x.new_tensor([x.shape[0]],dtype=torch.long) + + x = x.contiguous() + weight = weight.contiguous() if ctx.affine else x.new_empty(0) + bias = bias.contiguous() if ctx.affine else x.new_empty(0) + + if ctx.training: + mean, var = _backend.mean_var(x) + if ctx.world_size>1: + # get global batch size + if equal_batches: + batch_size *= ctx.world_size + else: + dist.all_reduce(batch_size, dist.ReduceOp.SUM) + + ctx.factor = x.shape[0]/float(batch_size.item()) + + mean_all = mean.clone() * ctx.factor + dist.all_reduce(mean_all, dist.ReduceOp.SUM) + + var_all = (var + (mean - mean_all) ** 2) * ctx.factor + dist.all_reduce(var_all, dist.ReduceOp.SUM) + + mean = mean_all + var = var_all + + # Update running stats + running_mean.mul_((1 - ctx.momentum)).add_(ctx.momentum * mean) + count = batch_size.item() * x.view(x.shape[0],x.shape[1],-1).shape[-1] + running_var.mul_((1 - ctx.momentum)).add_(ctx.momentum * var * (float(count) / (count - 1))) + + # Mark in-place modified tensors + ctx.mark_dirty(x, running_mean, running_var) + else: + mean, var = running_mean.contiguous(), running_var.contiguous() + ctx.mark_dirty(x) + + # BN forward + activation + _backend.forward(x, mean, var, weight, bias, ctx.affine, ctx.eps) + _act_forward(ctx, x) + + # Output + ctx.var = var + ctx.save_for_backward(x, var, weight, bias) + return x + + @staticmethod + @once_differentiable + def backward(ctx, dz): + z, var, weight, bias = ctx.saved_tensors + dz = dz.contiguous() + + # Undo activation + _act_backward(ctx, z, dz) + + if ctx.training: + edz, eydz = _backend.edz_eydz(z, dz, weight, bias, ctx.affine, ctx.eps) + edz_local = edz.clone() + eydz_local = eydz.clone() + + if ctx.world_size>1: + edz *= ctx.factor + dist.all_reduce(edz, dist.ReduceOp.SUM) + + eydz *= ctx.factor + dist.all_reduce(eydz, dist.ReduceOp.SUM) + else: + edz_local = edz = dz.new_zeros(dz.size(1)) + eydz_local = eydz = dz.new_zeros(dz.size(1)) + + dx = _backend.backward(z, dz, var, weight, bias, edz, eydz, ctx.affine, ctx.eps) + dweight = eydz_local * weight.sign() if ctx.affine else None + dbias = edz_local if ctx.affine else None + + return dx, dweight, dbias, None, None, None, None, None, None, None + +inplace_abn = InPlaceABN.apply +inplace_abn_sync = InPlaceABNSync.apply + +__all__ = ["inplace_abn", "inplace_abn_sync", "ACT_RELU", "ACT_LEAKY_RELU", "ACT_ELU", "ACT_NONE"] diff --git a/modules/misc.py b/modules/misc.py new file mode 100644 index 0000000..3c50b69 --- /dev/null +++ b/modules/misc.py @@ -0,0 +1,21 @@ +import torch.nn as nn +import torch +import torch.distributed as dist + +class GlobalAvgPool2d(nn.Module): + def __init__(self): + """Global average pooling over the input's spatial dimensions""" + super(GlobalAvgPool2d, self).__init__() + + def forward(self, inputs): + in_size = inputs.size() + return inputs.view((in_size[0], in_size[1], -1)).mean(dim=2) + +class SingleGPU(nn.Module): + def __init__(self, module): + super(SingleGPU, self).__init__() + self.module=module + + def forward(self, input): + return self.module(input.cuda(non_blocking=True)) + diff --git a/modules/residual.py b/modules/residual.py new file mode 100644 index 0000000..b7d51ad --- /dev/null +++ b/modules/residual.py @@ -0,0 +1,88 @@ +from collections import OrderedDict + +import torch.nn as nn + +from .bn import ABN + + +class IdentityResidualBlock(nn.Module): + def __init__(self, + in_channels, + channels, + stride=1, + dilation=1, + groups=1, + norm_act=ABN, + dropout=None): + """Configurable identity-mapping residual block + + Parameters + ---------- + in_channels : int + Number of input channels. + channels : list of int + Number of channels in the internal feature maps. Can either have two or three elements: if three construct + a residual block with two `3 x 3` convolutions, otherwise construct a bottleneck block with `1 x 1`, then + `3 x 3` then `1 x 1` convolutions. + stride : int + Stride of the first `3 x 3` convolution + dilation : int + Dilation to apply to the `3 x 3` convolutions. + groups : int + Number of convolution groups. This is used to create ResNeXt-style blocks and is only compatible with + bottleneck blocks. + norm_act : callable + Function to create normalization / activation Module. + dropout: callable + Function to create Dropout Module. + """ + super(IdentityResidualBlock, self).__init__() + + # Check parameters for inconsistencies + if len(channels) != 2 and len(channels) != 3: + raise ValueError("channels must contain either two or three values") + if len(channels) == 2 and groups != 1: + raise ValueError("groups > 1 are only valid if len(channels) == 3") + + is_bottleneck = len(channels) == 3 + need_proj_conv = stride != 1 or in_channels != channels[-1] + + self.bn1 = norm_act(in_channels) + if not is_bottleneck: + layers = [ + ("conv1", nn.Conv2d(in_channels, channels[0], 3, stride=stride, padding=dilation, bias=False, + dilation=dilation)), + ("bn2", norm_act(channels[0])), + ("conv2", nn.Conv2d(channels[0], channels[1], 3, stride=1, padding=dilation, bias=False, + dilation=dilation)) + ] + if dropout is not None: + layers = layers[0:2] + [("dropout", dropout())] + layers[2:] + else: + layers = [ + ("conv1", nn.Conv2d(in_channels, channels[0], 1, stride=stride, padding=0, bias=False)), + ("bn2", norm_act(channels[0])), + ("conv2", nn.Conv2d(channels[0], channels[1], 3, stride=1, padding=dilation, bias=False, + groups=groups, dilation=dilation)), + ("bn3", norm_act(channels[1])), + ("conv3", nn.Conv2d(channels[1], channels[2], 1, stride=1, padding=0, bias=False)) + ] + if dropout is not None: + layers = layers[0:4] + [("dropout", dropout())] + layers[4:] + self.convs = nn.Sequential(OrderedDict(layers)) + + if need_proj_conv: + self.proj_conv = nn.Conv2d(in_channels, channels[-1], 1, stride=stride, padding=0, bias=False) + + def forward(self, x): + if hasattr(self, "proj_conv"): + bn1 = self.bn1(x) + shortcut = self.proj_conv(bn1) + else: + shortcut = x.clone() + bn1 = self.bn1(x) + + out = self.convs(bn1) + out.add_(shortcut) + + return out diff --git a/modules/src/checks.h b/modules/src/checks.h new file mode 100644 index 0000000..e761a6f --- /dev/null +++ b/modules/src/checks.h @@ -0,0 +1,15 @@ +#pragma once + +#include + +// Define AT_CHECK for old version of ATen where the same function was called AT_ASSERT +#ifndef AT_CHECK +#define AT_CHECK AT_ASSERT +#endif + +#define CHECK_CUDA(x) AT_CHECK((x).type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CPU(x) AT_CHECK(!(x).type().is_cuda(), #x " must be a CPU tensor") +#define CHECK_CONTIGUOUS(x) AT_CHECK((x).is_contiguous(), #x " must be contiguous") + +#define CHECK_CUDA_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) +#define CHECK_CPU_INPUT(x) CHECK_CPU(x); CHECK_CONTIGUOUS(x) \ No newline at end of file diff --git a/modules/src/inplace_abn.cpp b/modules/src/inplace_abn.cpp new file mode 100644 index 0000000..0a6b112 --- /dev/null +++ b/modules/src/inplace_abn.cpp @@ -0,0 +1,95 @@ +#include + +#include + +#include "inplace_abn.h" + +std::vector mean_var(at::Tensor x) { + if (x.is_cuda()) { + if (x.type().scalarType() == at::ScalarType::Half) { + return mean_var_cuda_h(x); + } else { + return mean_var_cuda(x); + } + } else { + return mean_var_cpu(x); + } +} + +at::Tensor forward(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias, + bool affine, float eps) { + if (x.is_cuda()) { + if (x.type().scalarType() == at::ScalarType::Half) { + return forward_cuda_h(x, mean, var, weight, bias, affine, eps); + } else { + return forward_cuda(x, mean, var, weight, bias, affine, eps); + } + } else { + return forward_cpu(x, mean, var, weight, bias, affine, eps); + } +} + +std::vector edz_eydz(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias, + bool affine, float eps) { + if (z.is_cuda()) { + if (z.type().scalarType() == at::ScalarType::Half) { + return edz_eydz_cuda_h(z, dz, weight, bias, affine, eps); + } else { + return edz_eydz_cuda(z, dz, weight, bias, affine, eps); + } + } else { + return edz_eydz_cpu(z, dz, weight, bias, affine, eps); + } +} + +at::Tensor backward(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias, + at::Tensor edz, at::Tensor eydz, bool affine, float eps) { + if (z.is_cuda()) { + if (z.type().scalarType() == at::ScalarType::Half) { + return backward_cuda_h(z, dz, var, weight, bias, edz, eydz, affine, eps); + } else { + return backward_cuda(z, dz, var, weight, bias, edz, eydz, affine, eps); + } + } else { + return backward_cpu(z, dz, var, weight, bias, edz, eydz, affine, eps); + } +} + +void leaky_relu_forward(at::Tensor z, float slope) { + at::leaky_relu_(z, slope); +} + +void leaky_relu_backward(at::Tensor z, at::Tensor dz, float slope) { + if (z.is_cuda()) { + if (z.type().scalarType() == at::ScalarType::Half) { + return leaky_relu_backward_cuda_h(z, dz, slope); + } else { + return leaky_relu_backward_cuda(z, dz, slope); + } + } else { + return leaky_relu_backward_cpu(z, dz, slope); + } +} + +void elu_forward(at::Tensor z) { + at::elu_(z); +} + +void elu_backward(at::Tensor z, at::Tensor dz) { + if (z.is_cuda()) { + return elu_backward_cuda(z, dz); + } else { + return elu_backward_cpu(z, dz); + } +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("mean_var", &mean_var, "Mean and variance computation"); + m.def("forward", &forward, "In-place forward computation"); + m.def("edz_eydz", &edz_eydz, "First part of backward computation"); + m.def("backward", &backward, "Second part of backward computation"); + m.def("leaky_relu_forward", &leaky_relu_forward, "Leaky relu forward computation"); + m.def("leaky_relu_backward", &leaky_relu_backward, "Leaky relu backward computation and inversion"); + m.def("elu_forward", &elu_forward, "Elu forward computation"); + m.def("elu_backward", &elu_backward, "Elu backward computation and inversion"); +} diff --git a/modules/src/inplace_abn.h b/modules/src/inplace_abn.h new file mode 100644 index 0000000..17afd11 --- /dev/null +++ b/modules/src/inplace_abn.h @@ -0,0 +1,88 @@ +#pragma once + +#include + +#include + +std::vector mean_var_cpu(at::Tensor x); +std::vector mean_var_cuda(at::Tensor x); +std::vector mean_var_cuda_h(at::Tensor x); + +at::Tensor forward_cpu(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias, + bool affine, float eps); +at::Tensor forward_cuda(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias, + bool affine, float eps); +at::Tensor forward_cuda_h(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias, + bool affine, float eps); + +std::vector edz_eydz_cpu(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias, + bool affine, float eps); +std::vector edz_eydz_cuda(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias, + bool affine, float eps); +std::vector edz_eydz_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias, + bool affine, float eps); + +at::Tensor backward_cpu(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias, + at::Tensor edz, at::Tensor eydz, bool affine, float eps); +at::Tensor backward_cuda(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias, + at::Tensor edz, at::Tensor eydz, bool affine, float eps); +at::Tensor backward_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias, + at::Tensor edz, at::Tensor eydz, bool affine, float eps); + +void leaky_relu_backward_cpu(at::Tensor z, at::Tensor dz, float slope); +void leaky_relu_backward_cuda(at::Tensor z, at::Tensor dz, float slope); +void leaky_relu_backward_cuda_h(at::Tensor z, at::Tensor dz, float slope); + +void elu_backward_cpu(at::Tensor z, at::Tensor dz); +void elu_backward_cuda(at::Tensor z, at::Tensor dz); + +static void get_dims(at::Tensor x, int64_t& num, int64_t& chn, int64_t& sp) { + num = x.size(0); + chn = x.size(1); + sp = 1; + for (int64_t i = 2; i < x.ndimension(); ++i) + sp *= x.size(i); +} + +/* + * Specialized CUDA reduction functions for BN + */ +#ifdef __CUDACC__ + +#include "utils/cuda.cuh" + +template +__device__ T reduce(Op op, int plane, int N, int S) { + T sum = (T)0; + for (int batch = 0; batch < N; ++batch) { + for (int x = threadIdx.x; x < S; x += blockDim.x) { + sum += op(batch, plane, x); + } + } + + // sum over NumThreads within a warp + sum = warpSum(sum); + + // 'transpose', and reduce within warp again + __shared__ T shared[32]; + __syncthreads(); + if (threadIdx.x % WARP_SIZE == 0) { + shared[threadIdx.x / WARP_SIZE] = sum; + } + if (threadIdx.x >= blockDim.x / WARP_SIZE && threadIdx.x < WARP_SIZE) { + // zero out the other entries in shared + shared[threadIdx.x] = (T)0; + } + __syncthreads(); + if (threadIdx.x / WARP_SIZE == 0) { + sum = warpSum(shared[threadIdx.x]); + if (threadIdx.x == 0) { + shared[0] = sum; + } + } + __syncthreads(); + + // Everyone picks it up, should be broadcast into the whole gradInput + return shared[0]; +} +#endif diff --git a/modules/src/inplace_abn_cpu.cpp b/modules/src/inplace_abn_cpu.cpp new file mode 100644 index 0000000..ffc6d38 --- /dev/null +++ b/modules/src/inplace_abn_cpu.cpp @@ -0,0 +1,119 @@ +#include + +#include + +#include "utils/checks.h" +#include "inplace_abn.h" + +at::Tensor reduce_sum(at::Tensor x) { + if (x.ndimension() == 2) { + return x.sum(0); + } else { + auto x_view = x.view({x.size(0), x.size(1), -1}); + return x_view.sum(-1).sum(0); + } +} + +at::Tensor broadcast_to(at::Tensor v, at::Tensor x) { + if (x.ndimension() == 2) { + return v; + } else { + std::vector broadcast_size = {1, -1}; + for (int64_t i = 2; i < x.ndimension(); ++i) + broadcast_size.push_back(1); + + return v.view(broadcast_size); + } +} + +int64_t count(at::Tensor x) { + int64_t count = x.size(0); + for (int64_t i = 2; i < x.ndimension(); ++i) + count *= x.size(i); + + return count; +} + +at::Tensor invert_affine(at::Tensor z, at::Tensor weight, at::Tensor bias, bool affine, float eps) { + if (affine) { + return (z - broadcast_to(bias, z)) / broadcast_to(at::abs(weight) + eps, z); + } else { + return z; + } +} + +std::vector mean_var_cpu(at::Tensor x) { + auto num = count(x); + auto mean = reduce_sum(x) / num; + auto diff = x - broadcast_to(mean, x); + auto var = reduce_sum(diff.pow(2)) / num; + + return {mean, var}; +} + +at::Tensor forward_cpu(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias, + bool affine, float eps) { + auto gamma = affine ? at::abs(weight) + eps : at::ones_like(var); + auto mul = at::rsqrt(var + eps) * gamma; + + x.sub_(broadcast_to(mean, x)); + x.mul_(broadcast_to(mul, x)); + if (affine) x.add_(broadcast_to(bias, x)); + + return x; +} + +std::vector edz_eydz_cpu(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias, + bool affine, float eps) { + auto edz = reduce_sum(dz); + auto y = invert_affine(z, weight, bias, affine, eps); + auto eydz = reduce_sum(y * dz); + + return {edz, eydz}; +} + +at::Tensor backward_cpu(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias, + at::Tensor edz, at::Tensor eydz, bool affine, float eps) { + auto y = invert_affine(z, weight, bias, affine, eps); + auto mul = affine ? at::rsqrt(var + eps) * (at::abs(weight) + eps) : at::rsqrt(var + eps); + + auto num = count(z); + auto dx = (dz - broadcast_to(edz / num, dz) - y * broadcast_to(eydz / num, dz)) * broadcast_to(mul, dz); + return dx; +} + +void leaky_relu_backward_cpu(at::Tensor z, at::Tensor dz, float slope) { + CHECK_CPU_INPUT(z); + CHECK_CPU_INPUT(dz); + + AT_DISPATCH_FLOATING_TYPES(z.type(), "leaky_relu_backward_cpu", ([&] { + int64_t count = z.numel(); + auto *_z = z.data(); + auto *_dz = dz.data(); + + for (int64_t i = 0; i < count; ++i) { + if (_z[i] < 0) { + _z[i] *= 1 / slope; + _dz[i] *= slope; + } + } + })); +} + +void elu_backward_cpu(at::Tensor z, at::Tensor dz) { + CHECK_CPU_INPUT(z); + CHECK_CPU_INPUT(dz); + + AT_DISPATCH_FLOATING_TYPES(z.type(), "elu_backward_cpu", ([&] { + int64_t count = z.numel(); + auto *_z = z.data(); + auto *_dz = dz.data(); + + for (int64_t i = 0; i < count; ++i) { + if (_z[i] < 0) { + _z[i] = log1p(_z[i]); + _dz[i] *= (_z[i] + 1.f); + } + } + })); +} diff --git a/modules/src/inplace_abn_cuda.cu b/modules/src/inplace_abn_cuda.cu new file mode 100644 index 0000000..b157b06 --- /dev/null +++ b/modules/src/inplace_abn_cuda.cu @@ -0,0 +1,333 @@ +#include + +#include +#include + +#include + +#include "utils/checks.h" +#include "utils/cuda.cuh" +#include "inplace_abn.h" + +#include + +// Operations for reduce +template +struct SumOp { + __device__ SumOp(const T *t, int c, int s) + : tensor(t), chn(c), sp(s) {} + __device__ __forceinline__ T operator()(int batch, int plane, int n) { + return tensor[(batch * chn + plane) * sp + n]; + } + const T *tensor; + const int chn; + const int sp; +}; + +template +struct VarOp { + __device__ VarOp(T m, const T *t, int c, int s) + : mean(m), tensor(t), chn(c), sp(s) {} + __device__ __forceinline__ T operator()(int batch, int plane, int n) { + T val = tensor[(batch * chn + plane) * sp + n]; + return (val - mean) * (val - mean); + } + const T mean; + const T *tensor; + const int chn; + const int sp; +}; + +template +struct GradOp { + __device__ GradOp(T _weight, T _bias, const T *_z, const T *_dz, int c, int s) + : weight(_weight), bias(_bias), z(_z), dz(_dz), chn(c), sp(s) {} + __device__ __forceinline__ Pair operator()(int batch, int plane, int n) { + T _y = (z[(batch * chn + plane) * sp + n] - bias) / weight; + T _dz = dz[(batch * chn + plane) * sp + n]; + return Pair(_dz, _y * _dz); + } + const T weight; + const T bias; + const T *z; + const T *dz; + const int chn; + const int sp; +}; + +/*********** + * mean_var + ***********/ + +template +__global__ void mean_var_kernel(const T *x, T *mean, T *var, int num, int chn, int sp) { + int plane = blockIdx.x; + T norm = T(1) / T(num * sp); + + T _mean = reduce>(SumOp(x, chn, sp), plane, num, sp) * norm; + __syncthreads(); + T _var = reduce>(VarOp(_mean, x, chn, sp), plane, num, sp) * norm; + + if (threadIdx.x == 0) { + mean[plane] = _mean; + var[plane] = _var; + } +} + +std::vector mean_var_cuda(at::Tensor x) { + CHECK_CUDA_INPUT(x); + + // Extract dimensions + int64_t num, chn, sp; + get_dims(x, num, chn, sp); + + // Prepare output tensors + auto mean = at::empty({chn}, x.options()); + auto var = at::empty({chn}, x.options()); + + // Run kernel + dim3 blocks(chn); + dim3 threads(getNumThreads(sp)); + auto stream = at::cuda::getCurrentCUDAStream(); + AT_DISPATCH_FLOATING_TYPES(x.type(), "mean_var_cuda", ([&] { + mean_var_kernel<<>>( + x.data(), + mean.data(), + var.data(), + num, chn, sp); + })); + + return {mean, var}; +} + +/********** + * forward + **********/ + +template +__global__ void forward_kernel(T *x, const T *mean, const T *var, const T *weight, const T *bias, + bool affine, float eps, int num, int chn, int sp) { + int plane = blockIdx.x; + + T _mean = mean[plane]; + T _var = var[plane]; + T _weight = affine ? abs(weight[plane]) + eps : T(1); + T _bias = affine ? bias[plane] : T(0); + + T mul = rsqrt(_var + eps) * _weight; + + for (int batch = 0; batch < num; ++batch) { + for (int n = threadIdx.x; n < sp; n += blockDim.x) { + T _x = x[(batch * chn + plane) * sp + n]; + T _y = (_x - _mean) * mul + _bias; + + x[(batch * chn + plane) * sp + n] = _y; + } + } +} + +at::Tensor forward_cuda(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias, + bool affine, float eps) { + CHECK_CUDA_INPUT(x); + CHECK_CUDA_INPUT(mean); + CHECK_CUDA_INPUT(var); + CHECK_CUDA_INPUT(weight); + CHECK_CUDA_INPUT(bias); + + // Extract dimensions + int64_t num, chn, sp; + get_dims(x, num, chn, sp); + + // Run kernel + dim3 blocks(chn); + dim3 threads(getNumThreads(sp)); + auto stream = at::cuda::getCurrentCUDAStream(); + AT_DISPATCH_FLOATING_TYPES(x.type(), "forward_cuda", ([&] { + forward_kernel<<>>( + x.data(), + mean.data(), + var.data(), + weight.data(), + bias.data(), + affine, eps, num, chn, sp); + })); + + return x; +} + +/*********** + * edz_eydz + ***********/ + +template +__global__ void edz_eydz_kernel(const T *z, const T *dz, const T *weight, const T *bias, + T *edz, T *eydz, bool affine, float eps, int num, int chn, int sp) { + int plane = blockIdx.x; + + T _weight = affine ? abs(weight[plane]) + eps : 1.f; + T _bias = affine ? bias[plane] : 0.f; + + Pair res = reduce, GradOp>(GradOp(_weight, _bias, z, dz, chn, sp), plane, num, sp); + __syncthreads(); + + if (threadIdx.x == 0) { + edz[plane] = res.v1; + eydz[plane] = res.v2; + } +} + +std::vector edz_eydz_cuda(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias, + bool affine, float eps) { + CHECK_CUDA_INPUT(z); + CHECK_CUDA_INPUT(dz); + CHECK_CUDA_INPUT(weight); + CHECK_CUDA_INPUT(bias); + + // Extract dimensions + int64_t num, chn, sp; + get_dims(z, num, chn, sp); + + auto edz = at::empty({chn}, z.options()); + auto eydz = at::empty({chn}, z.options()); + + // Run kernel + dim3 blocks(chn); + dim3 threads(getNumThreads(sp)); + auto stream = at::cuda::getCurrentCUDAStream(); + AT_DISPATCH_FLOATING_TYPES(z.type(), "edz_eydz_cuda", ([&] { + edz_eydz_kernel<<>>( + z.data(), + dz.data(), + weight.data(), + bias.data(), + edz.data(), + eydz.data(), + affine, eps, num, chn, sp); + })); + + return {edz, eydz}; +} + +/*********** + * backward + ***********/ + +template +__global__ void backward_kernel(const T *z, const T *dz, const T *var, const T *weight, const T *bias, const T *edz, + const T *eydz, T *dx, bool affine, float eps, int num, int chn, int sp) { + int plane = blockIdx.x; + + T _weight = affine ? abs(weight[plane]) + eps : 1.f; + T _bias = affine ? bias[plane] : 0.f; + T _var = var[plane]; + T _edz = edz[plane]; + T _eydz = eydz[plane]; + + T _mul = _weight * rsqrt(_var + eps); + T count = T(num * sp); + + for (int batch = 0; batch < num; ++batch) { + for (int n = threadIdx.x; n < sp; n += blockDim.x) { + T _dz = dz[(batch * chn + plane) * sp + n]; + T _y = (z[(batch * chn + plane) * sp + n] - _bias) / _weight; + + dx[(batch * chn + plane) * sp + n] = (_dz - _edz / count - _y * _eydz / count) * _mul; + } + } +} + +at::Tensor backward_cuda(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias, + at::Tensor edz, at::Tensor eydz, bool affine, float eps) { + CHECK_CUDA_INPUT(z); + CHECK_CUDA_INPUT(dz); + CHECK_CUDA_INPUT(var); + CHECK_CUDA_INPUT(weight); + CHECK_CUDA_INPUT(bias); + CHECK_CUDA_INPUT(edz); + CHECK_CUDA_INPUT(eydz); + + // Extract dimensions + int64_t num, chn, sp; + get_dims(z, num, chn, sp); + + auto dx = at::zeros_like(z); + + // Run kernel + dim3 blocks(chn); + dim3 threads(getNumThreads(sp)); + auto stream = at::cuda::getCurrentCUDAStream(); + AT_DISPATCH_FLOATING_TYPES(z.type(), "backward_cuda", ([&] { + backward_kernel<<>>( + z.data(), + dz.data(), + var.data(), + weight.data(), + bias.data(), + edz.data(), + eydz.data(), + dx.data(), + affine, eps, num, chn, sp); + })); + + return dx; +} + +/************** + * activations + **************/ + +template +inline void leaky_relu_backward_impl(T *z, T *dz, float slope, int64_t count) { + // Create thrust pointers + thrust::device_ptr th_z = thrust::device_pointer_cast(z); + thrust::device_ptr th_dz = thrust::device_pointer_cast(dz); + + auto stream = at::cuda::getCurrentCUDAStream(); + thrust::transform_if(thrust::cuda::par.on(stream), + th_dz, th_dz + count, th_z, th_dz, + [slope] __device__ (const T& dz) { return dz * slope; }, + [] __device__ (const T& z) { return z < 0; }); + thrust::transform_if(thrust::cuda::par.on(stream), + th_z, th_z + count, th_z, + [slope] __device__ (const T& z) { return z / slope; }, + [] __device__ (const T& z) { return z < 0; }); +} + +void leaky_relu_backward_cuda(at::Tensor z, at::Tensor dz, float slope) { + CHECK_CUDA_INPUT(z); + CHECK_CUDA_INPUT(dz); + + int64_t count = z.numel(); + + AT_DISPATCH_FLOATING_TYPES(z.type(), "leaky_relu_backward_cuda", ([&] { + leaky_relu_backward_impl(z.data(), dz.data(), slope, count); + })); +} + +template +inline void elu_backward_impl(T *z, T *dz, int64_t count) { + // Create thrust pointers + thrust::device_ptr th_z = thrust::device_pointer_cast(z); + thrust::device_ptr th_dz = thrust::device_pointer_cast(dz); + + auto stream = at::cuda::getCurrentCUDAStream(); + thrust::transform_if(thrust::cuda::par.on(stream), + th_dz, th_dz + count, th_z, th_z, th_dz, + [] __device__ (const T& dz, const T& z) { return dz * (z + 1.); }, + [] __device__ (const T& z) { return z < 0; }); + thrust::transform_if(thrust::cuda::par.on(stream), + th_z, th_z + count, th_z, + [] __device__ (const T& z) { return log1p(z); }, + [] __device__ (const T& z) { return z < 0; }); +} + +void elu_backward_cuda(at::Tensor z, at::Tensor dz) { + CHECK_CUDA_INPUT(z); + CHECK_CUDA_INPUT(dz); + + int64_t count = z.numel(); + + AT_DISPATCH_FLOATING_TYPES(z.type(), "leaky_relu_backward_cuda", ([&] { + elu_backward_impl(z.data(), dz.data(), count); + })); +} diff --git a/modules/src/inplace_abn_cuda_half.cu b/modules/src/inplace_abn_cuda_half.cu new file mode 100644 index 0000000..bb63e73 --- /dev/null +++ b/modules/src/inplace_abn_cuda_half.cu @@ -0,0 +1,275 @@ +#include + +#include + +#include + +#include "utils/checks.h" +#include "utils/cuda.cuh" +#include "inplace_abn.h" + +#include + +// Operations for reduce +struct SumOpH { + __device__ SumOpH(const half *t, int c, int s) + : tensor(t), chn(c), sp(s) {} + __device__ __forceinline__ float operator()(int batch, int plane, int n) { + return __half2float(tensor[(batch * chn + plane) * sp + n]); + } + const half *tensor; + const int chn; + const int sp; +}; + +struct VarOpH { + __device__ VarOpH(float m, const half *t, int c, int s) + : mean(m), tensor(t), chn(c), sp(s) {} + __device__ __forceinline__ float operator()(int batch, int plane, int n) { + const auto t = __half2float(tensor[(batch * chn + plane) * sp + n]); + return (t - mean) * (t - mean); + } + const float mean; + const half *tensor; + const int chn; + const int sp; +}; + +struct GradOpH { + __device__ GradOpH(float _weight, float _bias, const half *_z, const half *_dz, int c, int s) + : weight(_weight), bias(_bias), z(_z), dz(_dz), chn(c), sp(s) {} + __device__ __forceinline__ Pair operator()(int batch, int plane, int n) { + float _y = (__half2float(z[(batch * chn + plane) * sp + n]) - bias) / weight; + float _dz = __half2float(dz[(batch * chn + plane) * sp + n]); + return Pair(_dz, _y * _dz); + } + const float weight; + const float bias; + const half *z; + const half *dz; + const int chn; + const int sp; +}; + +/*********** + * mean_var + ***********/ + +__global__ void mean_var_kernel_h(const half *x, float *mean, float *var, int num, int chn, int sp) { + int plane = blockIdx.x; + float norm = 1.f / static_cast(num * sp); + + float _mean = reduce(SumOpH(x, chn, sp), plane, num, sp) * norm; + __syncthreads(); + float _var = reduce(VarOpH(_mean, x, chn, sp), plane, num, sp) * norm; + + if (threadIdx.x == 0) { + mean[plane] = _mean; + var[plane] = _var; + } +} + +std::vector mean_var_cuda_h(at::Tensor x) { + CHECK_CUDA_INPUT(x); + + // Extract dimensions + int64_t num, chn, sp; + get_dims(x, num, chn, sp); + + // Prepare output tensors + auto mean = at::empty({chn},x.options().dtype(at::kFloat)); + auto var = at::empty({chn},x.options().dtype(at::kFloat)); + + // Run kernel + dim3 blocks(chn); + dim3 threads(getNumThreads(sp)); + auto stream = at::cuda::getCurrentCUDAStream(); + mean_var_kernel_h<<>>( + reinterpret_cast(x.data()), + mean.data(), + var.data(), + num, chn, sp); + + return {mean, var}; +} + +/********** + * forward + **********/ + +__global__ void forward_kernel_h(half *x, const float *mean, const float *var, const float *weight, const float *bias, + bool affine, float eps, int num, int chn, int sp) { + int plane = blockIdx.x; + + const float _mean = mean[plane]; + const float _var = var[plane]; + const float _weight = affine ? abs(weight[plane]) + eps : 1.f; + const float _bias = affine ? bias[plane] : 0.f; + + const float mul = rsqrt(_var + eps) * _weight; + + for (int batch = 0; batch < num; ++batch) { + for (int n = threadIdx.x; n < sp; n += blockDim.x) { + half *x_ptr = x + (batch * chn + plane) * sp + n; + float _x = __half2float(*x_ptr); + float _y = (_x - _mean) * mul + _bias; + + *x_ptr = __float2half(_y); + } + } +} + +at::Tensor forward_cuda_h(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias, + bool affine, float eps) { + CHECK_CUDA_INPUT(x); + CHECK_CUDA_INPUT(mean); + CHECK_CUDA_INPUT(var); + CHECK_CUDA_INPUT(weight); + CHECK_CUDA_INPUT(bias); + + // Extract dimensions + int64_t num, chn, sp; + get_dims(x, num, chn, sp); + + // Run kernel + dim3 blocks(chn); + dim3 threads(getNumThreads(sp)); + auto stream = at::cuda::getCurrentCUDAStream(); + forward_kernel_h<<>>( + reinterpret_cast(x.data()), + mean.data(), + var.data(), + weight.data(), + bias.data(), + affine, eps, num, chn, sp); + + return x; +} + +__global__ void edz_eydz_kernel_h(const half *z, const half *dz, const float *weight, const float *bias, + float *edz, float *eydz, bool affine, float eps, int num, int chn, int sp) { + int plane = blockIdx.x; + + float _weight = affine ? abs(weight[plane]) + eps : 1.f; + float _bias = affine ? bias[plane] : 0.f; + + Pair res = reduce, GradOpH>(GradOpH(_weight, _bias, z, dz, chn, sp), plane, num, sp); + __syncthreads(); + + if (threadIdx.x == 0) { + edz[plane] = res.v1; + eydz[plane] = res.v2; + } +} + +std::vector edz_eydz_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias, + bool affine, float eps) { + CHECK_CUDA_INPUT(z); + CHECK_CUDA_INPUT(dz); + CHECK_CUDA_INPUT(weight); + CHECK_CUDA_INPUT(bias); + + // Extract dimensions + int64_t num, chn, sp; + get_dims(z, num, chn, sp); + + auto edz = at::empty({chn},z.options().dtype(at::kFloat)); + auto eydz = at::empty({chn},z.options().dtype(at::kFloat)); + + // Run kernel + dim3 blocks(chn); + dim3 threads(getNumThreads(sp)); + auto stream = at::cuda::getCurrentCUDAStream(); + edz_eydz_kernel_h<<>>( + reinterpret_cast(z.data()), + reinterpret_cast(dz.data()), + weight.data(), + bias.data(), + edz.data(), + eydz.data(), + affine, eps, num, chn, sp); + + return {edz, eydz}; +} + +__global__ void backward_kernel_h(const half *z, const half *dz, const float *var, const float *weight, const float *bias, const float *edz, + const float *eydz, half *dx, bool affine, float eps, int num, int chn, int sp) { + int plane = blockIdx.x; + + float _weight = affine ? abs(weight[plane]) + eps : 1.f; + float _bias = affine ? bias[plane] : 0.f; + float _var = var[plane]; + float _edz = edz[plane]; + float _eydz = eydz[plane]; + + float _mul = _weight * rsqrt(_var + eps); + float count = float(num * sp); + + for (int batch = 0; batch < num; ++batch) { + for (int n = threadIdx.x; n < sp; n += blockDim.x) { + float _dz = __half2float(dz[(batch * chn + plane) * sp + n]); + float _y = (__half2float(z[(batch * chn + plane) * sp + n]) - _bias) / _weight; + + dx[(batch * chn + plane) * sp + n] = __float2half((_dz - _edz / count - _y * _eydz / count) * _mul); + } + } +} + +at::Tensor backward_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias, + at::Tensor edz, at::Tensor eydz, bool affine, float eps) { + CHECK_CUDA_INPUT(z); + CHECK_CUDA_INPUT(dz); + CHECK_CUDA_INPUT(var); + CHECK_CUDA_INPUT(weight); + CHECK_CUDA_INPUT(bias); + CHECK_CUDA_INPUT(edz); + CHECK_CUDA_INPUT(eydz); + + // Extract dimensions + int64_t num, chn, sp; + get_dims(z, num, chn, sp); + + auto dx = at::zeros_like(z); + + // Run kernel + dim3 blocks(chn); + dim3 threads(getNumThreads(sp)); + auto stream = at::cuda::getCurrentCUDAStream(); + backward_kernel_h<<>>( + reinterpret_cast(z.data()), + reinterpret_cast(dz.data()), + var.data(), + weight.data(), + bias.data(), + edz.data(), + eydz.data(), + reinterpret_cast(dx.data()), + affine, eps, num, chn, sp); + + return dx; +} + +__global__ void leaky_relu_backward_impl_h(half *z, half *dz, float slope, int64_t count) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < count; i += blockDim.x * gridDim.x){ + float _z = __half2float(z[i]); + if (_z < 0) { + dz[i] = __float2half(__half2float(dz[i]) * slope); + z[i] = __float2half(_z / slope); + } + } +} + +void leaky_relu_backward_cuda_h(at::Tensor z, at::Tensor dz, float slope) { + CHECK_CUDA_INPUT(z); + CHECK_CUDA_INPUT(dz); + + int64_t count = z.numel(); + dim3 threads(getNumThreads(count)); + dim3 blocks = (count + threads.x - 1) / threads.x; + auto stream = at::cuda::getCurrentCUDAStream(); + leaky_relu_backward_impl_h<<>>( + reinterpret_cast(z.data()), + reinterpret_cast(dz.data()), + slope, count); +} + diff --git a/modules/src/utils/checks.h b/modules/src/utils/checks.h new file mode 100644 index 0000000..e761a6f --- /dev/null +++ b/modules/src/utils/checks.h @@ -0,0 +1,15 @@ +#pragma once + +#include + +// Define AT_CHECK for old version of ATen where the same function was called AT_ASSERT +#ifndef AT_CHECK +#define AT_CHECK AT_ASSERT +#endif + +#define CHECK_CUDA(x) AT_CHECK((x).type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CPU(x) AT_CHECK(!(x).type().is_cuda(), #x " must be a CPU tensor") +#define CHECK_CONTIGUOUS(x) AT_CHECK((x).is_contiguous(), #x " must be contiguous") + +#define CHECK_CUDA_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) +#define CHECK_CPU_INPUT(x) CHECK_CPU(x); CHECK_CONTIGUOUS(x) \ No newline at end of file diff --git a/modules/src/utils/common.h b/modules/src/utils/common.h new file mode 100644 index 0000000..e8403ee --- /dev/null +++ b/modules/src/utils/common.h @@ -0,0 +1,49 @@ +#pragma once + +#include + +/* + * Functions to share code between CPU and GPU + */ + +#ifdef __CUDACC__ +// CUDA versions + +#define HOST_DEVICE __host__ __device__ +#define INLINE_HOST_DEVICE __host__ __device__ inline +#define FLOOR(x) floor(x) + +#if __CUDA_ARCH__ >= 600 +// Recent compute capabilities have block-level atomicAdd for all data types, so we use that +#define ACCUM(x,y) atomicAdd_block(&(x),(y)) +#else +// Older architectures don't have block-level atomicAdd, nor atomicAdd for doubles, so we defer to atomicAdd for float +// and use the known atomicCAS-based implementation for double +template +__device__ inline data_t atomic_add(data_t *address, data_t val) { + return atomicAdd(address, val); +} + +template<> +__device__ inline double atomic_add(double *address, double val) { + unsigned long long int* address_as_ull = (unsigned long long int*)address; + unsigned long long int old = *address_as_ull, assumed; + do { + assumed = old; + old = atomicCAS(address_as_ull, assumed, __double_as_longlong(val + __longlong_as_double(assumed))); + } while (assumed != old); + return __longlong_as_double(old); +} + +#define ACCUM(x,y) atomic_add(&(x),(y)) +#endif // #if __CUDA_ARCH__ >= 600 + +#else +// CPU versions + +#define HOST_DEVICE +#define INLINE_HOST_DEVICE inline +#define FLOOR(x) std::floor(x) +#define ACCUM(x,y) (x) += (y) + +#endif // #ifdef __CUDACC__ \ No newline at end of file diff --git a/modules/src/utils/cuda.cuh b/modules/src/utils/cuda.cuh new file mode 100644 index 0000000..60c0023 --- /dev/null +++ b/modules/src/utils/cuda.cuh @@ -0,0 +1,71 @@ +#pragma once + +/* + * General settings and functions + */ +const int WARP_SIZE = 32; +const int MAX_BLOCK_SIZE = 1024; + +static int getNumThreads(int nElem) { + int threadSizes[6] = {32, 64, 128, 256, 512, MAX_BLOCK_SIZE}; + for (int i = 0; i < 6; ++i) { + if (nElem <= threadSizes[i]) { + return threadSizes[i]; + } + } + return MAX_BLOCK_SIZE; +} + +/* + * Reduction utilities + */ +template +__device__ __forceinline__ T WARP_SHFL_XOR(T value, int laneMask, int width = warpSize, + unsigned int mask = 0xffffffff) { +#if CUDART_VERSION >= 9000 + return __shfl_xor_sync(mask, value, laneMask, width); +#else + return __shfl_xor(value, laneMask, width); +#endif +} + +__device__ __forceinline__ int getMSB(int val) { return 31 - __clz(val); } + +template +struct Pair { + T v1, v2; + __device__ Pair() {} + __device__ Pair(T _v1, T _v2) : v1(_v1), v2(_v2) {} + __device__ Pair(T v) : v1(v), v2(v) {} + __device__ Pair(int v) : v1(v), v2(v) {} + __device__ Pair &operator+=(const Pair &a) { + v1 += a.v1; + v2 += a.v2; + return *this; + } +}; + +template +static __device__ __forceinline__ T warpSum(T val) { +#if __CUDA_ARCH__ >= 300 + for (int i = 0; i < getMSB(WARP_SIZE); ++i) { + val += WARP_SHFL_XOR(val, 1 << i, WARP_SIZE); + } +#else + __shared__ T values[MAX_BLOCK_SIZE]; + values[threadIdx.x] = val; + __threadfence_block(); + const int base = (threadIdx.x / WARP_SIZE) * WARP_SIZE; + for (int i = 1; i < WARP_SIZE; i++) { + val += values[base + ((i + threadIdx.x) % WARP_SIZE)]; + } +#endif + return val; +} + +template +static __device__ __forceinline__ Pair warpSum(Pair value) { + value.v1 = warpSum(value.v1); + value.v2 = warpSum(value.v2); + return value; +} \ No newline at end of file diff --git a/nets/__init__.py b/nets/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/nets/__pycache__/__init__.cpython-37.pyc b/nets/__pycache__/__init__.cpython-37.pyc new file mode 100644 index 0000000..181ba39 Binary files /dev/null and b/nets/__pycache__/__init__.cpython-37.pyc differ diff --git a/nets/__pycache__/__init__.cpython-38.pyc b/nets/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000..5c32dae Binary files /dev/null and b/nets/__pycache__/__init__.cpython-38.pyc differ diff --git a/nets/__pycache__/stdcnet.cpython-37.pyc b/nets/__pycache__/stdcnet.cpython-37.pyc new file mode 100644 index 0000000..38d1de6 Binary files /dev/null and b/nets/__pycache__/stdcnet.cpython-37.pyc differ diff --git a/nets/__pycache__/stdcnet.cpython-38.pyc b/nets/__pycache__/stdcnet.cpython-38.pyc new file mode 100644 index 0000000..2f3c1e0 Binary files /dev/null and b/nets/__pycache__/stdcnet.cpython-38.pyc differ diff --git a/nets/cat.pth b/nets/cat.pth new file mode 100644 index 0000000..fb15d63 Binary files /dev/null and b/nets/cat.pth differ diff --git a/nets/stdcnet.py b/nets/stdcnet.py new file mode 100644 index 0000000..f9a9271 --- /dev/null +++ b/nets/stdcnet.py @@ -0,0 +1,306 @@ +import torch +import torch.nn as nn +from torch.nn import init +import math + + + +class ConvX(nn.Module): + def __init__(self, in_planes, out_planes, kernel=3, stride=1): + super(ConvX, self).__init__() + self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=kernel, stride=stride, padding=kernel//2, bias=False) + self.bn = nn.BatchNorm2d(out_planes) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + out = self.relu(self.bn(self.conv(x))) + return out + + +class AddBottleneck(nn.Module): + def __init__(self, in_planes, out_planes, block_num=3, stride=1): + super(AddBottleneck, self).__init__() + assert block_num > 1, print("block number should be larger than 1.") + self.conv_list = nn.ModuleList() + self.stride = stride + if stride == 2: + self.avd_layer = nn.Sequential( + nn.Conv2d(out_planes//2, out_planes//2, kernel_size=3, stride=2, padding=1, groups=out_planes//2, bias=False), + nn.BatchNorm2d(out_planes//2), + ) + self.skip = nn.Sequential( + nn.Conv2d(in_planes, in_planes, kernel_size=3, stride=2, padding=1, groups=in_planes, bias=False), + nn.BatchNorm2d(in_planes), + nn.Conv2d(in_planes, out_planes, kernel_size=1, bias=False), + nn.BatchNorm2d(out_planes), + ) + stride = 1 + + for idx in range(block_num): + if idx == 0: + self.conv_list.append(ConvX(in_planes, out_planes//2, kernel=1)) + elif idx == 1 and block_num == 2: + self.conv_list.append(ConvX(out_planes//2, out_planes//2, stride=stride)) + elif idx == 1 and block_num > 2: + self.conv_list.append(ConvX(out_planes//2, out_planes//4, stride=stride)) + elif idx < block_num - 1: + self.conv_list.append(ConvX(out_planes//int(math.pow(2, idx)), out_planes//int(math.pow(2, idx+1)))) + else: + self.conv_list.append(ConvX(out_planes//int(math.pow(2, idx)), out_planes//int(math.pow(2, idx)))) + + def forward(self, x): + out_list = [] + out = x + + for idx, conv in enumerate(self.conv_list): + if idx == 0 and self.stride == 2: + out = self.avd_layer(conv(out)) + else: + out = conv(out) + out_list.append(out) + + if self.stride == 2: + x = self.skip(x) + + return torch.cat(out_list, dim=1) + x + + + +class CatBottleneck(nn.Module): + def __init__(self, in_planes, out_planes, block_num=3, stride=1): + super(CatBottleneck, self).__init__() + assert block_num > 1, print("block number should be larger than 1.") + self.conv_list = nn.ModuleList() + self.stride = stride + if stride == 2: + self.avd_layer = nn.Sequential( + nn.Conv2d(out_planes//2, out_planes//2, kernel_size=3, stride=2, padding=1, groups=out_planes//2, bias=False), + nn.BatchNorm2d(out_planes//2), + ) + self.skip = nn.AvgPool2d(kernel_size=3, stride=2, padding=1) + stride = 1 + + for idx in range(block_num): + if idx == 0: + self.conv_list.append(ConvX(in_planes, out_planes//2, kernel=1)) + elif idx == 1 and block_num == 2: + self.conv_list.append(ConvX(out_planes//2, out_planes//2, stride=stride)) + elif idx == 1 and block_num > 2: + self.conv_list.append(ConvX(out_planes//2, out_planes//4, stride=stride)) + elif idx < block_num - 1: + self.conv_list.append(ConvX(out_planes//int(math.pow(2, idx)), out_planes//int(math.pow(2, idx+1)))) + else: + self.conv_list.append(ConvX(out_planes//int(math.pow(2, idx)), out_planes//int(math.pow(2, idx)))) + + def forward(self, x): + out_list = [] + out1 = self.conv_list[0](x) + + for idx, conv in enumerate(self.conv_list[1:]): + if idx == 0: + if self.stride == 2: + out = conv(self.avd_layer(out1)) + else: + out = conv(out1) + else: + out = conv(out) + out_list.append(out) + + if self.stride == 2: + out1 = self.skip(out1) + out_list.insert(0, out1) + + out = torch.cat(out_list, dim=1) + return out + +#STDC2Net +class STDCNet1446(nn.Module): + def __init__(self, base=64, layers=[4,5,3], block_num=4, type="cat", num_classes=1000, dropout=0.20, pretrain_model='', use_conv_last=False): + super(STDCNet1446, self).__init__() + if type == "cat": + block = CatBottleneck + elif type == "add": + block = AddBottleneck + self.use_conv_last = use_conv_last + self.features = self._make_layers(base, layers, block_num, block) + self.conv_last = ConvX(base*16, max(1024, base*16), 1, 1) + self.gap = nn.AdaptiveAvgPool2d(1) + self.fc = nn.Linear(max(1024, base*16), max(1024, base*16), bias=False) + self.bn = nn.BatchNorm1d(max(1024, base*16)) + self.relu = nn.ReLU(inplace=True) + self.dropout = nn.Dropout(p=dropout) + self.linear = nn.Linear(max(1024, base*16), num_classes, bias=False) + + self.x2 = nn.Sequential(self.features[:1]) + self.x4 = nn.Sequential(self.features[1:2]) + self.x8 = nn.Sequential(self.features[2:6]) + self.x16 = nn.Sequential(self.features[6:11]) + self.x32 = nn.Sequential(self.features[11:]) + + if pretrain_model: + print('use pretrain model {}'.format(pretrain_model)) + self.init_weight(pretrain_model) + else: + self.init_params() + + def init_weight(self, pretrain_model): + + state_dict = torch.load(pretrain_model)["state_dict"] + self_state_dict = self.state_dict() + for k, v in state_dict.items(): + self_state_dict.update({k: v}) + self.load_state_dict(self_state_dict) + + def init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + init.kaiming_normal_(m.weight, mode='fan_out') + if m.bias is not None: + init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + init.constant_(m.weight, 1) + init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + init.normal_(m.weight, std=0.001) + if m.bias is not None: + init.constant_(m.bias, 0) + + def _make_layers(self, base, layers, block_num, block): + features = [] + features += [ConvX(3, base//2, 3, 2)] + features += [ConvX(base//2, base, 3, 2)] + + for i, layer in enumerate(layers): + for j in range(layer): + if i == 0 and j == 0: + features.append(block(base, base*4, block_num, 2)) + elif j == 0: + features.append(block(base*int(math.pow(2,i+1)), base*int(math.pow(2,i+2)), block_num, 2)) + else: + features.append(block(base*int(math.pow(2,i+2)), base*int(math.pow(2,i+2)), block_num, 1)) + + return nn.Sequential(*features) + + def forward(self, x): + feat2 = self.x2(x) + feat4 = self.x4(feat2) + feat8 = self.x8(feat4) + feat16 = self.x16(feat8) + feat32 = self.x32(feat16) + if self.use_conv_last: + feat32 = self.conv_last(feat32) + + return feat2, feat4, feat8, feat16, feat32 + + def forward_impl(self, x): + out = self.features(x) + out = self.conv_last(out).pow(2) + out = self.gap(out).flatten(1) + out = self.fc(out) + # out = self.bn(out) + out = self.relu(out) + # out = self.relu(self.bn(self.fc(out))) + out = self.dropout(out) + out = self.linear(out) + return out + +# STDC1Net +class STDCNet813(nn.Module): + def __init__(self, base=64, layers=[2,2,2], block_num=4, type="cat", num_classes=1000, dropout=0.20, pretrain_model='', use_conv_last=False): + super(STDCNet813, self).__init__() + if type == "cat": + block = CatBottleneck + elif type == "add": + block = AddBottleneck + self.use_conv_last = use_conv_last + self.features = self._make_layers(base, layers, block_num, block) + self.conv_last = ConvX(base*16, max(1024, base*16), 1, 1) + self.gap = nn.AdaptiveAvgPool2d(1) + self.fc = nn.Linear(max(1024, base*16), max(1024, base*16), bias=False) + self.bn = nn.BatchNorm1d(max(1024, base*16)) + self.relu = nn.ReLU(inplace=True) + self.dropout = nn.Dropout(p=dropout) + self.linear = nn.Linear(max(1024, base*16), num_classes, bias=False) + + self.x2 = nn.Sequential(self.features[:1]) + self.x4 = nn.Sequential(self.features[1:2]) + self.x8 = nn.Sequential(self.features[2:4]) + self.x16 = nn.Sequential(self.features[4:6]) + self.x32 = nn.Sequential(self.features[6:]) + + if pretrain_model: + print('use pretrain model {}'.format(pretrain_model)) + self.init_weight(pretrain_model) + else: + self.init_params() + + def init_weight(self, pretrain_model): + try: + state_dict = torch.load(pretrain_model)["state_dict"] + self_state_dict = self.state_dict() + for k, v in state_dict.items(): + self_state_dict.update({k: v}) + self.load_state_dict(self_state_dict) + except Exception as e: + self.load_state_dict(torch.load(pretrain_model )) + + def init_params(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + init.kaiming_normal_(m.weight, mode='fan_out') + if m.bias is not None: + init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + init.constant_(m.weight, 1) + init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + init.normal_(m.weight, std=0.001) + if m.bias is not None: + init.constant_(m.bias, 0) + + def _make_layers(self, base, layers, block_num, block): + features = [] + features += [ConvX(3, base//2, 3, 2)] + features += [ConvX(base//2, base, 3, 2)] + + for i, layer in enumerate(layers): + for j in range(layer): + if i == 0 and j == 0: + features.append(block(base, base*4, block_num, 2)) + elif j == 0: + features.append(block(base*int(math.pow(2,i+1)), base*int(math.pow(2,i+2)), block_num, 2)) + else: + features.append(block(base*int(math.pow(2,i+2)), base*int(math.pow(2,i+2)), block_num, 1)) + + return nn.Sequential(*features) + + def forward(self, x): + feat2 = self.x2(x) + feat4 = self.x4(feat2) + feat8 = self.x8(feat4) + feat16 = self.x16(feat8) + feat32 = self.x32(feat16) + if self.use_conv_last: + feat32 = self.conv_last(feat32) + + return feat2, feat4, feat8, feat16, feat32 + + def forward_impl(self, x): + out = self.features(x) + out = self.conv_last(out).pow(2) + out = self.gap(out).flatten(1) + out = self.fc(out) + # out = self.bn(out) + out = self.relu(out) + # out = self.relu(self.bn(self.fc(out))) + out = self.dropout(out) + out = self.linear(out) + return out + +if __name__ == "__main__": + model = STDCNet813(num_classes=1000, dropout=0.00, block_num=4) + model.eval() + x = torch.randn(1,3,224,224) + y = model(x) + torch.save(model.state_dict(), 'cat.pth') + print(y.size()) diff --git a/optimizer_loss.py b/optimizer_loss.py new file mode 100644 index 0000000..966edea --- /dev/null +++ b/optimizer_loss.py @@ -0,0 +1,76 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- + +import torch +import logging + +logger = logging.getLogger() + +class Optimizer(object): + def __init__(self, + model, + loss, + lr0, + momentum, + wd, + warmup_steps, + warmup_start_lr, + max_iter, + power, + *args, **kwargs): + self.warmup_steps = warmup_steps + self.warmup_start_lr = warmup_start_lr + self.lr0 = lr0 + self.lr = self.lr0 + self.max_iter = float(max_iter) + self.power = power + self.it = 0 + wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params = model.get_params() + loss_nowd_params = loss.get_params() + # print(wd_params) + # print(nowd_params) + # print(loss_nowd_params) + # exit(0) + param_list = [ + {'params': wd_params}, + {'params': nowd_params, 'weight_decay': 0}, + {'params': lr_mul_wd_params, 'lr_mul': True}, + {'params': lr_mul_nowd_params, 'weight_decay': 0, 'lr_mul': True}, + {'params': loss_nowd_params}] + # {'params': loss_nowd_params, 'weight_decay': 0, 'lr': 0.000001}] + self.optim = torch.optim.SGD( + param_list, + lr = lr0, + momentum = momentum, + weight_decay = wd) + self.warmup_factor = (self.lr0/self.warmup_start_lr)**(1./self.warmup_steps) + + + def get_lr(self): + if self.it <= self.warmup_steps: + lr = self.warmup_start_lr*(self.warmup_factor**self.it) + else: + factor = (1-(self.it-self.warmup_steps)/(self.max_iter-self.warmup_steps))**self.power + lr = self.lr0 * factor + return lr + + + def step(self): + self.lr = self.get_lr() + for pg in self.optim.param_groups: + if pg.get('lr_mul', False): + pg['lr'] = self.lr * 10 + else: + pg['lr'] = self.lr + if self.optim.defaults.get('lr_mul', False): + self.optim.defaults['lr'] = self.lr * 10 + else: + self.optim.defaults['lr'] = self.lr + self.it += 1 + self.optim.step() + if self.it == self.warmup_steps+2: + logger.info('==> warmup done, start to implement poly lr strategy') + + def zero_grad(self): + self.optim.zero_grad() + diff --git a/run.sh b/run.sh new file mode 100644 index 0000000..9c50e14 --- /dev/null +++ b/run.sh @@ -0,0 +1,3 @@ + +#python train.py --parJson ./data/RoadLane.json --respath ./checkpooints/0430pm --gpuId 0 +python train.py --parJson ./data/RoadLane.json --gpuId 0 diff --git a/scripts/cityscaptes_train_STDC1Seg.sh b/scripts/cityscaptes_train_STDC1Seg.sh new file mode 100644 index 0000000..92bb0e3 --- /dev/null +++ b/scripts/cityscaptes_train_STDC1Seg.sh @@ -0,0 +1,11 @@ +export CUDA_VISIBLE_DEVICES=0,1,2 +python -m torch.distributed.launch \ +--nproc_per_node=3 train.py \ +--respath checkpoints/train_STDC1-Seg/ \ +--backbone STDCNet813 \ +--mode train \ +--n_workers_train 12 \ +--n_workers_val 1 \ +--max_iter 60000 \ +--use_boundary_8 True \ +--pretrain_path checkpoints/STDCNet813M_73.91.tar diff --git a/scripts/cityscaptes_train_STDC2Seg.sh b/scripts/cityscaptes_train_STDC2Seg.sh new file mode 100644 index 0000000..cb6332e --- /dev/null +++ b/scripts/cityscaptes_train_STDC2Seg.sh @@ -0,0 +1,11 @@ +export CUDA_VISIBLE_DEVICES=0,1,2 +python -m torch.distributed.launch \ +--nproc_per_node=3 train.py \ +--respath checkpoints/train_STDC2-Seg/ \ +--backbone STDCNet1446 \ +--mode train \ +--n_workers_train 12 \ +--n_workers_val 1 \ +--max_iter 60000 \ +--use_boundary_8 True \ +--pretrain_path checkpoints/STDCNet1446_76.47.tar diff --git a/train.py b/train.py new file mode 100644 index 0000000..12fcfd2 --- /dev/null +++ b/train.py @@ -0,0 +1,348 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- +import os +#os.environ['CUDA_VISIBLE_DEVICES'] = '1' +from logger import setup_logger +from models.model_stages import BiSeNet +# from heliushuju import Heliushuju +from heliushuju_process import Heliushuju +from loss.loss import OhemCELoss +from loss.detail_loss import DetailAggregateLoss +# from evaluation import MscEvalV0 +from evaluation_process import MscEvalV0 +from optimizer_loss import Optimizer +import sys +import torch +import torch.nn as nn +from torch.utils.data import DataLoader +import torch.nn.functional as F +import torch.distributed as dist + + +import os.path as osp +import logging +import time +import datetime +import argparse +import json +logger = logging.getLogger() + +def str2bool(v): + if v.lower() in ('yes', 'true', 't', 'y', '1'): + return True + elif v.lower() in ('no', 'false', 'f', 'n', '0'): + return False + else: + raise argparse.ArgumentTypeError('Unsupported value encountered.') + + +def parse_args(): + parse = argparse.ArgumentParser() + parse.add_argument('--parJson',dest = 'parJson',type = str,default = './data/Car',) + parse.add_argument('--gpuId',dest = 'gpuId',type = str,default = '0',) + parse.add_argument('--local_rank',dest = 'local_rank',type = int,default = -1 ) + parse.add_argument('--n_workers_train',dest = 'n_workers_train',type = int,default = 8) + parse.add_argument('--n_workers_val',dest = 'n_workers_val',type = int,default = 2) + parse.add_argument('--n_img_per_gpu',dest = 'n_img_per_gpu',type = int,default = 8,) + parse.add_argument('--max_iter',dest = 'max_iter',type = int,default = 43000) + parse.add_argument('--save_iter_sep',dest = 'save_iter_sep',type = int,default = 1000,) + parse.add_argument('--warmup_steps',dest = 'warmup_steps',type = int,default = 1000,) + parse.add_argument('--mode',dest = 'mode',type = str,default = 'train',) + parse.add_argument('--ckpt',dest = 'ckpt',type = str,default = None,) + parse.add_argument('--respath',dest = 'respath',type = str,default='./model_save') + parse.add_argument('--backbone',dest = 'backbone',type = str,default = 'STDCNet813') + parse.add_argument('--pretrain_path',dest = 'pretrain_path',type = str,default='./checkpoints2/STDCNet813M_73.91.tar',) + parse.add_argument('--use_conv_last',dest = 'use_conv_last',type = str2bool,default = False,) + parse.add_argument('--use_boundary_2',dest = 'use_boundary_2',type = str2bool,default = False,) + parse.add_argument('--use_boundary_4',dest = 'use_boundary_4',type = str2bool,default = False,) + parse.add_argument('--use_boundary_8',dest = 'use_boundary_8',type = str2bool,default = True) + parse.add_argument('--use_boundary_16',dest = 'use_boundary_16',type = str2bool,default = False,) + return parse.parse_args() + + +def train(): + + + + args = parse_args() + with open(args.parJson.strip(),'r') as fp: + par=json.load(fp) + #print('args.gpuId.strip :',args.gpuId.strip()) + #os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpuId.strip()) + + + torch.cuda.set_device(int(args.gpuId.strip())) + #print('line72:',torch.cuda.current_device()) + save_pth_path = os.path.join(args.respath, 'pths') + #dspth = '../../data/CarRoadLane/' + dspth = par['dspth'] + + if not osp.exists(save_pth_path): + os.makedirs(save_pth_path) + + #torch.cuda.set_device(args.local_rank) + ########################################################################fenbushi + + setup_logger(args.respath) + + #n_classes = 4 # 改动 + n_classes = par['n_classes'] + + n_img_per_gpu = args.n_img_per_gpu + n_workers_train = args.n_workers_train + n_workers_val = args.n_workers_val + use_boundary_16 = args.use_boundary_16 + use_boundary_8 = args.use_boundary_8 + use_boundary_4 = args.use_boundary_4 + use_boundary_2 = args.use_boundary_2 + + mode = args.mode # train + #cropsize = [640, 360] + cropsize = eval(par['cropsize']) + #labelJson='./heliushuju_info.json' + labelJson = par['labelJson'] + randomscale = (0.125, 0.25, 0.375, 0.5, 0.625, 0.75, 0.875, 1.0, 1.125, 1.25, 1.375, 1.5) + #print('line102:',torch.cuda.current_device()) + + + ds = Heliushuju(dspth, cropsize=cropsize, mode=mode,labelJson=labelJson, randomscale=randomscale) + sampler = None + # #################################################################################################fenbushi + # sampler = torch.utils.data.distributed.DistributedSampler(ds) + dl = DataLoader(ds, + batch_size = n_img_per_gpu, + shuffle = False, + sampler = sampler, + num_workers = n_workers_train, + pin_memory = False, + drop_last = True) + dsval = Heliushuju(dspth, mode='val',labelJson=labelJson, randomscale=randomscale) + sampler_val = None + dlval = DataLoader(dsval, + batch_size = 1, + shuffle = False, + sampler = sampler_val, + num_workers = n_workers_val, + drop_last = False) + #print('line124:',torch.cuda.current_device()) + ## model + #ignore_idx = 255 + ignore_idx = par['ignore_idx'] + net = BiSeNet(backbone=args.backbone, n_classes=n_classes, pretrain_model=args.pretrain_path, + use_boundary_2=use_boundary_2, use_boundary_4=use_boundary_4, use_boundary_8=use_boundary_8, + use_boundary_16=use_boundary_16, use_conv_last=args.use_conv_last) + + if not args.ckpt is None: + net.load_state_dict(torch.load(args.ckpt, map_location='cpu')) + print('load checkpoints model:',args.ckpt) + + #print('line136:',torch.cuda.current_device()) + net.cuda() + net.train() + #print('line142:',net.device) + #net = nn.DataParallel(net, device_ids=[0])########################################################################### + #print('line144:',net.device) + score_thres = 0.7 + n_min = n_img_per_gpu*cropsize[0]*cropsize[1]//16 + criteria_p = OhemCELoss(thresh=score_thres, n_min=n_min, ignore_lb=ignore_idx) + criteria_16 = OhemCELoss(thresh=score_thres, n_min=n_min, ignore_lb=ignore_idx) + criteria_32 = OhemCELoss(thresh=score_thres, n_min=n_min, ignore_lb=ignore_idx) + boundary_loss_func = DetailAggregateLoss() + ## optimizer + maxmIOU50 = 0. + maxmIOU75 = 0. + momentum = 0.9 + weight_decay = 5e-4 + lr_start = 1e-2 + max_iter = args.max_iter + save_iter_sep = args.save_iter_sep + power = 0.9 + warmup_steps = args.warmup_steps + warmup_start_lr = 1e-5 + print('max_iter: ', max_iter) + print('save_iter_sep: ', save_iter_sep) + print('warmup_steps: ', warmup_steps) + optim = Optimizer( + #model = net.module, + model = net, + loss = boundary_loss_func, + lr0 = lr_start, + momentum = momentum, + wd = weight_decay, + warmup_steps = warmup_steps, + warmup_start_lr = warmup_start_lr, + max_iter = max_iter, + power = power) + + ## train loop + msg_iter = 50 + loss_avg = [] + loss_boundery_bce = [] + loss_boundery_dice = [] + st = glob_st = time.time() + diter = iter(dl) + # diter = enumerate(dl) + epoch = 0 + for it in range(max_iter): + try: + im, lb = diter.__next__() + # print(im.size()[0]) + # im, lb = next(diter) + if not im.size()[0]==n_img_per_gpu: raise StopIteration + except StopIteration: + epoch += 1 + # sampler.set_epoch(epoch) + diter = iter(dl) + im, lb = next(diter) + im = im.cuda() + lb = lb.cuda() + H, W = im.size()[2:] + + lb = torch.squeeze(lb, 1) # lb.shape : torch.Size([8, 360, 640]) + + # lb = torch.argmax(lb, dim=3) # 添加(训练高速路时，需要添加这行代码，训练水域分割时，将这行代码注释掉) + + optim.zero_grad() + + + if use_boundary_2 and use_boundary_4 and use_boundary_8: + out, out16, out32, detail2, detail4, detail8 = net(im) + + if (not use_boundary_2) and use_boundary_4 and use_boundary_8: + out, out16, out32, detail4, detail8 = net(im) + + if (not use_boundary_2) and (not use_boundary_4) and use_boundary_8:#######True + out, out16, out32, detail8 = net(im) + + if (not use_boundary_2) and (not use_boundary_4) and (not use_boundary_8): + out, out16, out32 = net(im) + + lossp = criteria_p(out, lb) + loss2 = criteria_16(out16, lb) + loss3 = criteria_32(out32, lb) + #print(' line220 : img.size:', H,W,'label.size：',lb.size(),' pred size:',out.size() ) + boundery_bce_loss = 0. + boundery_dice_loss = 0. + + if use_boundary_2: + boundery_bce_loss2, boundery_dice_loss2 = boundary_loss_func(detail2, lb) + boundery_bce_loss += boundery_bce_loss2 + boundery_dice_loss += boundery_dice_loss2 + + if use_boundary_4: + boundery_bce_loss4, boundery_dice_loss4 = boundary_loss_func(detail4, lb) + boundery_bce_loss += boundery_bce_loss4 + boundery_dice_loss += boundery_dice_loss4 + + if use_boundary_8:###### + boundery_bce_loss8, boundery_dice_loss8 = boundary_loss_func(detail8, lb) + boundery_bce_loss += boundery_bce_loss8 + boundery_dice_loss += boundery_dice_loss8 + + loss = lossp + loss2 + loss3 + boundery_bce_loss + boundery_dice_loss + + loss.backward() + optim.step() + + loss_avg.append(loss.item()) + + loss_boundery_bce.append(boundery_bce_loss.item()) + loss_boundery_dice.append(boundery_dice_loss.item()) + + ## print training log message + if (it+1)%msg_iter==0: + loss_avg = sum(loss_avg) / len(loss_avg) + lr = optim.lr + ed = time.time() + t_intv, glob_t_intv = ed - st, ed - glob_st + eta = int((max_iter - it) * (glob_t_intv / it)) + eta = str(datetime.timedelta(seconds=eta)) + + loss_boundery_bce_avg = sum(loss_boundery_bce) / len(loss_boundery_bce) + loss_boundery_dice_avg = sum(loss_boundery_dice) / len(loss_boundery_dice) + msg = ', '.join([ + 'it: {it}/{max_it}', + 'lr: {lr:4f}', + 'loss: {loss:.4f}', + 'boundery_bce_loss: {boundery_bce_loss:.4f}', + 'boundery_dice_loss: {boundery_dice_loss:.4f}', + 'eta: {eta}', + 'time: {time:.4f}', + ]).format( + it = it+1, + max_it = max_iter, + lr = lr, + loss = loss_avg, + boundery_bce_loss = loss_boundery_bce_avg, + boundery_dice_loss = loss_boundery_dice_avg, + time = t_intv, + eta = eta + ) + + logger.info(msg) + loss_avg = [] + loss_boundery_bce = [] + loss_boundery_dice = [] + st = ed + # print(boundary_loss_func.get_params()) + if (it+1)%save_iter_sep==0:# and it != 0: + #if (it+1)%50==0:# and it != 0: + + ## model + logger.info('evaluating the model ...') + logger.info('setup and restore model') + + net.eval() + + # ## evaluator + logger.info('compute the mIOU') + with torch.no_grad(): + single_scale1 = MscEvalV0() + mIOU50 = single_scale1(net, dlval, n_classes) + + single_scale2 = MscEvalV0(scale=0.75) + mIOU75 = single_scale2(net, dlval, n_classes) + + save_pth = osp.join(save_pth_path, 'model_iter{}_mIOU50_{}_mIOU75_{}.pth' + .format(it+1, str(round(mIOU50, 4)), str(round(mIOU75, 4)))) + + state = net.module.state_dict() if hasattr(net, 'module') else net.state_dict() + # if dist.get_rank()==0: + torch.save(state, save_pth) + + logger.info('training iteration {}, model saved to: {}'.format(it+1, save_pth)) + + if mIOU50 > maxmIOU50: + maxmIOU50 = mIOU50 + save_pth = osp.join(save_pth_path, 'model_maxmIOU50.pth'.format(it+1)) + state = net.module.state_dict() if hasattr(net, 'module') else net.state_dict() + # if dist.get_rank()==0: + torch.save(state, save_pth) + + logger.info('max mIOU model saved to: {}'.format(save_pth)) + + if mIOU75 > maxmIOU75: + maxmIOU75 = mIOU75 + save_pth = osp.join(save_pth_path, 'model_maxmIOU75.pth'.format(it+1)) + state = net.module.state_dict() if hasattr(net, 'module') else net.state_dict() + # if dist.get_rank()==0: torch.save(state, save_pth) + torch.save(state, save_pth) + logger.info('max mIOU model saved to: {}'.format(save_pth)) + + logger.info('mIOU50 is: {}, mIOU75 is: {}'.format(mIOU50, mIOU75)) + logger.info('maxmIOU50 is: {}, maxmIOU75 is: {}.'.format(maxmIOU50, maxmIOU75)) + + net.train() + + ## dump the final model + save_pth = osp.join(save_pth_path, 'model_final.pth') + net.cpu() + state = net.module.state_dict() if hasattr(net, 'module') else net.state_dict() + # if dist.get_rank()==0: torch.save(state, save_pth) + torch.save(state, save_pth) + logger.info('training done, model saved to: {}'.format(save_pth)) + print('epoch: ', epoch) + + +if __name__ == "__main__": + train() diff --git a/transform.py b/transform.py new file mode 100644 index 0000000..3447a74 --- /dev/null +++ b/transform.py @@ -0,0 +1,118 @@ +#!/usr/bin/python +# -*- encoding: utf-8 -*- + + +from PIL import Image +import PIL.ImageEnhance as ImageEnhance +import random +import numpy as np + + +class RandomCrop(object): + def __init__(self, size, *args, **kwargs): + self.size = size + + def __call__(self, im_lb): + im = im_lb['im'] + lb = im_lb['lb'] + assert im.size == lb.size + W, H = self.size + w, h = im.size + + if (W, H) == (w, h): return dict(im=im, lb=lb) + if w < W or h < H: + scale = float(W) / w if w < h else float(H) / h + w, h = int(scale * w + 1), int(scale * h + 1) + im = im.resize((w, h), Image.BILINEAR) + lb = lb.resize((w, h), Image.NEAREST) + sw, sh = random.random() * (w - W), random.random() * (h - H) + crop = int(sw), int(sh), int(sw) + W, int(sh) + H + return dict( + im = im.crop(crop), + lb = lb.crop(crop) + ) + + +class HorizontalFlip(object): + def __init__(self, p=0.5, *args, **kwargs): + self.p = p + + def __call__(self, im_lb): + if random.random() > self.p: + return im_lb + else: + im = im_lb['im'] + lb = im_lb['lb'] + return dict(im = im.transpose(Image.FLIP_LEFT_RIGHT), + lb = lb.transpose(Image.FLIP_LEFT_RIGHT), + ) + + +class RandomScale(object): + def __init__(self, scales=(1, ), *args, **kwargs): + self.scales = scales + # print('scales: ', scales) + + def __call__(self, im_lb): + im = im_lb['im'] + lb = im_lb['lb'] + W, H = im.size + scale = random.choice(self.scales) + # scale = np.random.uniform(min(self.scales), max(self.scales)) + w, h = int(W * scale), int(H * scale) + return dict(im = im.resize((w, h), Image.BILINEAR), + lb = lb.resize((w, h), Image.NEAREST), + ) + + +class ColorJitter(object): + def __init__(self, brightness=None, contrast=None, saturation=None, *args, **kwargs): + if not brightness is None and brightness>0: + self.brightness = [max(1-brightness, 0), 1+brightness] + if not contrast is None and contrast>0: + self.contrast = [max(1-contrast, 0), 1+contrast] + if not saturation is None and saturation>0: + self.saturation = [max(1-saturation, 0), 1+saturation] + + def __call__(self, im_lb): + im = im_lb['im'] + lb = im_lb['lb'] + r_brightness = random.uniform(self.brightness[0], self.brightness[1]) + r_contrast = random.uniform(self.contrast[0], self.contrast[1]) + r_saturation = random.uniform(self.saturation[0], self.saturation[1]) + im = ImageEnhance.Brightness(im).enhance(r_brightness) + im = ImageEnhance.Contrast(im).enhance(r_contrast) + im = ImageEnhance.Color(im).enhance(r_saturation) + return dict(im = im, + lb = lb, + ) + + +class MultiScale(object): + def __init__(self, scales): + self.scales = scales + + def __call__(self, img): + W, H = img.size + sizes = [(int(W*ratio), int(H*ratio)) for ratio in self.scales] + imgs = [] + [imgs.append(img.resize(size, Image.BILINEAR)) for size in sizes] + return imgs + + +class Compose(object): + def __init__(self, do_list): + self.do_list = do_list + + def __call__(self, im_lb): + for comp in self.do_list: + im_lb = comp(im_lb) + return im_lb + + +if __name__ == '__main__': + flip = HorizontalFlip(p = 1) + crop = RandomCrop((321, 321)) + rscales = RandomScale((0.75, 1.0, 1.5, 1.75, 2.0)) + img = Image.open('data/img.jpg') + lb = Image.open('data/label.png') diff --git a/utils/__init__.py b/utils/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/utils/__pycache__/__init__.cpython-38.pyc b/utils/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000..0b0d094 Binary files /dev/null and b/utils/__pycache__/__init__.cpython-38.pyc differ diff --git a/utils/__pycache__/autoanchor.cpython-38.pyc b/utils/__pycache__/autoanchor.cpython-38.pyc new file mode 100644 index 0000000..465ee4b Binary files /dev/null and b/utils/__pycache__/autoanchor.cpython-38.pyc differ diff --git a/utils/__pycache__/common.cpython-38.pyc b/utils/__pycache__/common.cpython-38.pyc new file mode 100644 index 0000000..cd613a1 Binary files /dev/null and b/utils/__pycache__/common.cpython-38.pyc differ diff --git a/utils/__pycache__/datasets.cpython-38.pyc b/utils/__pycache__/datasets.cpython-38.pyc new file mode 100644 index 0000000..f066286 Binary files /dev/null and b/utils/__pycache__/datasets.cpython-38.pyc differ diff --git a/utils/__pycache__/experimental.cpython-38.pyc b/utils/__pycache__/experimental.cpython-38.pyc new file mode 100644 index 0000000..618ce55 Binary files /dev/null and b/utils/__pycache__/experimental.cpython-38.pyc differ diff --git a/utils/__pycache__/general.cpython-38.pyc b/utils/__pycache__/general.cpython-38.pyc new file mode 100644 index 0000000..53308e0 Binary files /dev/null and b/utils/__pycache__/general.cpython-38.pyc differ diff --git a/utils/__pycache__/google_utils.cpython-38.pyc b/utils/__pycache__/google_utils.cpython-38.pyc new file mode 100644 index 0000000..b61c113 Binary files /dev/null and b/utils/__pycache__/google_utils.cpython-38.pyc differ diff --git a/utils/__pycache__/loss.cpython-38.pyc b/utils/__pycache__/loss.cpython-38.pyc new file mode 100644 index 0000000..b736813 Binary files /dev/null and b/utils/__pycache__/loss.cpython-38.pyc differ diff --git a/utils/__pycache__/metrics.cpython-38.pyc b/utils/__pycache__/metrics.cpython-38.pyc new file mode 100644 index 0000000..5b43c7c Binary files /dev/null and b/utils/__pycache__/metrics.cpython-38.pyc differ diff --git a/utils/__pycache__/plots.cpython-38.pyc b/utils/__pycache__/plots.cpython-38.pyc new file mode 100644 index 0000000..b6d43c4 Binary files /dev/null and b/utils/__pycache__/plots.cpython-38.pyc differ diff --git a/utils/__pycache__/torch_utils.cpython-38.pyc b/utils/__pycache__/torch_utils.cpython-38.pyc new file mode 100644 index 0000000..f8b699d Binary files /dev/null and b/utils/__pycache__/torch_utils.cpython-38.pyc differ diff --git a/utils/activations.py b/utils/activations.py new file mode 100644 index 0000000..aa3ddf0 --- /dev/null +++ b/utils/activations.py @@ -0,0 +1,72 @@ +# Activation functions + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +# SiLU https://arxiv.org/pdf/1606.08415.pdf ---------------------------------------------------------------------------- +class SiLU(nn.Module): # export-friendly version of nn.SiLU() + @staticmethod + def forward(x): + return x * torch.sigmoid(x) + + +class Hardswish(nn.Module): # export-friendly version of nn.Hardswish() + @staticmethod + def forward(x): + # return x * F.hardsigmoid(x) # for torchscript and CoreML + return x * F.hardtanh(x + 3, 0., 6.) / 6. # for torchscript, CoreML and ONNX + + +class MemoryEfficientSwish(nn.Module): + class F(torch.autograd.Function): + @staticmethod + def forward(ctx, x): + ctx.save_for_backward(x) + return x * torch.sigmoid(x) + + @staticmethod + def backward(ctx, grad_output): + x = ctx.saved_tensors[0] + sx = torch.sigmoid(x) + return grad_output * (sx * (1 + x * (1 - sx))) + + def forward(self, x): + return self.F.apply(x) + + +# Mish https://github.com/digantamisra98/Mish -------------------------------------------------------------------------- +class Mish(nn.Module): + @staticmethod + def forward(x): + return x * F.softplus(x).tanh() + + +class MemoryEfficientMish(nn.Module): + class F(torch.autograd.Function): + @staticmethod + def forward(ctx, x): + ctx.save_for_backward(x) + return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x))) + + @staticmethod + def backward(ctx, grad_output): + x = ctx.saved_tensors[0] + sx = torch.sigmoid(x) + fx = F.softplus(x).tanh() + return grad_output * (fx + x * sx * (1 - fx * fx)) + + def forward(self, x): + return self.F.apply(x) + + +# FReLU https://arxiv.org/abs/2007.11824 ------------------------------------------------------------------------------- +class FReLU(nn.Module): + def __init__(self, c1, k=3): # ch_in, kernel + super().__init__() + self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False) + self.bn = nn.BatchNorm2d(c1) + + def forward(self, x): + return torch.max(x, self.bn(self.conv(x))) diff --git a/utils/autoanchor.py b/utils/autoanchor.py new file mode 100644 index 0000000..5777746 --- /dev/null +++ b/utils/autoanchor.py @@ -0,0 +1,160 @@ +# Auto-anchor utils + +import numpy as np +import torch +import yaml +from scipy.cluster.vq import kmeans +from tqdm import tqdm + +from utils.general import colorstr + + +def check_anchor_order(m): + # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary + a = m.anchor_grid.prod(-1).view(-1) # anchor area + da = a[-1] - a[0] # delta a + ds = m.stride[-1] - m.stride[0] # delta s + if da.sign() != ds.sign(): # same order + print('Reversing anchor order') + m.anchors[:] = m.anchors.flip(0) + m.anchor_grid[:] = m.anchor_grid.flip(0) + + +def check_anchors(dataset, model, thr=4.0, imgsz=640): + # Check anchor fit to data, recompute if necessary + prefix = colorstr('autoanchor: ') + print(f'\n{prefix}Analyzing anchors... ', end='') + m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect() + shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True) + scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale + wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() # wh + + def metric(k): # compute metric + r = wh[:, None] / k[None] + x = torch.min(r, 1. / r).min(2)[0] # ratio metric + best = x.max(1)[0] # best_x + aat = (x > 1. / thr).float().sum(1).mean() # anchors above threshold + bpr = (best > 1. / thr).float().mean() # best possible recall + return bpr, aat + + anchors = m.anchor_grid.clone().cpu().view(-1, 2) # current anchors + bpr, aat = metric(anchors) + print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='') + if bpr < 0.98: # threshold to recompute + print('. Attempting to improve anchors, please wait...') + na = m.anchor_grid.numel() // 2 # number of anchors + try: + anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False) + except Exception as e: + print(f'{prefix}ERROR: {e}') + new_bpr = metric(anchors)[0] + if new_bpr > bpr: # replace anchors + anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors) + m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid) # for inference + m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1) # loss + check_anchor_order(m) + print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.') + else: + print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.') + print('') # newline + + +def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True): + """ Creates kmeans-evolved anchors from training dataset + + Arguments: + path: path to dataset *.yaml, or a loaded dataset + n: number of anchors + img_size: image size used for training + thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0 + gen: generations to evolve anchors using genetic algorithm + verbose: print all results + + Return: + k: kmeans evolved anchors + + Usage: + from utils.autoanchor import *; _ = kmean_anchors() + """ + thr = 1. / thr + prefix = colorstr('autoanchor: ') + + def metric(k, wh): # compute metrics + r = wh[:, None] / k[None] + x = torch.min(r, 1. / r).min(2)[0] # ratio metric + # x = wh_iou(wh, torch.tensor(k)) # iou metric + return x, x.max(1)[0] # x, best_x + + def anchor_fitness(k): # mutation fitness + _, best = metric(torch.tensor(k, dtype=torch.float32), wh) + return (best * (best > thr).float()).mean() # fitness + + def print_results(k): + k = k[np.argsort(k.prod(1))] # sort small to large + x, best = metric(k, wh0) + bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr + print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr') + print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, ' + f'past_thr={x[x > thr].mean():.3f}-mean: ', end='') + for i, x in enumerate(k): + print('%i,%i' % (round(x[0]), round(x[1])), end=', ' if i < len(k) - 1 else '\n') # use in *.cfg + return k + + if isinstance(path, str): # *.yaml file + with open(path) as f: + data_dict = yaml.load(f, Loader=yaml.SafeLoader) # model dict + from utils.datasets import LoadImagesAndLabels + dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True) + else: + dataset = path # dataset + + # Get label wh + shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True) + wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh + + # Filter + i = (wh0 < 3.0).any(1).sum() + if i: + print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.') + wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels + # wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1) # multiply by random scale 0-1 + + # Kmeans calculation + print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...') + s = wh.std(0) # sigmas for whitening + k, dist = kmeans(wh / s, n, iter=30) # points, mean distance + assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}') + k *= s + wh = torch.tensor(wh, dtype=torch.float32) # filtered + wh0 = torch.tensor(wh0, dtype=torch.float32) # unfiltered + k = print_results(k) + + # Plot + # k, d = [None] * 20, [None] * 20 + # for i in tqdm(range(1, 21)): + # k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance + # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True) + # ax = ax.ravel() + # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.') + # fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh + # ax[0].hist(wh[wh[:, 0]<100, 0],400) + # ax[1].hist(wh[wh[:, 1]<100, 1],400) + # fig.savefig('wh.png', dpi=200) + + # Evolve + npr = np.random + f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma + pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:') # progress bar + for _ in pbar: + v = np.ones(sh) + while (v == 1).all(): # mutate until a change occurs (prevent duplicates) + v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0) + kg = (k.copy() * v).clip(min=2.0) + fg = anchor_fitness(kg) + if fg > f: + f, k = fg, kg.copy() + pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}' + if verbose: + print_results(k) + + return print_results(k) diff --git a/utils/aws/__init__.py b/utils/aws/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/utils/aws/mime.sh b/utils/aws/mime.sh new file mode 100644 index 0000000..c319a83 --- /dev/null +++ b/utils/aws/mime.sh @@ -0,0 +1,26 @@ +# AWS EC2 instance startup 'MIME' script https://aws.amazon.com/premiumsupport/knowledge-center/execute-user-data-ec2/ +# This script will run on every instance restart, not only on first start +# --- DO NOT COPY ABOVE COMMENTS WHEN PASTING INTO USERDATA --- + +Content-Type: multipart/mixed; boundary="//" +MIME-Version: 1.0 + +--// +Content-Type: text/cloud-config; charset="us-ascii" +MIME-Version: 1.0 +Content-Transfer-Encoding: 7bit +Content-Disposition: attachment; filename="cloud-config.txt" + +#cloud-config +cloud_final_modules: +- [scripts-user, always] + +--// +Content-Type: text/x-shellscript; charset="us-ascii" +MIME-Version: 1.0 +Content-Transfer-Encoding: 7bit +Content-Disposition: attachment; filename="userdata.txt" + +#!/bin/bash +# --- paste contents of userdata.sh here --- +--// diff --git a/utils/aws/resume.py b/utils/aws/resume.py new file mode 100644 index 0000000..faad8d2 --- /dev/null +++ b/utils/aws/resume.py @@ -0,0 +1,37 @@ +# Resume all interrupted trainings in yolov5/ dir including DDP trainings +# Usage: $ python utils/aws/resume.py + +import os +import sys +from pathlib import Path + +import torch +import yaml + +sys.path.append('./') # to run '$ python *.py' files in subdirectories + +port = 0 # --master_port +path = Path('').resolve() +for last in path.rglob('*/**/last.pt'): + ckpt = torch.load(last) + if ckpt['optimizer'] is None: + continue + + # Load opt.yaml + with open(last.parent.parent / 'opt.yaml') as f: + opt = yaml.load(f, Loader=yaml.SafeLoader) + + # Get device count + d = opt['device'].split(',') # devices + nd = len(d) # number of devices + ddp = nd > 1 or (nd == 0 and torch.cuda.device_count() > 1) # distributed data parallel + + if ddp: # multi-GPU + port += 1 + cmd = f'python -m torch.distributed.launch --nproc_per_node {nd} --master_port {port} train.py --resume {last}' + else: # single-GPU + cmd = f'python train.py --resume {last}' + + cmd += ' > /dev/null 2>&1 &' # redirect output to dev/null and run in daemon thread + print(cmd) + os.system(cmd) diff --git a/utils/aws/userdata.sh b/utils/aws/userdata.sh new file mode 100644 index 0000000..890606b --- /dev/null +++ b/utils/aws/userdata.sh @@ -0,0 +1,27 @@ +#!/bin/bash +# AWS EC2 instance startup script https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/user-data.html +# This script will run only once on first instance start (for a re-start script see mime.sh) +# /home/ubuntu (ubuntu) or /home/ec2-user (amazon-linux) is working dir +# Use >300 GB SSD + +cd home/ubuntu +if [ ! -d yolov5 ]; then + echo "Running first-time script." # install dependencies, download COCO, pull Docker + git clone https://github.com/ultralytics/yolov5 && sudo chmod -R 777 yolov5 + cd yolov5 + bash data/scripts/get_coco.sh && echo "Data done." & + sudo docker pull ultralytics/yolov5:latest && echo "Docker done." & + python -m pip install --upgrade pip && pip install -r requirements.txt && python detect.py && echo "Requirements done." & + wait && echo "All tasks done." # finish background tasks +else + echo "Running re-start script." # resume interrupted runs + i=0 + list=$(sudo docker ps -qa) # container list i.e. $'one\ntwo\nthree\nfour' + while IFS= read -r id; do + ((i++)) + echo "restarting container $i: $id" + sudo docker start $id + # sudo docker exec -it $id python train.py --resume # single-GPU + sudo docker exec -d $id python utils/aws/resume.py # multi-scenario + done <<<"$list" +fi diff --git a/utils/datasets.py b/utils/datasets.py new file mode 100644 index 0000000..6080237 --- /dev/null +++ b/utils/datasets.py @@ -0,0 +1,1072 @@ +# Dataset utils and dataloaders + +import glob +import logging +import math +import os +import random +import shutil +import time +from itertools import repeat +from multiprocessing.pool import ThreadPool +from pathlib import Path +from threading import Thread + +import cv2 +import numpy as np +import torch +import torch.nn.functional as F +from PIL import Image, ExifTags +from torch.utils.data import Dataset +from tqdm import tqdm + +from utils.general import check_requirements, xyxy2xywh, xywh2xyxy, xywhn2xyxy, xyn2xy, segment2box, segments2boxes, \ + resample_segments, clean_str +from utils.torch_utils import torch_distributed_zero_first + +# Parameters +help_url = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data' +img_formats = ['bmp', 'jpg', 'jpeg', 'png', 'tif', 'tiff', 'dng', 'webp', 'mpo'] # acceptable image suffixes +vid_formats = ['mov', 'avi', 'mp4', 'mpg', 'mpeg', 'm4v', 'wmv', 'mkv'] # acceptable video suffixes +logger = logging.getLogger(__name__) + +# Get orientation exif tag +for orientation in ExifTags.TAGS.keys(): + if ExifTags.TAGS[orientation] == 'Orientation': + break + + +def get_hash(files): + # Returns a single hash value of a list of files + return sum(os.path.getsize(f) for f in files if os.path.isfile(f)) + + +def exif_size(img): + # Returns exif-corrected PIL size + s = img.size # (width, height) + try: + rotation = dict(img._getexif().items())[orientation] + if rotation == 6: # rotation 270 + s = (s[1], s[0]) + elif rotation == 8: # rotation 90 + s = (s[1], s[0]) + except: + pass + + return s + + +def create_dataloader(path, imgsz, batch_size, stride, opt, hyp=None, augment=False, cache=False, pad=0.0, rect=False, + rank=-1, world_size=1, workers=8, image_weights=False, quad=False, prefix=''): + # Make sure only the first process in DDP process the dataset first, and the following others can use the cache + with torch_distributed_zero_first(rank): + dataset = LoadImagesAndLabels(path, imgsz, batch_size, + augment=augment, # augment images + hyp=hyp, # augmentation hyperparameters + rect=rect, # rectangular training + cache_images=cache, + single_cls=opt.single_cls, + stride=int(stride), + pad=pad, + image_weights=image_weights, + prefix=prefix) + + batch_size = min(batch_size, len(dataset)) + nw = min([os.cpu_count() // world_size, batch_size if batch_size > 1 else 0, workers]) # number of workers + sampler = torch.utils.data.distributed.DistributedSampler(dataset) if rank != -1 else None + loader = torch.utils.data.DataLoader if image_weights else InfiniteDataLoader + # Use torch.utils.data.DataLoader() if dataset.properties will update during training else InfiniteDataLoader() + dataloader = loader(dataset, + batch_size=batch_size, + num_workers=nw, + sampler=sampler, + pin_memory=True, + collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn) + return dataloader, dataset + + +class InfiniteDataLoader(torch.utils.data.dataloader.DataLoader): + """ Dataloader that reuses workers + + Uses same syntax as vanilla DataLoader + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + object.__setattr__(self, 'batch_sampler', _RepeatSampler(self.batch_sampler)) + self.iterator = super().__iter__() + + def __len__(self): + return len(self.batch_sampler.sampler) + + def __iter__(self): + for i in range(len(self)): + yield next(self.iterator) + + +class _RepeatSampler(object): + """ Sampler that repeats forever + + Args: + sampler (Sampler) + """ + + def __init__(self, sampler): + self.sampler = sampler + + def __iter__(self): + while True: + yield from iter(self.sampler) + + +class LoadImages: # for inference + def __init__(self, path, img_size=640, stride=32): + + p = str(Path(path).absolute()) # os-agnostic absolute path + # p = str(Path(path)) + + if '*' in p: + files = sorted(glob.glob(p, recursive=True)) # glob + elif os.path.isdir(p): + files = sorted(glob.glob(os.path.join(p, '*.*'))) # dir + elif os.path.isfile(p): + files = [p] # files + else: + raise Exception(f'ERROR: {p} does not exist') + + images = [x for x in files if x.split('.')[-1].lower() in img_formats] + videos = [x for x in files if x.split('.')[-1].lower() in vid_formats] + ni, nv = len(images), len(videos) + + self.img_size = img_size + self.stride = stride + self.files = images + videos + self.nf = ni + nv # number of files + self.video_flag = [False] * ni + [True] * nv + self.mode = 'image' + if any(videos): + self.new_video(videos[0]) # new video + else: + self.cap = None + assert self.nf > 0, f'No images or videos found in {p}. ' \ + f'Supported formats are:\nimages: {img_formats}\nvideos: {vid_formats}' + + def __iter__(self): + self.count = 0 + return self + + def __next__(self): + if self.count == self.nf: + raise StopIteration + path = self.files[self.count] + + if self.video_flag[self.count]: + # Read video + self.mode = 'video' + ret_val, img0 = self.cap.read() + if not ret_val: + self.count += 1 + self.cap.release() + if self.count == self.nf: # last video + raise StopIteration + else: + path = self.files[self.count] + self.new_video(path) + ret_val, img0 = self.cap.read() + + self.frame += 1 + print(f'video {self.count + 1}/{self.nf} ({self.frame}/{self.nframes}) {path}: ', end='') + + else: + # Read image + self.count += 1 + img0 = cv2.imread(path) # BGR + + # print("line185", path) + # print("line186", img0.shape) + + assert img0 is not None, 'Image Not Found ' + path + print(f'image {self.count}/{self.nf} {path}: ', end='') + + # Padded resize + img = letterbox(img0, self.img_size, stride=self.stride)[0] + + # Convert + img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 + img = np.ascontiguousarray(img) + # print("line197", img0.shape) + + return path, img, img0, self.cap + + def new_video(self, path): + self.frame = 0 + self.cap = cv2.VideoCapture(path) + self.nframes = int(self.cap.get(cv2.CAP_PROP_FRAME_COUNT)) + + def __len__(self): + return self.nf # number of files + + +class LoadWebcam: # for inference + def __init__(self, pipe='0', img_size=640, stride=32): + self.img_size = img_size + self.stride = stride + + if pipe.isnumeric(): + pipe = eval(pipe) # local camera + # pipe = 'rtsp://192.168.1.64/1' # IP camera + # pipe = 'rtsp://username:password@192.168.1.64/1' # IP camera with login + # pipe = 'http://wmccpinetop.axiscam.net/mjpg/video.mjpg' # IP golf camera + + self.pipe = pipe + self.cap = cv2.VideoCapture(pipe) # video capture object + self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 3) # set buffer size + + def __iter__(self): + self.count = -1 + return self + + def __next__(self): + self.count += 1 + if cv2.waitKey(1) == ord('q'): # q to quit + self.cap.release() + cv2.destroyAllWindows() + raise StopIteration + + # Read frame + if self.pipe == 0: # local camera + ret_val, img0 = self.cap.read() + img0 = cv2.flip(img0, 1) # flip left-right + else: # IP camera + n = 0 + while True: + n += 1 + self.cap.grab() + if n % 30 == 0: # skip frames + ret_val, img0 = self.cap.retrieve() + if ret_val: + break + + # Print + assert ret_val, f'Camera Error {self.pipe}' + img_path = 'webcam.jpg' + print(f'webcam {self.count}: ', end='') + + # Padded resize + img = letterbox(img0, self.img_size, stride=self.stride)[0] + + # Convert + img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 + img = np.ascontiguousarray(img) + + return img_path, img, img0, None + + def __len__(self): + return 0 + + +class LoadStreams: # multiple IP or RTSP cameras + def __init__(self, sources='streams.txt', img_size=640, stride=32): + self.mode = 'stream' + self.img_size = img_size + self.stride = stride + + if os.path.isfile(sources): + with open(sources, 'r') as f: + sources = [x.strip() for x in f.read().strip().splitlines() if len(x.strip())] + else: + sources = [sources] + + n = len(sources) + self.imgs = [None] * n + self.sources = [clean_str(x) for x in sources] # clean source names for later + for i, s in enumerate(sources): + # Start the thread to read frames from the video stream + print(f'{i + 1}/{n}: {s}... ', end='') + url = eval(s) if s.isnumeric() else s + if 'youtube.com/' in url or 'youtu.be/' in url: # if source is YouTube video + check_requirements(('pafy', 'youtube_dl')) + import pafy + url = pafy.new(url).getbest(preftype="mp4").url + cap = cv2.VideoCapture(url) + assert cap.isOpened(), f'Failed to open {s}' + w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) + h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) + self.fps = cap.get(cv2.CAP_PROP_FPS) % 100 + + _, self.imgs[i] = cap.read() # guarantee first frame + thread = Thread(target=self.update, args=([i, cap]), daemon=True) + print(f' success ({w}x{h} at {self.fps:.2f} FPS).') + thread.start() + print('') # newline + + # check for common shapes + s = np.stack([letterbox(x, self.img_size, stride=self.stride)[0].shape for x in self.imgs], 0) # shapes + self.rect = np.unique(s, axis=0).shape[0] == 1 # rect inference if all shapes equal + if not self.rect: + print('WARNING: Different stream shapes detected. For optimal performance supply similarly-shaped streams.') + + def update(self, index, cap): + # Read next stream frame in a daemon thread + n = 0 + while cap.isOpened(): + n += 1 + # _, self.imgs[index] = cap.read() + cap.grab() + if n == 4: # read every 4th frame + success, im = cap.retrieve() + self.imgs[index] = im if success else self.imgs[index] * 0 + n = 0 + time.sleep(1 / self.fps) # wait time + + def __iter__(self): + self.count = -1 + return self + + def __next__(self): + self.count += 1 + img0 = self.imgs.copy() + if cv2.waitKey(1) == ord('q'): # q to quit + cv2.destroyAllWindows() + raise StopIteration + + # Letterbox + img = [letterbox(x, self.img_size, auto=self.rect, stride=self.stride)[0] for x in img0] + + # Stack + img = np.stack(img, 0) + + # Convert + img = img[:, :, :, ::-1].transpose(0, 3, 1, 2) # BGR to RGB, to bsx3x416x416 + img = np.ascontiguousarray(img) + + return self.sources, img, img0, None + + def __len__(self): + return 0 # 1E12 frames = 32 streams at 30 FPS for 30 years + + +def img2label_paths(img_paths): + # Define label paths as a function of image paths + sa, sb = os.sep + 'images' + os.sep, os.sep + 'labels' + os.sep # /images/, /labels/ substrings + return ['txt'.join(x.replace(sa, sb, 1).rsplit(x.split('.')[-1], 1)) for x in img_paths] + + +class LoadImagesAndLabels(Dataset): # for training/testing + def __init__(self, path, img_size=640, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False, + cache_images=False, single_cls=False, stride=32, pad=0.0, prefix=''): + self.img_size = img_size + self.augment = augment + self.hyp = hyp + self.image_weights = image_weights + self.rect = False if image_weights else rect + self.mosaic = self.augment and not self.rect # load 4 images at a time into a mosaic (only during training) + self.mosaic_border = [-img_size // 2, -img_size // 2] + self.stride = stride + self.path = path + + try: + f = [] # image files + for p in path if isinstance(path, list) else [path]: + p = Path(p) # os-agnostic + if p.is_dir(): # dir + f += glob.glob(str(p / '**' / '*.*'), recursive=True) + # f = list(p.rglob('**/*.*')) # pathlib + elif p.is_file(): # file + with open(p, 'r') as t: + t = t.read().strip().splitlines() + parent = str(p.parent) + os.sep + f += [x.replace('./', parent) if x.startswith('./') else x for x in t] # local to global path + # f += [p.parent / x.lstrip(os.sep) for x in t] # local to global path (pathlib) + else: + raise Exception(f'{prefix}{p} does not exist') + self.img_files = sorted([x.replace('/', os.sep) for x in f if x.split('.')[-1].lower() in img_formats]) + # self.img_files = sorted([x for x in f if x.suffix[1:].lower() in img_formats]) # pathlib + assert self.img_files, f'{prefix}No images found' + except Exception as e: + raise Exception(f'{prefix}Error loading data from {path}: {e}\nSee {help_url}') + + # Check cache + self.label_files = img2label_paths(self.img_files) # labels + cache_path = (p if p.is_file() else Path(self.label_files[0]).parent).with_suffix('.cache') # cached labels + if cache_path.is_file(): + cache, exists = torch.load(cache_path), True # load + if cache['hash'] != get_hash(self.label_files + self.img_files) or 'version' not in cache: # changed + cache, exists = self.cache_labels(cache_path, prefix), False # re-cache + else: + cache, exists = self.cache_labels(cache_path, prefix), False # cache + + # Display cache + nf, nm, ne, nc, n = cache.pop('results') # found, missing, empty, corrupted, total + if exists: + d = f"Scanning '{cache_path}' images and labels... {nf} found, {nm} missing, {ne} empty, {nc} corrupted" + tqdm(None, desc=prefix + d, total=n, initial=n) # display cache results + assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {help_url}' + + # Read cache + cache.pop('hash') # remove hash + cache.pop('version') # remove version + labels, shapes, self.segments = zip(*cache.values()) + self.labels = list(labels) + self.shapes = np.array(shapes, dtype=np.float64) + self.img_files = list(cache.keys()) # update + self.label_files = img2label_paths(cache.keys()) # update + if single_cls: + for x in self.labels: + x[:, 0] = 0 + + n = len(shapes) # number of images + bi = np.floor(np.arange(n) / batch_size).astype(np.int) # batch index + nb = bi[-1] + 1 # number of batches + self.batch = bi # batch index of image + self.n = n + self.indices = range(n) + + # Rectangular Training + if self.rect: + # Sort by aspect ratio + s = self.shapes # wh + ar = s[:, 1] / s[:, 0] # aspect ratio + irect = ar.argsort() + self.img_files = [self.img_files[i] for i in irect] + self.label_files = [self.label_files[i] for i in irect] + self.labels = [self.labels[i] for i in irect] + self.shapes = s[irect] # wh + ar = ar[irect] + + # Set training image shapes + shapes = [[1, 1]] * nb + for i in range(nb): + ari = ar[bi == i] + mini, maxi = ari.min(), ari.max() + if maxi < 1: + shapes[i] = [maxi, 1] + elif mini > 1: + shapes[i] = [1, 1 / mini] + + self.batch_shapes = np.ceil(np.array(shapes) * img_size / stride + pad).astype(np.int) * stride + + # Cache images into memory for faster training (WARNING: large datasets may exceed system RAM) + self.imgs = [None] * n + if cache_images: + gb = 0 # Gigabytes of cached images + self.img_hw0, self.img_hw = [None] * n, [None] * n + results = ThreadPool(8).imap(lambda x: load_image(*x), zip(repeat(self), range(n))) # 8 threads + pbar = tqdm(enumerate(results), total=n) + for i, x in pbar: + self.imgs[i], self.img_hw0[i], self.img_hw[i] = x # img, hw_original, hw_resized = load_image(self, i) + gb += self.imgs[i].nbytes + pbar.desc = f'{prefix}Caching images ({gb / 1E9:.1f}GB)' + pbar.close() + + def cache_labels(self, path=Path('./labels.cache'), prefix=''): + # Cache dataset labels, check images and read shapes + x = {} # dict + nm, nf, ne, nc = 0, 0, 0, 0 # number missing, found, empty, duplicate + pbar = tqdm(zip(self.img_files, self.label_files), desc='Scanning images', total=len(self.img_files)) + for i, (im_file, lb_file) in enumerate(pbar): + try: + # verify images + im = Image.open(im_file) + im.verify() # PIL verify + shape = exif_size(im) # image size + segments = [] # instance segments + assert (shape[0] > 9) & (shape[1] > 9), f'image size {shape} <10 pixels' + assert im.format.lower() in img_formats, f'invalid image format {im.format}' + + # verify labels + if os.path.isfile(lb_file): + nf += 1 # label found + with open(lb_file, 'r') as f: + l = [x.split() for x in f.read().strip().splitlines()] + if any([len(x) > 8 for x in l]): # is segment + classes = np.array([x[0] for x in l], dtype=np.float32) + segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in l] # (cls, xy1...) + l = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1) # (cls, xywh) + l = np.array(l, dtype=np.float32) + if len(l): + assert l.shape[1] == 5, 'labels require 5 columns each' + assert (l >= 0).all(), 'negative labels' + assert (l[:, 1:] <= 1).all(), 'non-normalized or out of bounds coordinate labels' + assert np.unique(l, axis=0).shape[0] == l.shape[0], 'duplicate labels' + else: + ne += 1 # label empty + l = np.zeros((0, 5), dtype=np.float32) + else: + nm += 1 # label missing + l = np.zeros((0, 5), dtype=np.float32) + x[im_file] = [l, shape, segments] + except Exception as e: + nc += 1 + print(f'{prefix}WARNING: Ignoring corrupted image and/or label {im_file}: {e}') + + pbar.desc = f"{prefix}Scanning '{path.parent / path.stem}' images and labels... " \ + f"{nf} found, {nm} missing, {ne} empty, {nc} corrupted" + pbar.close() + + if nf == 0: + print(f'{prefix}WARNING: No labels found in {path}. See {help_url}') + + x['hash'] = get_hash(self.label_files + self.img_files) + x['results'] = nf, nm, ne, nc, i + 1 + x['version'] = 0.1 # cache version + torch.save(x, path) # save for next time + logging.info(f'{prefix}New cache created: {path}') + return x + + def __len__(self): + return len(self.img_files) + + # def __iter__(self): + # self.count = -1 + # print('ran dataset iter') + # #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF) + # return self + + def __getitem__(self, index): + index = self.indices[index] # linear, shuffled, or image_weights + + hyp = self.hyp + mosaic = self.mosaic and random.random() < hyp['mosaic'] + if mosaic: + # Load mosaic + img, labels = load_mosaic(self, index) + shapes = None + + # MixUp https://arxiv.org/pdf/1710.09412.pdf + if random.random() < hyp['mixup']: + img2, labels2 = load_mosaic(self, random.randint(0, self.n - 1)) + r = np.random.beta(8.0, 8.0) # mixup ratio, alpha=beta=8.0 + img = (img * r + img2 * (1 - r)).astype(np.uint8) + labels = np.concatenate((labels, labels2), 0) + + else: + # Load image + img, (h0, w0), (h, w) = load_image(self, index) + + # Letterbox + shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size # final letterboxed shape + img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment) + shapes = (h0, w0), ((h / h0, w / w0), pad) # for COCO mAP rescaling + + labels = self.labels[index].copy() + if labels.size: # normalized xywh to pixel xyxy format + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1]) + + if self.augment: + # Augment imagespace + if not mosaic: + img, labels = random_perspective(img, labels, + degrees=hyp['degrees'], + translate=hyp['translate'], + scale=hyp['scale'], + shear=hyp['shear'], + perspective=hyp['perspective']) + + # Augment colorspace + augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v']) + + # Apply cutouts + # if random.random() < 0.9: + # labels = cutout(img, labels) + + nL = len(labels) # number of labels + if nL: + labels[:, 1:5] = xyxy2xywh(labels[:, 1:5]) # convert xyxy to xywh + labels[:, [2, 4]] /= img.shape[0] # normalized height 0-1 + labels[:, [1, 3]] /= img.shape[1] # normalized width 0-1 + + if self.augment: + # flip up-down + if random.random() < hyp['flipud']: + img = np.flipud(img) + if nL: + labels[:, 2] = 1 - labels[:, 2] + + # flip left-right + if random.random() < hyp['fliplr']: + img = np.fliplr(img) + if nL: + labels[:, 1] = 1 - labels[:, 1] + + labels_out = torch.zeros((nL, 6)) + if nL: + labels_out[:, 1:] = torch.from_numpy(labels) + + # Convert + img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 + img = np.ascontiguousarray(img) + + return torch.from_numpy(img), labels_out, self.img_files[index], shapes + + @staticmethod + def collate_fn(batch): + img, label, path, shapes = zip(*batch) # transposed + for i, l in enumerate(label): + l[:, 0] = i # add target image index for build_targets() + return torch.stack(img, 0), torch.cat(label, 0), path, shapes + + @staticmethod + def collate_fn4(batch): + img, label, path, shapes = zip(*batch) # transposed + n = len(shapes) // 4 + img4, label4, path4, shapes4 = [], [], path[:n], shapes[:n] + + ho = torch.tensor([[0., 0, 0, 1, 0, 0]]) + wo = torch.tensor([[0., 0, 1, 0, 0, 0]]) + s = torch.tensor([[1, 1, .5, .5, .5, .5]]) # scale + for i in range(n): # zidane torch.zeros(16,3,720,1280) # BCHW + i *= 4 + if random.random() < 0.5: + im = F.interpolate(img[i].unsqueeze(0).float(), scale_factor=2., mode='bilinear', align_corners=False)[ + 0].type(img[i].type()) + l = label[i] + else: + im = torch.cat((torch.cat((img[i], img[i + 1]), 1), torch.cat((img[i + 2], img[i + 3]), 1)), 2) + l = torch.cat((label[i], label[i + 1] + ho, label[i + 2] + wo, label[i + 3] + ho + wo), 0) * s + img4.append(im) + label4.append(l) + + for i, l in enumerate(label4): + l[:, 0] = i # add target image index for build_targets() + + return torch.stack(img4, 0), torch.cat(label4, 0), path4, shapes4 + + +# Ancillary functions -------------------------------------------------------------------------------------------------- +def load_image(self, index): + # loads 1 image from dataset, returns img, original hw, resized hw + img = self.imgs[index] + if img is None: # not cached + path = self.img_files[index] + img = cv2.imread(path) # BGR + assert img is not None, 'Image Not Found ' + path + h0, w0 = img.shape[:2] # orig hw + r = self.img_size / max(h0, w0) # resize image to img_size + if r != 1: # always resize down, only resize up if training with augmentation + interp = cv2.INTER_AREA if r < 1 and not self.augment else cv2.INTER_LINEAR + img = cv2.resize(img, (int(w0 * r), int(h0 * r)), interpolation=interp) + return img, (h0, w0), img.shape[:2] # img, hw_original, hw_resized + else: + return self.imgs[index], self.img_hw0[index], self.img_hw[index] # img, hw_original, hw_resized + + +def augment_hsv(img, hgain=0.5, sgain=0.5, vgain=0.5): + r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1 # random gains + hue, sat, val = cv2.split(cv2.cvtColor(img, cv2.COLOR_BGR2HSV)) + dtype = img.dtype # uint8 + + x = np.arange(0, 256, dtype=np.int16) + lut_hue = ((x * r[0]) % 180).astype(dtype) + lut_sat = np.clip(x * r[1], 0, 255).astype(dtype) + lut_val = np.clip(x * r[2], 0, 255).astype(dtype) + + img_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val))).astype(dtype) + cv2.cvtColor(img_hsv, cv2.COLOR_HSV2BGR, dst=img) # no return needed + + +def hist_equalize(img, clahe=True, bgr=False): + # Equalize histogram on BGR image 'img' with img.shape(n,m,3) and range 0-255 + yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV) + if clahe: + c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8)) + yuv[:, :, 0] = c.apply(yuv[:, :, 0]) + else: + yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0]) # equalize Y channel histogram + return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB) # convert YUV image to RGB + + +def load_mosaic(self, index): + # loads images in a 4-mosaic + + labels4, segments4 = [], [] + s = self.img_size + yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border] # mosaic center x, y + indices = [index] + random.choices(self.indices, k=3) # 3 additional image indices + for i, index in enumerate(indices): + # Load image + img, _, (h, w) = load_image(self, index) + + # place img in img4 + if i == 0: # top left + img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles + x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image) + x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image) + elif i == 1: # top right + x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc + x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h + elif i == 2: # bottom left + x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h) + x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h) + elif i == 3: # bottom right + x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h) + x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h) + + img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax] + padw = x1a - x1b + padh = y1a - y1b + + # Labels + labels, segments = self.labels[index].copy(), self.segments[index].copy() + if labels.size: + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh) # normalized xywh to pixel xyxy format + segments = [xyn2xy(x, w, h, padw, padh) for x in segments] + labels4.append(labels) + segments4.extend(segments) + + # Concat/clip labels + labels4 = np.concatenate(labels4, 0) + for x in (labels4[:, 1:], *segments4): + np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective() + # img4, labels4 = replicate(img4, labels4) # replicate + + # Augment + img4, labels4 = random_perspective(img4, labels4, segments4, + degrees=self.hyp['degrees'], + translate=self.hyp['translate'], + scale=self.hyp['scale'], + shear=self.hyp['shear'], + perspective=self.hyp['perspective'], + border=self.mosaic_border) # border to remove + + return img4, labels4 + + +def load_mosaic9(self, index): + # loads images in a 9-mosaic + + labels9, segments9 = [], [] + s = self.img_size + indices = [index] + random.choices(self.indices, k=8) # 8 additional image indices + for i, index in enumerate(indices): + # Load image + img, _, (h, w) = load_image(self, index) + + # place img in img9 + if i == 0: # center + img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles + h0, w0 = h, w + c = s, s, s + w, s + h # xmin, ymin, xmax, ymax (base) coordinates + elif i == 1: # top + c = s, s - h, s + w, s + elif i == 2: # top right + c = s + wp, s - h, s + wp + w, s + elif i == 3: # right + c = s + w0, s, s + w0 + w, s + h + elif i == 4: # bottom right + c = s + w0, s + hp, s + w0 + w, s + hp + h + elif i == 5: # bottom + c = s + w0 - w, s + h0, s + w0, s + h0 + h + elif i == 6: # bottom left + c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h + elif i == 7: # left + c = s - w, s + h0 - h, s, s + h0 + elif i == 8: # top left + c = s - w, s + h0 - hp - h, s, s + h0 - hp + + padx, pady = c[:2] + x1, y1, x2, y2 = [max(x, 0) for x in c] # allocate coords + + # Labels + labels, segments = self.labels[index].copy(), self.segments[index].copy() + if labels.size: + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padx, pady) # normalized xywh to pixel xyxy format + segments = [xyn2xy(x, w, h, padx, pady) for x in segments] + labels9.append(labels) + segments9.extend(segments) + + # Image + img9[y1:y2, x1:x2] = img[y1 - pady:, x1 - padx:] # img9[ymin:ymax, xmin:xmax] + hp, wp = h, w # height, width previous + + # Offset + yc, xc = [int(random.uniform(0, s)) for _ in self.mosaic_border] # mosaic center x, y + img9 = img9[yc:yc + 2 * s, xc:xc + 2 * s] + + # Concat/clip labels + labels9 = np.concatenate(labels9, 0) + labels9[:, [1, 3]] -= xc + labels9[:, [2, 4]] -= yc + c = np.array([xc, yc]) # centers + segments9 = [x - c for x in segments9] + + for x in (labels9[:, 1:], *segments9): + np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective() + # img9, labels9 = replicate(img9, labels9) # replicate + + # Augment + img9, labels9 = random_perspective(img9, labels9, segments9, + degrees=self.hyp['degrees'], + translate=self.hyp['translate'], + scale=self.hyp['scale'], + shear=self.hyp['shear'], + perspective=self.hyp['perspective'], + border=self.mosaic_border) # border to remove + + return img9, labels9 + + +def replicate(img, labels): + # Replicate labels + h, w = img.shape[:2] + boxes = labels[:, 1:].astype(int) + x1, y1, x2, y2 = boxes.T + s = ((x2 - x1) + (y2 - y1)) / 2 # side length (pixels) + for i in s.argsort()[:round(s.size * 0.5)]: # smallest indices + x1b, y1b, x2b, y2b = boxes[i] + bh, bw = y2b - y1b, x2b - x1b + yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw)) # offset x, y + x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh] + img[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax] + labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0) + + return img, labels + + +def letterbox(img, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32): + # Resize and pad image while meeting stride-multiple constraints + shape = img.shape[:2] # current shape [height, width] + if isinstance(new_shape, int): + new_shape = (new_shape, new_shape) + + # Scale ratio (new / old) + r = min(new_shape[0] / shape[0], new_shape[1] / shape[1]) + if not scaleup: # only scale down, do not scale up (for better test mAP) + r = min(r, 1.0) + + # Compute padding + ratio = r, r # width, height ratios + new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) + dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh padding + if auto: # minimum rectangle + dw, dh = np.mod(dw, stride), np.mod(dh, stride) # wh padding + elif scaleFill: # stretch + dw, dh = 0.0, 0.0 + new_unpad = (new_shape[1], new_shape[0]) + ratio = new_shape[1] / shape[1], new_shape[0] / shape[0] # width, height ratios + + dw /= 2 # divide padding into 2 sides + dh /= 2 + + if shape[::-1] != new_unpad: # resize + img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR) + top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1)) + left, right = int(round(dw - 0.1)), int(round(dw + 0.1)) + img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add border + return img, ratio, (dw, dh) + + +def random_perspective(img, targets=(), segments=(), degrees=10, translate=.1, scale=.1, shear=10, perspective=0.0, + border=(0, 0)): + # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10)) + # targets = [cls, xyxy] + + height = img.shape[0] + border[0] * 2 # shape(h,w,c) + width = img.shape[1] + border[1] * 2 + + # Center + C = np.eye(3) + C[0, 2] = -img.shape[1] / 2 # x translation (pixels) + C[1, 2] = -img.shape[0] / 2 # y translation (pixels) + + # Perspective + P = np.eye(3) + P[2, 0] = random.uniform(-perspective, perspective) # x perspective (about y) + P[2, 1] = random.uniform(-perspective, perspective) # y perspective (about x) + + # Rotation and Scale + R = np.eye(3) + a = random.uniform(-degrees, degrees) + # a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations + s = random.uniform(1 - scale, 1 + scale) + # s = 2 ** random.uniform(-scale, scale) + R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s) + + # Shear + S = np.eye(3) + S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg) + S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg) + + # Translation + T = np.eye(3) + T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width # x translation (pixels) + T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height # y translation (pixels) + + # Combined rotation matrix + M = T @ S @ R @ P @ C # order of operations (right to left) is IMPORTANT + if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any(): # image changed + if perspective: + img = cv2.warpPerspective(img, M, dsize=(width, height), borderValue=(114, 114, 114)) + else: # affine + img = cv2.warpAffine(img, M[:2], dsize=(width, height), borderValue=(114, 114, 114)) + + # Visualize + # import matplotlib.pyplot as plt + # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel() + # ax[0].imshow(img[:, :, ::-1]) # base + # ax[1].imshow(img2[:, :, ::-1]) # warped + + # Transform label coordinates + n = len(targets) + if n: + use_segments = any(x.any() for x in segments) + new = np.zeros((n, 4)) + if use_segments: # warp segments + segments = resample_segments(segments) # upsample + for i, segment in enumerate(segments): + xy = np.ones((len(segment), 3)) + xy[:, :2] = segment + xy = xy @ M.T # transform + xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2] # perspective rescale or affine + + # clip + new[i] = segment2box(xy, width, height) + + else: # warp boxes + xy = np.ones((n * 4, 3)) + xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2) # x1y1, x2y2, x1y2, x2y1 + xy = xy @ M.T # transform + xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8) # perspective rescale or affine + + # create new boxes + x = xy[:, [0, 2, 4, 6]] + y = xy[:, [1, 3, 5, 7]] + new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T + + # clip + new[:, [0, 2]] = new[:, [0, 2]].clip(0, width) + new[:, [1, 3]] = new[:, [1, 3]].clip(0, height) + + # filter candidates + i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10) + targets = targets[i] + targets[:, 1:5] = new[i] + + return img, targets + + +def box_candidates(box1, box2, wh_thr=2, ar_thr=20, area_thr=0.1, eps=1e-16): # box1(4,n), box2(4,n) + # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio + w1, h1 = box1[2] - box1[0], box1[3] - box1[1] + w2, h2 = box2[2] - box2[0], box2[3] - box2[1] + ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps)) # aspect ratio + return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr) # candidates + + +def cutout(image, labels): + # Applies image cutout augmentation https://arxiv.org/abs/1708.04552 + h, w = image.shape[:2] + + def bbox_ioa(box1, box2): + # Returns the intersection over box2 area given box1, box2. box1 is 4, box2 is nx4. boxes are x1y1x2y2 + box2 = box2.transpose() + + # Get the coordinates of bounding boxes + b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3] + b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3] + + # Intersection area + inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \ + (np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0) + + # box2 area + box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + 1e-16 + + # Intersection over box2 area + return inter_area / box2_area + + # create random masks + scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16 # image size fraction + for s in scales: + mask_h = random.randint(1, int(h * s)) + mask_w = random.randint(1, int(w * s)) + + # box + xmin = max(0, random.randint(0, w) - mask_w // 2) + ymin = max(0, random.randint(0, h) - mask_h // 2) + xmax = min(w, xmin + mask_w) + ymax = min(h, ymin + mask_h) + + # apply random color mask + image[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)] + + # return unobscured labels + if len(labels) and s > 0.03: + box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32) + ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area + labels = labels[ioa < 0.60] # remove >60% obscured labels + + return labels + + +def create_folder(path='./new'): + # Create folder + if os.path.exists(path): + shutil.rmtree(path) # delete output folder + os.makedirs(path) # make new output folder + + +def flatten_recursive(path='../coco128'): + # Flatten a recursive directory by bringing all files to top level + new_path = Path(path + '_flat') + create_folder(new_path) + for file in tqdm(glob.glob(str(Path(path)) + '/**/*.*', recursive=True)): + shutil.copyfile(file, new_path / Path(file).name) + + +def extract_boxes(path='../coco128/'): # from utils.datasets import *; extract_boxes('../coco128') + # Convert detection dataset into classification dataset, with one directory per class + + path = Path(path) # images dir + shutil.rmtree(path / 'classifier') if (path / 'classifier').is_dir() else None # remove existing + files = list(path.rglob('*.*')) + n = len(files) # number of files + for im_file in tqdm(files, total=n): + if im_file.suffix[1:] in img_formats: + # image + im = cv2.imread(str(im_file))[..., ::-1] # BGR to RGB + h, w = im.shape[:2] + + # labels + lb_file = Path(img2label_paths([str(im_file)])[0]) + if Path(lb_file).exists(): + with open(lb_file, 'r') as f: + lb = np.array([x.split() for x in f.read().strip().splitlines()], dtype=np.float32) # labels + + for j, x in enumerate(lb): + c = int(x[0]) # class + f = (path / 'classifier') / f'{c}' / f'{path.stem}_{im_file.stem}_{j}.jpg' # new filename + if not f.parent.is_dir(): + f.parent.mkdir(parents=True) + + b = x[1:] * [w, h, w, h] # box + # b[2:] = b[2:].max() # rectangle to square + b[2:] = b[2:] * 1.2 + 3 # pad + b = xywh2xyxy(b.reshape(-1, 4)).ravel().astype(np.int) + + b[[0, 2]] = np.clip(b[[0, 2]], 0, w) # clip boxes outside of image + b[[1, 3]] = np.clip(b[[1, 3]], 0, h) + assert cv2.imwrite(str(f), im[b[1]:b[3], b[0]:b[2]]), f'box failure in {f}' + + +def autosplit(path='../coco128', weights=(0.9, 0.1, 0.0), annotated_only=False): + """ Autosplit a dataset into train/val/test splits and save path/autosplit_*.txt files + Usage: from utils.datasets import *; autosplit('../coco128') + Arguments + path: Path to images directory + weights: Train, val, test weights (list) + annotated_only: Only use images with an annotated txt file + """ + path = Path(path) # images dir + files = sum([list(path.rglob(f"*.{img_ext}")) for img_ext in img_formats], []) # image files only + n = len(files) # number of files + indices = random.choices([0, 1, 2], weights=weights, k=n) # assign each image to a split + + txt = ['autosplit_train.txt', 'autosplit_val.txt', 'autosplit_test.txt'] # 3 txt files + [(path / x).unlink() for x in txt if (path / x).exists()] # remove existing + + print(f'Autosplitting images from {path}' + ', using *.txt labeled images only' * annotated_only) + for i, img in tqdm(zip(indices, files), total=n): + if not annotated_only or Path(img2label_paths([str(img)])[0]).exists(): # check label + with open(path / txt[i], 'a') as f: + f.write(str(img) + '\n') # add image to txt file diff --git a/utils/general.py b/utils/general.py new file mode 100644 index 0000000..7e21f62 --- /dev/null +++ b/utils/general.py @@ -0,0 +1,608 @@ +# YOLOv5 general utils + +import glob +import logging +import math +import os +import platform +import random +import re +import subprocess +import time +from pathlib import Path + +import cv2 +import numpy as np +import pandas as pd +import torch +import torchvision +import yaml + +from utils.google_utils import gsutil_getsize + +from utils.torch_utils import init_torch_seeds + +# Settings +torch.set_printoptions(linewidth=320, precision=5, profile='long') +np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format}) # format short g, %precision=5 +pd.options.display.max_columns = 10 +cv2.setNumThreads(0) # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader) +os.environ['NUMEXPR_MAX_THREADS'] = str(min(os.cpu_count(), 8)) # NumExpr max threads + +def fitness(x): + # Model fitness as a weighted combination of metrics + w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] + return (x[:, :4] * w).sum(1) + +def set_logging(rank=-1): + logging.basicConfig( + format="%(message)s", + level=logging.INFO if rank in [-1, 0] else logging.WARN) + + +def init_seeds(seed=0): + # Initialize random number generator (RNG) seeds + random.seed(seed) + np.random.seed(seed) + init_torch_seeds(seed) + + +def get_latest_run(search_dir='.'): + # Return path to most recent 'last.pt' in /runs (i.e. to --resume from) + last_list = glob.glob(f'{search_dir}/**/last*.pt', recursive=True) + return max(last_list, key=os.path.getctime) if last_list else '' + + +def isdocker(): + # Is environment a Docker container + return Path('/workspace').exists() # or Path('/.dockerenv').exists() + + +def emojis(str=''): + # Return platform-dependent emoji-safe version of string + return str.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else str + + +def check_online(): + # Check internet connectivity + import socket + try: + socket.create_connection(("1.1.1.1", 443), 5) # check host accesability + return True + except OSError: + return False + + +def check_git_status(): + # Recommend 'git pull' if code is out of date + print(colorstr('github: '), end='') + try: + assert Path('.git').exists(), 'skipping check (not a git repository)' + assert not isdocker(), 'skipping check (Docker image)' + assert check_online(), 'skipping check (offline)' + + cmd = 'git fetch && git config --get remote.origin.url' + url = subprocess.check_output(cmd, shell=True).decode().strip().rstrip('.git') # github repo url + branch = subprocess.check_output('git rev-parse --abbrev-ref HEAD', shell=True).decode().strip() # checked out + n = int(subprocess.check_output(f'git rev-list {branch}..origin/master --count', shell=True)) # commits behind + if n > 0: + s = f"⚠️ WARNING: code is out of date by {n} commit{'s' * (n > 1)}. " \ + f"Use 'git pull' to update or 'git clone {url}' to download latest." + else: + s = f'up to date with {url} ✅' + print(emojis(s)) # emoji-safe + except Exception as e: + print(e) + + +def check_requirements(requirements='requirements.txt', exclude=()): + # Check installed dependencies meet requirements (pass *.txt file or list of packages) + import pkg_resources as pkg + prefix = colorstr('red', 'bold', 'requirements:') + if isinstance(requirements, (str, Path)): # requirements.txt file + file = Path(requirements) + if not file.exists(): + print(f"{prefix} {file.resolve()} not found, check failed.") + return + requirements = [f'{x.name}{x.specifier}' for x in pkg.parse_requirements(file.open()) if x.name not in exclude] + else: # list or tuple of packages + requirements = [x for x in requirements if x not in exclude] + + n = 0 # number of packages updates + for r in requirements: + try: + pkg.require(r) + except Exception as e: # DistributionNotFound or VersionConflict if requirements not met + n += 1 + print(f"{prefix} {e.req} not found and is required by YOLOv5, attempting auto-update...") + print(subprocess.check_output(f"pip install '{e.req}'", shell=True).decode()) + + if n: # if packages updated + source = file.resolve() if 'file' in locals() else requirements + s = f"{prefix} {n} package{'s' * (n > 1)} updated per {source}\n" \ + f"{prefix} ⚠️ {colorstr('bold', 'Restart runtime or rerun command for updates to take effect')}\n" + print(emojis(s)) # emoji-safe + + +def check_img_size(img_size, s=32): + # Verify img_size is a multiple of stride s + new_size = make_divisible(img_size, int(s)) # ceil gs-multiple + if new_size != img_size: + print('WARNING: --img-size %g must be multiple of max stride %g, updating to %g' % (img_size, s, new_size)) + return new_size + + +def check_imshow(): + # Check if environment supports image displays + try: + assert not isdocker(), 'cv2.imshow() is disabled in Docker environments' + cv2.imshow('test', np.zeros((1, 1, 3))) + cv2.waitKey(1) + cv2.destroyAllWindows() + cv2.waitKey(1) + return True + except Exception as e: + print(f'WARNING: Environment does not support cv2.imshow() or PIL Image.show() image displays\n{e}') + return False + + +def check_file(file): + # Search for file if not found + if os.path.isfile(file) or file == '': + return file + else: + files = glob.glob('./**/' + file, recursive=True) # find file + assert len(files), 'File Not Found: %s' % file # assert file was found + assert len(files) == 1, "Multiple files match '%s', specify exact path: %s" % (file, files) # assert unique + return files[0] # return file + + +def check_dataset(dict): + # Download dataset if not found locally + val, s = dict.get('val'), dict.get('download') + if val and len(val): + val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])] # val path + if not all(x.exists() for x in val): + print('\nWARNING: Dataset not found, nonexistent paths: %s' % [str(x) for x in val if not x.exists()]) + if s and len(s): # download script + print('Downloading %s ...' % s) + if s.startswith('http') and s.endswith('.zip'): # URL + f = Path(s).name # filename + torch.hub.download_url_to_file(s, f) + r = os.system('unzip -q %s -d ../ && rm %s' % (f, f)) # unzip + else: # bash script + r = os.system(s) + print('Dataset autodownload %s\n' % ('success' if r == 0 else 'failure')) # analyze return value + else: + raise Exception('Dataset not found.') + + +def make_divisible(x, divisor): + # Returns x evenly divisible by divisor + return math.ceil(x / divisor) * divisor + + +def clean_str(s): + # Cleans a string by replacing special characters with underscore _ + return re.sub(pattern="[|@#!¡·$€%&()=?¿^*;:,¨´><+]", repl="_", string=s) + + +def one_cycle(y1=0.0, y2=1.0, steps=100): + # lambda function for sinusoidal ramp from y1 to y2 + return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1 + + +def colorstr(*input): + # Colors a string https://en.wikipedia.org/wiki/ANSI_escape_code, i.e. colorstr('blue', 'hello world') + *args, string = input if len(input) > 1 else ('blue', 'bold', input[0]) # color arguments, string + colors = {'black': '\033[30m', # basic colors + 'red': '\033[31m', + 'green': '\033[32m', + 'yellow': '\033[33m', + 'blue': '\033[34m', + 'magenta': '\033[35m', + 'cyan': '\033[36m', + 'white': '\033[37m', + 'bright_black': '\033[90m', # bright colors + 'bright_red': '\033[91m', + 'bright_green': '\033[92m', + 'bright_yellow': '\033[93m', + 'bright_blue': '\033[94m', + 'bright_magenta': '\033[95m', + 'bright_cyan': '\033[96m', + 'bright_white': '\033[97m', + 'end': '\033[0m', # misc + 'bold': '\033[1m', + 'underline': '\033[4m'} + return ''.join(colors[x] for x in args) + f'{string}' + colors['end'] + + +def labels_to_class_weights(labels, nc=80): + # Get class weights (inverse frequency) from training labels + if labels[0] is None: # no labels loaded + return torch.Tensor() + + labels = np.concatenate(labels, 0) # labels.shape = (866643, 5) for COCO + classes = labels[:, 0].astype(np.int) # labels = [class xywh] + weights = np.bincount(classes, minlength=nc) # occurrences per class + + # Prepend gridpoint count (for uCE training) + # gpi = ((320 / 32 * np.array([1, 2, 4])) ** 2 * 3).sum() # gridpoints per image + # weights = np.hstack([gpi * len(labels) - weights.sum() * 9, weights * 9]) ** 0.5 # prepend gridpoints to start + + weights[weights == 0] = 1 # replace empty bins with 1 + weights = 1 / weights # number of targets per class + weights /= weights.sum() # normalize + return torch.from_numpy(weights) + + +def labels_to_image_weights(labels, nc=80, class_weights=np.ones(80)): + # Produces image weights based on class_weights and image contents + class_counts = np.array([np.bincount(x[:, 0].astype(np.int), minlength=nc) for x in labels]) + image_weights = (class_weights.reshape(1, nc) * class_counts).sum(1) + # index = random.choices(range(n), weights=image_weights, k=1) # weight image sample + return image_weights + + +def coco80_to_coco91_class(): # converts 80-index (val2014) to 91-index (paper) + # https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/ + # a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n') + # b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n') + # x1 = [list(a[i] == b).index(True) + 1 for i in range(80)] # darknet to coco + # x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)] # coco to darknet + x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, + 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90] + return x + + +def xyxy2xywh(x): + # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center + y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center + y[:, 2] = x[:, 2] - x[:, 0] # width + y[:, 3] = x[:, 3] - x[:, 1] # height + return y + + +def xywh2xyxy(x): + # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x + y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y + y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x + y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y + return y + + +def xywhn2xyxy(x, w=640, h=640, padw=0, padh=0): + # Convert nx4 boxes from [x, y, w, h] normalized to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = w * (x[:, 0] - x[:, 2] / 2) + padw # top left x + y[:, 1] = h * (x[:, 1] - x[:, 3] / 2) + padh # top left y + y[:, 2] = w * (x[:, 0] + x[:, 2] / 2) + padw # bottom right x + y[:, 3] = h * (x[:, 1] + x[:, 3] / 2) + padh # bottom right y + return y + + +def xyn2xy(x, w=640, h=640, padw=0, padh=0): + # Convert normalized segments into pixel segments, shape (n,2) + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = w * x[:, 0] + padw # top left x + y[:, 1] = h * x[:, 1] + padh # top left y + return y + + +def segment2box(segment, width=640, height=640): + # Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy) + x, y = segment.T # segment xy + inside = (x >= 0) & (y >= 0) & (x <= width) & (y <= height) + x, y, = x[inside], y[inside] + return np.array([x.min(), y.min(), x.max(), y.max()]) if any(x) else np.zeros((1, 4)) # xyxy + + +def segments2boxes(segments): + # Convert segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh) + boxes = [] + for s in segments: + x, y = s.T # segment xy + boxes.append([x.min(), y.min(), x.max(), y.max()]) # cls, xyxy + return xyxy2xywh(np.array(boxes)) # cls, xywh + + +def resample_segments(segments, n=1000): + # Up-sample an (n,2) segment + for i, s in enumerate(segments): + x = np.linspace(0, len(s) - 1, n) + xp = np.arange(len(s)) + segments[i] = np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)]).reshape(2, -1).T # segment xy + return segments + + +def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None): + # Rescale coords (xyxy) from img1_shape to img0_shape + if ratio_pad is None: # calculate from img0_shape + gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1]) # gain = old / new + pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2 # wh padding + else: + gain = ratio_pad[0][0] + pad = ratio_pad[1] + + coords[:, [0, 2]] -= pad[0] # x padding + coords[:, [1, 3]] -= pad[1] # y padding + coords[:, :4] /= gain + clip_coords(coords, img0_shape) + return coords + + +def clip_coords(boxes, img_shape): + # Clip bounding xyxy bounding boxes to image shape (height, width) + boxes[:, 0].clamp_(0, img_shape[1]) # x1 + boxes[:, 1].clamp_(0, img_shape[0]) # y1 + boxes[:, 2].clamp_(0, img_shape[1]) # x2 + boxes[:, 3].clamp_(0, img_shape[0]) # y2 + + +def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7): + # Returns the IoU of box1 to box2. box1 is 4, box2 is nx4 + box2 = box2.T + + # Get the coordinates of bounding boxes + if x1y1x2y2: # x1, y1, x2, y2 = box1 + b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3] + b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3] + else: # transform from xywh to xyxy + b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2 + b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2 + b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2 + b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2 + + # Intersection area + inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \ + (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0) + + # Union Area + w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps + w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps + union = w1 * h1 + w2 * h2 - inter + eps + + iou = inter / union + if GIoU or DIoU or CIoU: + cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex (smallest enclosing box) width + ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height + if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1 + c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared + rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + + (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center distance squared + if DIoU: + return iou - rho2 / c2 # DIoU + elif CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47 + v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2) + with torch.no_grad(): + alpha = v / (v - iou + (1 + eps)) + return iou - (rho2 / c2 + v * alpha) # CIoU + else: # GIoU https://arxiv.org/pdf/1902.09630.pdf + c_area = cw * ch + eps # convex area + return iou - (c_area - union) / c_area # GIoU + else: + return iou # IoU + + +def box_iou(box1, box2): + # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py + """ + Return intersection-over-union (Jaccard index) of boxes. + Both sets of boxes are expected to be in (x1, y1, x2, y2) format. + Arguments: + box1 (Tensor[N, 4]) + box2 (Tensor[M, 4]) + Returns: + iou (Tensor[N, M]): the NxM matrix containing the pairwise + IoU values for every element in boxes1 and boxes2 + """ + + def box_area(box): + # box = 4xn + return (box[2] - box[0]) * (box[3] - box[1]) + + area1 = box_area(box1.T) + area2 = box_area(box2.T) + + # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2) + inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2) + return inter / (area1[:, None] + area2 - inter) # iou = inter / (area1 + area2 - inter) + + +def wh_iou(wh1, wh2): + # Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2 + wh1 = wh1[:, None] # [N,1,2] + wh2 = wh2[None] # [1,M,2] + inter = torch.min(wh1, wh2).prod(2) # [N,M] + return inter / (wh1.prod(2) + wh2.prod(2) - inter) # iou = inter / (area1 + area2 - inter) + + +def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False, + labels=()): + """Runs Non-Maximum Suppression (NMS) on inference results + + Returns: + list of detections, on (n,6) tensor per image [xyxy, conf, cls] + """ + + nc = prediction.shape[2] - 5 # number of classes + xc = prediction[..., 4] > conf_thres # candidates + + # Settings + min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height + max_det = 300 # maximum number of detections per image + max_nms = 30000 # maximum number of boxes into torchvision.ops.nms() + time_limit = 10.0 # seconds to quit after + redundant = True # require redundant detections + multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img) + merge = False # use merge-NMS + + t = time.time() + output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0] + for xi, x in enumerate(prediction): # image index, image inference + # Apply constraints + # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height + x = x[xc[xi]] # confidence + + # Cat apriori labels if autolabelling + if labels and len(labels[xi]): + l = labels[xi] + v = torch.zeros((len(l), nc + 5), device=x.device) + v[:, :4] = l[:, 1:5] # box + v[:, 4] = 1.0 # conf + v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls + x = torch.cat((x, v), 0) + + # If none remain process next image + if not x.shape[0]: + continue + + # Compute conf + x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf + + # Box (center x, center y, width, height) to (x1, y1, x2, y2) + box = xywh2xyxy(x[:, :4]) + + # Detections matrix nx6 (xyxy, conf, cls) + if multi_label: + i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T + x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1) + else: # best class only + conf, j = x[:, 5:].max(1, keepdim=True) + x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres] + + # Filter by class + if classes is not None: + x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)] + + # Apply finite constraint + # if not torch.isfinite(x).all(): + # x = x[torch.isfinite(x).all(1)] + + # Check shape + n = x.shape[0] # number of boxes + if not n: # no boxes + continue + elif n > max_nms: # excess boxes + x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence + + # Batched NMS + c = x[:, 5:6] * (0 if agnostic else max_wh) # classes + boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores + i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS + if i.shape[0] > max_det: # limit detections + i = i[:max_det] + if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean) + # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4) + iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix + weights = iou * scores[None] # box weights + x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes + if redundant: + i = i[iou.sum(1) > 1] # require redundancy + + output[xi] = x[i] + if (time.time() - t) > time_limit: + print(f'WARNING: NMS time limit {time_limit}s exceeded') + break # time limit exceeded + + return output + + +def strip_optimizer(f='best.pt', s=''): # from utils.general import *; strip_optimizer() + # Strip optimizer from 'f' to finalize training, optionally save as 's' + x = torch.load(f, map_location=torch.device('cpu')) + if x.get('ema'): + x['model'] = x['ema'] # replace model with ema + for k in 'optimizer', 'training_results', 'wandb_id', 'ema', 'updates': # keys + x[k] = None + x['epoch'] = -1 + x['model'].half() # to FP16 + for p in x['model'].parameters(): + p.requires_grad = False + torch.save(x, s or f) + mb = os.path.getsize(s or f) / 1E6 # filesize + print(f"Optimizer stripped from {f},{(' saved as %s,' % s) if s else ''} {mb:.1f}MB") + + +def print_mutation(hyp, results, yaml_file='hyp_evolved.yaml', bucket=''): + # Print mutation results to evolve.txt (for use with train.py --evolve) + a = '%10s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys + b = '%10.3g' * len(hyp) % tuple(hyp.values()) # hyperparam values + c = '%10.4g' * len(results) % results # results (P, R, mAP@0.5, mAP@0.5:0.95, val_losses x 3) + print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c)) + + if bucket: + url = 'gs://%s/evolve.txt' % bucket + if gsutil_getsize(url) > (os.path.getsize('evolve.txt') if os.path.exists('evolve.txt') else 0): + os.system('gsutil cp %s .' % url) # download evolve.txt if larger than local + + with open('evolve.txt', 'a') as f: # append result + f.write(c + b + '\n') + x = np.unique(np.loadtxt('evolve.txt', ndmin=2), axis=0) # load unique rows + x = x[np.argsort(-fitness(x))] # sort + np.savetxt('evolve.txt', x, '%10.3g') # save sort by fitness + + # Save yaml + for i, k in enumerate(hyp.keys()): + hyp[k] = float(x[0, i + 7]) + with open(yaml_file, 'w') as f: + results = tuple(x[0, :7]) + c = '%10.4g' * len(results) % results # results (P, R, mAP@0.5, mAP@0.5:0.95, val_losses x 3) + f.write('# Hyperparameter Evolution Results\n# Generations: %g\n# Metrics: ' % len(x) + c + '\n\n') + yaml.dump(hyp, f, sort_keys=False) + + if bucket: + os.system('gsutil cp evolve.txt %s gs://%s' % (yaml_file, bucket)) # upload + + +def apply_classifier(x, model, img, im0): + # applies a second stage classifier to yolo outputs + im0 = [im0] if isinstance(im0, np.ndarray) else im0 + for i, d in enumerate(x): # per image + if d is not None and len(d): + d = d.clone() + + # Reshape and pad cutouts + b = xyxy2xywh(d[:, :4]) # boxes + b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # rectangle to square + b[:, 2:] = b[:, 2:] * 1.3 + 30 # pad + d[:, :4] = xywh2xyxy(b).long() + + # Rescale boxes from img_size to im0 size + scale_coords(img.shape[2:], d[:, :4], im0[i].shape) + + # Classes + pred_cls1 = d[:, 5].long() + ims = [] + for j, a in enumerate(d): # per item + cutout = im0[i][int(a[1]):int(a[3]), int(a[0]):int(a[2])] + im = cv2.resize(cutout, (224, 224)) # BGR + # cv2.imwrite('test%i.jpg' % j, cutout) + + im = im[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 + im = np.ascontiguousarray(im, dtype=np.float32) # uint8 to float32 + im /= 255.0 # 0 - 255 to 0.0 - 1.0 + ims.append(im) + + pred_cls2 = model(torch.Tensor(ims).to(d.device)).argmax(1) # classifier prediction + x[i] = x[i][pred_cls1 == pred_cls2] # retain matching class detections + + return x + + +def increment_path(path, exist_ok=True, sep=''): + # Increment path, i.e. runs/exp --> runs/exp{sep}0, runs/exp{sep}1 etc. + path = Path(path) # os-agnostic + if (path.exists() and exist_ok) or (not path.exists()): + return str(path) + else: + dirs = glob.glob(f"{path}{sep}*") # similar paths + matches = [re.search(rf"%s{sep}(\d+)" % path.stem, d) for d in dirs] + i = [int(m.groups()[0]) for m in matches if m] # indices + n = max(i) + 1 if i else 2 # increment number + return f"{path}{sep}{n}" # update path diff --git a/utils/google_app_engine/Dockerfile b/utils/google_app_engine/Dockerfile new file mode 100644 index 0000000..0155618 --- /dev/null +++ b/utils/google_app_engine/Dockerfile @@ -0,0 +1,25 @@ +FROM gcr.io/google-appengine/python + +# Create a virtualenv for dependencies. This isolates these packages from +# system-level packages. +# Use -p python3 or -p python3.7 to select python version. Default is version 2. +RUN virtualenv /env -p python3 + +# Setting these environment variables are the same as running +# source /env/bin/activate. +ENV VIRTUAL_ENV /env +ENV PATH /env/bin:$PATH + +RUN apt-get update && apt-get install -y python-opencv + +# Copy the application's requirements.txt and run pip to install all +# dependencies into the virtualenv. +ADD requirements.txt /app/requirements.txt +RUN pip install -r /app/requirements.txt + +# Add the application source code. +ADD . /app + +# Run a WSGI server to serve the application. gunicorn must be declared as +# a dependency in requirements.txt. +CMD gunicorn -b :$PORT main:app diff --git a/utils/google_app_engine/additional_requirements.txt b/utils/google_app_engine/additional_requirements.txt new file mode 100644 index 0000000..5fcc305 --- /dev/null +++ b/utils/google_app_engine/additional_requirements.txt @@ -0,0 +1,4 @@ +# add these requirements in your app on top of the existing ones +pip==18.1 +Flask==1.0.2 +gunicorn==19.9.0 diff --git a/utils/google_app_engine/app.yaml b/utils/google_app_engine/app.yaml new file mode 100644 index 0000000..ac29d10 --- /dev/null +++ b/utils/google_app_engine/app.yaml @@ -0,0 +1,14 @@ +runtime: custom +env: flex + +service: yolov5app + +liveness_check: + initial_delay_sec: 600 + +manual_scaling: + instances: 1 +resources: + cpu: 1 + memory_gb: 4 + disk_size_gb: 20 \ No newline at end of file diff --git a/utils/google_utils.py b/utils/google_utils.py new file mode 100644 index 0000000..0a7ca3b --- /dev/null +++ b/utils/google_utils.py @@ -0,0 +1,122 @@ +# Google utils: https://cloud.google.com/storage/docs/reference/libraries + +import os +import platform +import subprocess +import time +from pathlib import Path + +import requests +import torch + + +def gsutil_getsize(url=''): + # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du + s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8') + return eval(s.split(' ')[0]) if len(s) else 0 # bytes + + +def attempt_download(file, repo='ultralytics/yolov5'): + # Attempt file download if does not exist + file = Path(str(file).strip().replace("'", '').lower()) + + if not file.exists(): + try: + response = requests.get(f'https://api.github.com/repos/{repo}/releases/latest').json() # github api + assets = [x['name'] for x in response['assets']] # release assets, i.e. ['yolov5s.pt', 'yolov5m.pt', ...] + tag = response['tag_name'] # i.e. 'v1.0' + except: # fallback plan + assets = ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt'] + tag = subprocess.check_output('git tag', shell=True).decode().split()[-1] + + name = file.name + if name in assets: + msg = f'{file} missing, try downloading from https://github.com/{repo}/releases/' + redundant = False # second download option + try: # GitHub + url = f'https://github.com/{repo}/releases/download/{tag}/{name}' + print(f'Downloading {url} to {file}...') + torch.hub.download_url_to_file(url, file) + assert file.exists() and file.stat().st_size > 1E6 # check + except Exception as e: # GCP + print(f'Download error: {e}') + assert redundant, 'No secondary mirror' + url = f'https://storage.googleapis.com/{repo}/ckpt/{name}' + print(f'Downloading {url} to {file}...') + os.system(f'curl -L {url} -o {file}') # torch.hub.download_url_to_file(url, weights) + finally: + if not file.exists() or file.stat().st_size < 1E6: # check + file.unlink(missing_ok=True) # remove partial downloads + print(f'ERROR: Download failure: {msg}') + print('') + return + + +def gdrive_download(id='16TiPfZj7htmTyhntwcZyEEAejOUxuT6m', file='tmp.zip'): + # Downloads a file from Google Drive. from yolov5.utils.google_utils import *; gdrive_download() + t = time.time() + file = Path(file) + cookie = Path('cookie') # gdrive cookie + print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='') + file.unlink(missing_ok=True) # remove existing file + cookie.unlink(missing_ok=True) # remove existing cookie + + # Attempt file download + out = "NUL" if platform.system() == "Windows" else "/dev/null" + os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}') + if os.path.exists('cookie'): # large file + s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}' + else: # small file + s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"' + r = os.system(s) # execute, capture return + cookie.unlink(missing_ok=True) # remove existing cookie + + # Error check + if r != 0: + file.unlink(missing_ok=True) # remove partial + print('Download error ') # raise Exception('Download error') + return r + + # Unzip if archive + if file.suffix == '.zip': + print('unzipping... ', end='') + os.system(f'unzip -q {file}') # unzip + file.unlink() # remove zip to free space + + print(f'Done ({time.time() - t:.1f}s)') + return r + + +def get_token(cookie="./cookie"): + with open(cookie) as f: + for line in f: + if "download" in line: + return line.split()[-1] + return "" + +# def upload_blob(bucket_name, source_file_name, destination_blob_name): +# # Uploads a file to a bucket +# # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python +# +# storage_client = storage.Client() +# bucket = storage_client.get_bucket(bucket_name) +# blob = bucket.blob(destination_blob_name) +# +# blob.upload_from_filename(source_file_name) +# +# print('File {} uploaded to {}.'.format( +# source_file_name, +# destination_blob_name)) +# +# +# def download_blob(bucket_name, source_blob_name, destination_file_name): +# # Uploads a blob from a bucket +# storage_client = storage.Client() +# bucket = storage_client.get_bucket(bucket_name) +# blob = bucket.blob(source_blob_name) +# +# blob.download_to_filename(destination_file_name) +# +# print('Blob {} downloaded to {}.'.format( +# source_blob_name, +# destination_file_name)) diff --git a/utils/loss.py b/utils/loss.py new file mode 100644 index 0000000..9e78df1 --- /dev/null +++ b/utils/loss.py @@ -0,0 +1,216 @@ +# Loss functions + +import torch +import torch.nn as nn + +from utils.general import bbox_iou +from utils.torch_utils import is_parallel + + +def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441 + # return positive, negative label smoothing BCE targets + return 1.0 - 0.5 * eps, 0.5 * eps + + +class BCEBlurWithLogitsLoss(nn.Module): + # BCEwithLogitLoss() with reduced missing label effects. + def __init__(self, alpha=0.05): + super(BCEBlurWithLogitsLoss, self).__init__() + self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss() + self.alpha = alpha + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + pred = torch.sigmoid(pred) # prob from logits + dx = pred - true # reduce only missing label effects + # dx = (pred - true).abs() # reduce missing label and false label effects + alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4)) + loss *= alpha_factor + return loss.mean() + + +class FocalLoss(nn.Module): + # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) + def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): + super(FocalLoss, self).__init__() + self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() + self.gamma = gamma + self.alpha = alpha + self.reduction = loss_fcn.reduction + self.loss_fcn.reduction = 'none' # required to apply FL to each element + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + # p_t = torch.exp(-loss) + # loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability + + # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py + pred_prob = torch.sigmoid(pred) # prob from logits + p_t = true * pred_prob + (1 - true) * (1 - pred_prob) + alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) + modulating_factor = (1.0 - p_t) ** self.gamma + loss *= alpha_factor * modulating_factor + + if self.reduction == 'mean': + return loss.mean() + elif self.reduction == 'sum': + return loss.sum() + else: # 'none' + return loss + + +class QFocalLoss(nn.Module): + # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) + def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): + super(QFocalLoss, self).__init__() + self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() + self.gamma = gamma + self.alpha = alpha + self.reduction = loss_fcn.reduction + self.loss_fcn.reduction = 'none' # required to apply FL to each element + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + + pred_prob = torch.sigmoid(pred) # prob from logits + alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) + modulating_factor = torch.abs(true - pred_prob) ** self.gamma + loss *= alpha_factor * modulating_factor + + if self.reduction == 'mean': + return loss.mean() + elif self.reduction == 'sum': + return loss.sum() + else: # 'none' + return loss + + +class ComputeLoss: + # Compute losses + def __init__(self, model, autobalance=False): + super(ComputeLoss, self).__init__() + device = next(model.parameters()).device # get model device + h = model.hyp # hyperparameters + + # Define criteria + BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device)) + BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device)) + + # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3 + self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets + + # Focal loss + g = h['fl_gamma'] # focal loss gamma + if g > 0: + BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g) + + det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module + self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7 + self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index + self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance + for k in 'na', 'nc', 'nl', 'anchors': + setattr(self, k, getattr(det, k)) + + def __call__(self, p, targets): # predictions, targets, model + device = targets.device + lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device) + tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets + + # Losses + for i, pi in enumerate(p): # layer index, layer predictions + b, a, gj, gi = indices[i] # image, anchor, gridy, gridx + tobj = torch.zeros_like(pi[..., 0], device=device) # target obj + + n = b.shape[0] # number of targets + if n: + ps = pi[b, a, gj, gi] # prediction subset corresponding to targets + + # Regression + pxy = ps[:, :2].sigmoid() * 2. - 0.5 + pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i] + pbox = torch.cat((pxy, pwh), 1) # predicted box + iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # iou(prediction, target) + lbox += (1.0 - iou).mean() # iou loss + + # Objectness + tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio + + # Classification + if self.nc > 1: # cls loss (only if multiple classes) + t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets + t[range(n), tcls[i]] = self.cp + lcls += self.BCEcls(ps[:, 5:], t) # BCE + + # Append targets to text file + # with open('targets.txt', 'a') as file: + # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)] + + obji = self.BCEobj(pi[..., 4], tobj) + lobj += obji * self.balance[i] # obj loss + if self.autobalance: + self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item() + + if self.autobalance: + self.balance = [x / self.balance[self.ssi] for x in self.balance] + lbox *= self.hyp['box'] + lobj *= self.hyp['obj'] + lcls *= self.hyp['cls'] + bs = tobj.shape[0] # batch size + + loss = lbox + lobj + lcls + return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach() + + def build_targets(self, p, targets): + # Build targets for compute_loss(), input targets(image,class,x,y,w,h) + na, nt = self.na, targets.shape[0] # number of anchors, targets + tcls, tbox, indices, anch = [], [], [], [] + gain = torch.ones(7, device=targets.device) # normalized to gridspace gain + ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt) + targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices + + g = 0.5 # bias + off = torch.tensor([[0, 0], + [1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m + # [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm + ], device=targets.device).float() * g # offsets + + for i in range(self.nl): + anchors = self.anchors[i] + gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain + + # Match targets to anchors + t = targets * gain + if nt: + # Matches + r = t[:, :, 4:6] / anchors[:, None] # wh ratio + j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare + # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2)) + t = t[j] # filter + + # Offsets + gxy = t[:, 2:4] # grid xy + gxi = gain[[2, 3]] - gxy # inverse + j, k = ((gxy % 1. < g) & (gxy > 1.)).T + l, m = ((gxi % 1. < g) & (gxi > 1.)).T + j = torch.stack((torch.ones_like(j), j, k, l, m)) + t = t.repeat((5, 1, 1))[j] + offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j] + else: + t = targets[0] + offsets = 0 + + # Define + b, c = t[:, :2].long().T # image, class + gxy = t[:, 2:4] # grid xy + gwh = t[:, 4:6] # grid wh + gij = (gxy - offsets).long() + gi, gj = gij.T # grid xy indices + + # Append + a = t[:, 6].long() # anchor indices + indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices + tbox.append(torch.cat((gxy - gij, gwh), 1)) # box + anch.append(anchors[a]) # anchors + tcls.append(c) # class + + return tcls, tbox, indices, anch diff --git a/utils/metrics.py b/utils/metrics.py new file mode 100644 index 0000000..ec4b363 --- /dev/null +++ b/utils/metrics.py @@ -0,0 +1,276 @@ +# Model validation metrics + +from pathlib import Path + +import matplotlib.pyplot as plt +import numpy as np +import torch + +from . import general + +class Evaluator(object): + def __init__(self, num_class): + self.num_class = num_class + self.confusion_matrix = np.zeros((self.num_class,)*2) + + def Recall_Precision(self): + TP= np.diag(self.confusion_matrix) + TP_add_FN=self.confusion_matrix.sum(axis=1)#每一行的和 + TP_add_FP=self.confusion_matrix.sum(axis=0)#每一列的和 + Recall=TP / TP_add_FN + Precision=TP / TP_add_FP + F1=2*(Precision*Recall)/(Precision+Recall) + # recall= np.nanmean(Recall) + # precision = np.nanmean(Precision) + # f1= np.nanmean(F1) + + return Recall,Precision,F1 + + def Pixel_Accuracy(self): + Acc = np.diag(self.confusion_matrix).sum() / self.confusion_matrix.sum() + return Acc + + def Pixel_Accuracy_Class(self): + Acc = np.diag(self.confusion_matrix) / self.confusion_matrix.sum(axis=1) + Acc = np.nanmean(Acc) + return Acc + + def Mean_Intersection_over_Union(self): + class_IoU = np.diag(self.confusion_matrix) / ( + np.sum(self.confusion_matrix, axis=1) + np.sum(self.confusion_matrix, axis=0) - + np.diag(self.confusion_matrix)) + MIoU = np.nanmean(class_IoU) + return class_IoU,MIoU + + def Frequency_Weighted_Intersection_over_Union(self): + freq = np.sum(self.confusion_matrix, axis=1) / np.sum(self.confusion_matrix) + iu = np.diag(self.confusion_matrix) / ( + np.sum(self.confusion_matrix, axis=1) + np.sum(self.confusion_matrix, axis=0) - + np.diag(self.confusion_matrix)) + + FWIoU = (freq[freq > 0] * iu[freq > 0]).sum() + return FWIoU + + def _generate_matrix(self, gt_image, pre_image): + mask = (gt_image >= 0) & (gt_image < self.num_class) + label = self.num_class * gt_image[mask].astype('int') + pre_image[mask] + count = np.bincount(label, minlength=self.num_class**2) + confusion_matrix = count.reshape(self.num_class, self.num_class) + return confusion_matrix + + def add_batch(self, gt_image, pre_image): + assert gt_image.shape == pre_image.shape + self.confusion_matrix += self._generate_matrix(gt_image, pre_image) + + def reset(self): + self.confusion_matrix = np.zeros((self.num_class,) * 2) + + + + +def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=()): + """ Compute the average precision, given the recall and precision curves. + Source: https://github.com/rafaelpadilla/Object-Detection-Metrics. + # Arguments + tp: True positives (nparray, nx1 or nx10). + conf: Objectness value from 0-1 (nparray). + pred_cls: Predicted object classes (nparray). + target_cls: True object classes (nparray). + plot: Plot precision-recall curve at mAP@0.5 + save_dir: Plot save directory + # Returns + The average precision as computed in py-faster-rcnn. + """ + + # Sort by objectness + i = np.argsort(-conf) + tp, conf, pred_cls = tp[i], conf[i], pred_cls[i] + + # Find unique classes + unique_classes = np.unique(target_cls) + nc = unique_classes.shape[0] # number of classes, number of detections + + # Create Precision-Recall curve and compute AP for each class + px, py = np.linspace(0, 1, 1000), [] # for plotting + ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000)) + for ci, c in enumerate(unique_classes): + i = pred_cls == c + n_l = (target_cls == c).sum() # number of labels + n_p = i.sum() # number of predictions + + if n_p == 0 or n_l == 0: + continue + else: + # Accumulate FPs and TPs + fpc = (1 - tp[i]).cumsum(0) + tpc = tp[i].cumsum(0) + + # Recall + recall = tpc / (n_l + 1e-16) # recall curve + r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases + + # Precision + precision = tpc / (tpc + fpc) # precision curve + p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score + + # AP from recall-precision curve + for j in range(tp.shape[1]): + ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j]) + if plot and j == 0: + py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5 + + # Compute F1 (harmonic mean of precision and recall) + f1 = 2 * p * r / (p + r + 1e-16) + if plot: + plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names) + plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1') + plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision') + plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall') + + i = f1.mean(0).argmax() # max F1 index + return p[:, i], r[:, i], ap, f1[:, i], unique_classes.astype('int32') + + +def compute_ap(recall, precision): + """ Compute the average precision, given the recall and precision curves + # Arguments + recall: The recall curve (list) + precision: The precision curve (list) + # Returns + Average precision, precision curve, recall curve + """ + + # Append sentinel values to beginning and end + mrec = np.concatenate(([0.], recall, [recall[-1] + 0.01])) + mpre = np.concatenate(([1.], precision, [0.])) + + # Compute the precision envelope + mpre = np.flip(np.maximum.accumulate(np.flip(mpre))) + + # Integrate area under curve + method = 'interp' # methods: 'continuous', 'interp' + if method == 'interp': + x = np.linspace(0, 1, 101) # 101-point interp (COCO) + ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate + else: # 'continuous' + i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes + ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve + + return ap, mpre, mrec + + +class ConfusionMatrix: + # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix + def __init__(self, nc, conf=0.25, iou_thres=0.45): + self.matrix = np.zeros((nc + 1, nc + 1)) + self.nc = nc # number of classes + self.conf = conf + self.iou_thres = iou_thres + + def process_batch(self, detections, labels): + """ + Return intersection-over-union (Jaccard index) of boxes. + Both sets of boxes are expected to be in (x1, y1, x2, y2) format. + Arguments: + detections (Array[N, 6]), x1, y1, x2, y2, conf, class + labels (Array[M, 5]), class, x1, y1, x2, y2 + Returns: + None, updates confusion matrix accordingly + """ + detections = detections[detections[:, 4] > self.conf] + gt_classes = labels[:, 0].int() + detection_classes = detections[:, 5].int() + iou = general.box_iou(labels[:, 1:], detections[:, :4]) + + x = torch.where(iou > self.iou_thres) + if x[0].shape[0]: + matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() + if x[0].shape[0] > 1: + matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 1], return_index=True)[1]] + matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 0], return_index=True)[1]] + else: + matches = np.zeros((0, 3)) + + n = matches.shape[0] > 0 + m0, m1, _ = matches.transpose().astype(np.int16) + for i, gc in enumerate(gt_classes): + j = m0 == i + if n and sum(j) == 1: + self.matrix[gc, detection_classes[m1[j]]] += 1 # correct + else: + self.matrix[self.nc, gc] += 1 # background FP + + if n: + for i, dc in enumerate(detection_classes): + if not any(m1 == i): + self.matrix[dc, self.nc] += 1 # background FN + + def matrix(self): + return self.matrix + + def plot(self, save_dir='', names=()): + try: + import seaborn as sn + + array = self.matrix / (self.matrix.sum(0).reshape(1, self.nc + 1) + 1E-6) # normalize + array[array < 0.005] = np.nan # don't annotate (would appear as 0.00) + + fig = plt.figure(figsize=(12, 9), tight_layout=True) + sn.set(font_scale=1.0 if self.nc < 50 else 0.8) # for label size + labels = (0 < len(names) < 99) and len(names) == self.nc # apply names to ticklabels + sn.heatmap(array, annot=self.nc < 30, annot_kws={"size": 8}, cmap='Blues', fmt='.2f', square=True, + xticklabels=names + ['background FP'] if labels else "auto", + yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1)) + fig.axes[0].set_xlabel('True') + fig.axes[0].set_ylabel('Predicted') + fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250) + except Exception as e: + pass + + def print(self): + for i in range(self.nc + 1): + print(' '.join(map(str, self.matrix[i]))) + + +# Plots ---------------------------------------------------------------------------------------------------------------- + +def plot_pr_curve(px, py, ap, save_dir='pr_curve.png', names=()): + # Precision-recall curve + fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) + py = np.stack(py, axis=1) + + if 0 < len(names) < 21: # display per-class legend if < 21 classes + for i, y in enumerate(py.T): + ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}') # plot(recall, precision) + else: + ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision) + + ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean()) + ax.set_xlabel('Recall') + ax.set_ylabel('Precision') + ax.set_xlim(0, 1) + ax.set_ylim(0, 1) + plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") + fig.savefig(Path(save_dir), dpi=250) + + +def plot_mc_curve(px, py, save_dir='mc_curve.png', names=(), xlabel='Confidence', ylabel='Metric'): + # Metric-confidence curve + fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) + + if 0 < len(names) < 21: # display per-class legend if < 21 classes + for i, y in enumerate(py): + ax.plot(px, y, linewidth=1, label=f'{names[i]}') # plot(confidence, metric) + else: + ax.plot(px, py.T, linewidth=1, color='grey') # plot(confidence, metric) + + y = py.mean(0) + ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}') + ax.set_xlabel(xlabel) + ax.set_ylabel(ylabel) + ax.set_xlim(0, 1) + ax.set_ylim(0, 1) + plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") + fig.savefig(Path(save_dir), dpi=250) diff --git a/utils/plots.py b/utils/plots.py new file mode 100644 index 0000000..4042c45 --- /dev/null +++ b/utils/plots.py @@ -0,0 +1,436 @@ +# Plotting utils + +import glob +import math +import os +import random +from copy import copy +from pathlib import Path + +import cv2 +import matplotlib +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd +import seaborn as sns +import torch +import yaml +from PIL import Image, ImageDraw, ImageFont +from scipy.signal import butter, filtfilt + +from utils.general import xywh2xyxy, xyxy2xywh +from utils.metrics import fitness + +# Settings +matplotlib.rc('font', **{'size': 11}) +matplotlib.use('Agg') # for writing to files only + + +def color_list(): + # Return first 10 plt colors as (r,g,b) https://stackoverflow.com/questions/51350872/python-from-color-name-to-rgb + def hex2rgb(h): + return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4)) + + return [hex2rgb(h) for h in matplotlib.colors.TABLEAU_COLORS.values()] # or BASE_ (8), CSS4_ (148), XKCD_ (949) + + +def hist2d(x, y, n=100): + # 2d histogram used in labels.png and evolve.png + xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n) + hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges)) + xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1) + yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1) + return np.log(hist[xidx, yidx]) + + +def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5): + # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy + def butter_lowpass(cutoff, fs, order): + nyq = 0.5 * fs + normal_cutoff = cutoff / nyq + return butter(order, normal_cutoff, btype='low', analog=False) + + b, a = butter_lowpass(cutoff, fs, order=order) + return filtfilt(b, a, data) # forward-backward filter + + +def plot_one_box(x, img, color=None, label=None, line_thickness=3): + # Plots one bounding box on image img + # tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1 # line/font thickness + color = color or [random.randint(0, 255) for _ in range(3)] + c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3])) + print(6666666666666666) + print(c1, c2) + print(7777777777777777) + cv2.rectangle(img, c1, c2, color, thickness=1, lineType=cv2.LINE_AA) + # if label: + # tf = max(tl - 1, 1) # font thickness + # t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] + # c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3 + # cv2.rectangle(img, c1, c2, color, thickness=1, lineType=cv2.LINE_AA) # filled + # cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA) + + +def plot_one_box_PIL(box, img, color=None, label=None, line_thickness=None): + img = Image.fromarray(img) + draw = ImageDraw.Draw(img) + line_thickness = line_thickness or max(int(min(img.size) / 200), 2) + draw.rectangle(box, width=line_thickness, outline=tuple(color)) # plot + if label: + fontsize = max(round(max(img.size) / 40), 12) + font = ImageFont.truetype("Arial.ttf", fontsize) + txt_width, txt_height = font.getsize(label) + draw.rectangle([box[0], box[1] - txt_height + 4, box[0] + txt_width, box[1]], fill=tuple(color)) + draw.text((box[0], box[1] - txt_height + 1), label, fill=(255, 255, 255), font=font) + return np.asarray(img) + + +def plot_wh_methods(): # from utils.plots import *; plot_wh_methods() + # Compares the two methods for width-height anchor multiplication + # https://github.com/ultralytics/yolov3/issues/168 + x = np.arange(-4.0, 4.0, .1) + ya = np.exp(x) + yb = torch.sigmoid(torch.from_numpy(x)).numpy() * 2 + + fig = plt.figure(figsize=(6, 3), tight_layout=True) + plt.plot(x, ya, '.-', label='YOLOv3') + plt.plot(x, yb ** 2, '.-', label='YOLOv5 ^2') + plt.plot(x, yb ** 1.6, '.-', label='YOLOv5 ^1.6') + plt.xlim(left=-4, right=4) + plt.ylim(bottom=0, top=6) + plt.xlabel('input') + plt.ylabel('output') + plt.grid() + plt.legend() + fig.savefig('comparison.png', dpi=200) + + +def output_to_target(output): + # Convert model output to target format [batch_id, class_id, x, y, w, h, conf] + targets = [] + for i, o in enumerate(output): + for *box, conf, cls in o.cpu().numpy(): + targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf]) + return np.array(targets) + + +def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=640, max_subplots=16): + # Plot image grid with labels + + if isinstance(images, torch.Tensor): + images = images.cpu().float().numpy() + if isinstance(targets, torch.Tensor): + targets = targets.cpu().numpy() + + # un-normalise + if np.max(images[0]) <= 1: + images *= 255 + + tl = 3 # line thickness + tf = max(tl - 1, 1) # font thickness + bs, _, h, w = images.shape # batch size, _, height, width + bs = min(bs, max_subplots) # limit plot images + ns = np.ceil(bs ** 0.5) # number of subplots (square) + + # Check if we should resize + scale_factor = max_size / max(h, w) + if scale_factor < 1: + h = math.ceil(scale_factor * h) + w = math.ceil(scale_factor * w) + + colors = color_list() # list of colors + mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8) # init + for i, img in enumerate(images): + if i == max_subplots: # if last batch has fewer images than we expect + break + + block_x = int(w * (i // ns)) + block_y = int(h * (i % ns)) + + img = img.transpose(1, 2, 0) + if scale_factor < 1: + img = cv2.resize(img, (w, h)) + + mosaic[block_y:block_y + h, block_x:block_x + w, :] = img + if len(targets) > 0: + image_targets = targets[targets[:, 0] == i] + boxes = xywh2xyxy(image_targets[:, 2:6]).T + classes = image_targets[:, 1].astype('int') + labels = image_targets.shape[1] == 6 # labels if no conf column + conf = None if labels else image_targets[:, 6] # check for confidence presence (label vs pred) + + if boxes.shape[1]: + if boxes.max() <= 1.01: # if normalized with tolerance 0.01 + boxes[[0, 2]] *= w # scale to pixels + boxes[[1, 3]] *= h + elif scale_factor < 1: # absolute coords need scale if image scales + boxes *= scale_factor + boxes[[0, 2]] += block_x + boxes[[1, 3]] += block_y + for j, box in enumerate(boxes.T): + cls = int(classes[j]) + color = colors[cls % len(colors)] + cls = names[cls] if names else cls + if labels or conf[j] > 0.25: # 0.25 conf thresh + label = '%s' % cls if labels else '%s %.1f' % (cls, conf[j]) + plot_one_box(box, mosaic, label=label, color=color, line_thickness=tl) + + # Draw image filename labels + if paths: + label = Path(paths[i]).name[:40] # trim to 40 char + t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] + cv2.putText(mosaic, label, (block_x + 5, block_y + t_size[1] + 5), 0, tl / 3, [220, 220, 220], thickness=tf, + lineType=cv2.LINE_AA) + + # Image border + cv2.rectangle(mosaic, (block_x, block_y), (block_x + w, block_y + h), (255, 255, 255), thickness=3) + + if fname: + r = min(1280. / max(h, w) / ns, 1.0) # ratio to limit image size + mosaic = cv2.resize(mosaic, (int(ns * w * r), int(ns * h * r)), interpolation=cv2.INTER_AREA) + # cv2.imwrite(fname, cv2.cvtColor(mosaic, cv2.COLOR_BGR2RGB)) # cv2 save + Image.fromarray(mosaic).save(fname) # PIL save + return mosaic + + +def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''): + # Plot LR simulating training for full epochs + optimizer, scheduler = copy(optimizer), copy(scheduler) # do not modify originals + y = [] + for _ in range(epochs): + scheduler.step() + y.append(optimizer.param_groups[0]['lr']) + plt.plot(y, '.-', label='LR') + plt.xlabel('epoch') + plt.ylabel('LR') + plt.grid() + plt.xlim(0, epochs) + plt.ylim(0) + plt.savefig(Path(save_dir) / 'LR.png', dpi=200) + plt.close() + + +def plot_test_txt(): # from utils.plots import *; plot_test() + # Plot test.txt histograms + x = np.loadtxt('test.txt', dtype=np.float32) + box = xyxy2xywh(x[:, :4]) + cx, cy = box[:, 0], box[:, 1] + + fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True) + ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0) + ax.set_aspect('equal') + plt.savefig('hist2d.png', dpi=300) + + fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True) + ax[0].hist(cx, bins=600) + ax[1].hist(cy, bins=600) + plt.savefig('hist1d.png', dpi=200) + + +def plot_targets_txt(): # from utils.plots import *; plot_targets_txt() + # Plot targets.txt histograms + x = np.loadtxt('targets.txt', dtype=np.float32).T + s = ['x targets', 'y targets', 'width targets', 'height targets'] + fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True) + ax = ax.ravel() + for i in range(4): + ax[i].hist(x[i], bins=100, label='%.3g +/- %.3g' % (x[i].mean(), x[i].std())) + ax[i].legend() + ax[i].set_title(s[i]) + plt.savefig('targets.jpg', dpi=200) + + +def plot_study_txt(path='', x=None): # from utils.plots import *; plot_study_txt() + # Plot study.txt generated by test.py + fig, ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True) + # ax = ax.ravel() + + fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True) + # for f in [Path(path) / f'study_coco_{x}.txt' for x in ['yolov5s6', 'yolov5m6', 'yolov5l6', 'yolov5x6']]: + for f in sorted(Path(path).glob('study*.txt')): + y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T + x = np.arange(y.shape[1]) if x is None else np.array(x) + s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_inference (ms/img)', 't_NMS (ms/img)', 't_total (ms/img)'] + # for i in range(7): + # ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8) + # ax[i].set_title(s[i]) + + j = y[3].argmax() + 1 + ax2.plot(y[6, 1:j], y[3, 1:j] * 1E2, '.-', linewidth=2, markersize=8, + label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO')) + + ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5], + 'k.-', linewidth=2, markersize=8, alpha=.25, label='EfficientDet') + + ax2.grid(alpha=0.2) + ax2.set_yticks(np.arange(20, 60, 5)) + ax2.set_xlim(0, 57) + ax2.set_ylim(30, 55) + ax2.set_xlabel('GPU Speed (ms/img)') + ax2.set_ylabel('COCO AP val') + ax2.legend(loc='lower right') + plt.savefig(str(Path(path).name) + '.png', dpi=300) + + +def plot_labels(labels, names=(), save_dir=Path(''), loggers=None): + # plot dataset labels + print('Plotting labels... ') + c, b = labels[:, 0], labels[:, 1:].transpose() # classes, boxes + nc = int(c.max() + 1) # number of classes + colors = color_list() + x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height']) + + # seaborn correlogram + sns.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9)) + plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200) + plt.close() + + # matplotlib labels + matplotlib.use('svg') # faster + ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel() + ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8) + ax[0].set_ylabel('instances') + if 0 < len(names) < 30: + ax[0].set_xticks(range(len(names))) + ax[0].set_xticklabels(names, rotation=90, fontsize=10) + else: + ax[0].set_xlabel('classes') + sns.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9) + sns.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9) + + # rectangles + labels[:, 1:3] = 0.5 # center + labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000 + img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255) + for cls, *box in labels[:1000]: + ImageDraw.Draw(img).rectangle(box, width=1, outline=colors[int(cls) % 10]) # plot + ax[1].imshow(img) + ax[1].axis('off') + + for a in [0, 1, 2, 3]: + for s in ['top', 'right', 'left', 'bottom']: + ax[a].spines[s].set_visible(False) + + plt.savefig(save_dir / 'labels.jpg', dpi=200) + matplotlib.use('Agg') + plt.close() + + # loggers + for k, v in loggers.items() or {}: + if k == 'wandb' and v: + v.log({"Labels": [v.Image(str(x), caption=x.name) for x in save_dir.glob('*labels*.jpg')]}, commit=False) + + +def plot_evolution(yaml_file='data/hyp.finetune.yaml'): # from utils.plots import *; plot_evolution() + # Plot hyperparameter evolution results in evolve.txt + with open(yaml_file) as f: + hyp = yaml.load(f, Loader=yaml.SafeLoader) + x = np.loadtxt('evolve.txt', ndmin=2) + f = fitness(x) + # weights = (f - f.min()) ** 2 # for weighted results + plt.figure(figsize=(10, 12), tight_layout=True) + matplotlib.rc('font', **{'size': 8}) + for i, (k, v) in enumerate(hyp.items()): + y = x[:, i + 7] + # mu = (y * weights).sum() / weights.sum() # best weighted result + mu = y[f.argmax()] # best single result + plt.subplot(6, 5, i + 1) + plt.scatter(y, f, c=hist2d(y, f, 20), cmap='viridis', alpha=.8, edgecolors='none') + plt.plot(mu, f.max(), 'k+', markersize=15) + plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9}) # limit to 40 characters + if i % 5 != 0: + plt.yticks([]) + print('%15s: %.3g' % (k, mu)) + plt.savefig('evolve.png', dpi=200) + print('\nPlot saved as evolve.png') + + +def profile_idetection(start=0, stop=0, labels=(), save_dir=''): + # Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection() + ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel() + s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS'] + files = list(Path(save_dir).glob('frames*.txt')) + for fi, f in enumerate(files): + try: + results = np.loadtxt(f, ndmin=2).T[:, 90:-30] # clip first and last rows + n = results.shape[1] # number of rows + x = np.arange(start, min(stop, n) if stop else n) + results = results[:, x] + t = (results[0] - results[0].min()) # set t0=0s + results[0] = x + for i, a in enumerate(ax): + if i < len(results): + label = labels[fi] if len(labels) else f.stem.replace('frames_', '') + a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5) + a.set_title(s[i]) + a.set_xlabel('time (s)') + # if fi == len(files) - 1: + # a.set_ylim(bottom=0) + for side in ['top', 'right']: + a.spines[side].set_visible(False) + else: + a.remove() + except Exception as e: + print('Warning: Plotting error for %s; %s' % (f, e)) + + ax[1].legend() + plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200) + + +def plot_results_overlay(start=0, stop=0): # from utils.plots import *; plot_results_overlay() + # Plot training 'results*.txt', overlaying train and val losses + s = ['train', 'train', 'train', 'Precision', 'mAP@0.5', 'val', 'val', 'val', 'Recall', 'mAP@0.5:0.95'] # legends + t = ['Box', 'Objectness', 'Classification', 'P-R', 'mAP-F1'] # titles + for f in sorted(glob.glob('results*.txt') + glob.glob('../../Downloads/results*.txt')): + results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T + n = results.shape[1] # number of rows + x = range(start, min(stop, n) if stop else n) + fig, ax = plt.subplots(1, 5, figsize=(14, 3.5), tight_layout=True) + ax = ax.ravel() + for i in range(5): + for j in [i, i + 5]: + y = results[j, x] + ax[i].plot(x, y, marker='.', label=s[j]) + # y_smooth = butter_lowpass_filtfilt(y) + # ax[i].plot(x, np.gradient(y_smooth), marker='.', label=s[j]) + + ax[i].set_title(t[i]) + ax[i].legend() + ax[i].set_ylabel(f) if i == 0 else None # add filename + fig.savefig(f.replace('.txt', '.png'), dpi=200) + + +def plot_results(start=0, stop=0, bucket='', id=(), labels=(), save_dir=''): + # Plot training 'results*.txt'. from utils.plots import *; plot_results(save_dir='runs/train/exp') + fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True) + ax = ax.ravel() + s = ['Box', 'Objectness', 'Classification', 'Precision', 'Recall', + 'val Box', 'val Objectness', 'val Classification', 'mAP@0.5', 'mAP@0.5:0.95'] + if bucket: + # files = ['https://storage.googleapis.com/%s/results%g.txt' % (bucket, x) for x in id] + files = ['results%g.txt' % x for x in id] + c = ('gsutil cp ' + '%s ' * len(files) + '.') % tuple('gs://%s/results%g.txt' % (bucket, x) for x in id) + os.system(c) + else: + files = list(Path(save_dir).glob('results*.txt')) + assert len(files), 'No results.txt files found in %s, nothing to plot.' % os.path.abspath(save_dir) + for fi, f in enumerate(files): + try: + results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T + n = results.shape[1] # number of rows + x = range(start, min(stop, n) if stop else n) + for i in range(10): + y = results[i, x] + if i in [0, 1, 2, 5, 6, 7]: + y[y == 0] = np.nan # don't show zero loss values + # y /= y[0] # normalize + label = labels[fi] if len(labels) else f.stem + ax[i].plot(x, y, marker='.', label=label, linewidth=2, markersize=8) + ax[i].set_title(s[i]) + # if i in [5, 6, 7]: # share train and val loss y axes + # ax[i].get_shared_y_axes().join(ax[i], ax[i - 5]) + except Exception as e: + print('Warning: Plotting error for %s; %s' % (f, e)) + + ax[1].legend() + fig.savefig(Path(save_dir) / 'results.png', dpi=200) diff --git a/utils/torch_utils.py b/utils/torch_utils.py new file mode 100644 index 0000000..9991e5e --- /dev/null +++ b/utils/torch_utils.py @@ -0,0 +1,303 @@ +# YOLOv5 PyTorch utils + +import datetime +import logging +import math +import os +import platform +import subprocess +import time +from contextlib import contextmanager +from copy import deepcopy +from pathlib import Path + +import torch +import torch.backends.cudnn as cudnn +import torch.nn as nn +import torch.nn.functional as F +import torchvision + +try: + import thop # for FLOPS computation +except ImportError: + thop = None +logger = logging.getLogger(__name__) + + +@contextmanager +def torch_distributed_zero_first(local_rank: int): + """ + Decorator to make all processes in distributed training wait for each local_master to do something. + """ + if local_rank not in [-1, 0]: + torch.distributed.barrier() + yield + if local_rank == 0: + torch.distributed.barrier() + + +def init_torch_seeds(seed=0): + # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html + torch.manual_seed(seed) + if seed == 0: # slower, more reproducible + cudnn.benchmark, cudnn.deterministic = False, True + else: # faster, less reproducible + cudnn.benchmark, cudnn.deterministic = True, False + + +def date_modified(path=__file__): + # return human-readable file modification date, i.e. '2021-3-26' + t = datetime.datetime.fromtimestamp(Path(path).stat().st_mtime) + return f'{t.year}-{t.month}-{t.day}' + + +def git_describe(path=Path(__file__).parent): # path must be a directory + # return human-readable git description, i.e. v5.0-5-g3e25f1e https://git-scm.com/docs/git-describe + s = f'git -C {path} describe --tags --long --always' + try: + return subprocess.check_output(s, shell=True, stderr=subprocess.STDOUT).decode()[:-1] + except subprocess.CalledProcessError as e: + return '' # not a git repository + + +def select_device(device='', batch_size=None): + # device = 'cpu' or '0' or '0,1,2,3' + s = f'YOLOv5 🚀 {git_describe() or date_modified()} torch {torch.__version__} ' # string + cpu = device.lower() == 'cpu' + if cpu: + os.environ['CUDA_VISIBLE_DEVICES'] = '-1' # force torch.cuda.is_available() = False + elif device: # non-cpu device requested + os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable + assert torch.cuda.is_available(), f'CUDA unavailable, invalid device {device} requested' # check availability + + cuda = not cpu and torch.cuda.is_available() + if cuda: + n = torch.cuda.device_count() + if n > 1 and batch_size: # check that batch_size is compatible with device_count + assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}' + space = ' ' * len(s) + for i, d in enumerate(device.split(',') if device else range(n)): + p = torch.cuda.get_device_properties(i) + s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n" # bytes to MB + else: + s += 'CPU\n' + + logger.info(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s) # emoji-safe + return torch.device('cuda:0' if cuda else 'cpu') + + +def time_synchronized(): + # pytorch-accurate time + if torch.cuda.is_available(): + torch.cuda.synchronize() + return time.time() + + +def profile(x, ops, n=100, device=None): + # profile a pytorch module or list of modules. Example usage: + # x = torch.randn(16, 3, 640, 640) # input + # m1 = lambda x: x * torch.sigmoid(x) + # m2 = nn.SiLU() + # profile(x, [m1, m2], n=100) # profile speed over 100 iterations + + device = device or torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') + x = x.to(device) + x.requires_grad = True + print(torch.__version__, device.type, torch.cuda.get_device_properties(0) if device.type == 'cuda' else '') + print(f"\n{'Params':>12s}{'GFLOPS':>12s}{'forward (ms)':>16s}{'backward (ms)':>16s}{'input':>24s}{'output':>24s}") + for m in ops if isinstance(ops, list) else [ops]: + m = m.to(device) if hasattr(m, 'to') else m # device + m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m # type + dtf, dtb, t = 0., 0., [0., 0., 0.] # dt forward, backward + try: + flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 # GFLOPS + except: + flops = 0 + + for _ in range(n): + t[0] = time_synchronized() + y = m(x) + t[1] = time_synchronized() + try: + _ = y.sum().backward() + t[2] = time_synchronized() + except: # no backward method + t[2] = float('nan') + dtf += (t[1] - t[0]) * 1000 / n # ms per op forward + dtb += (t[2] - t[1]) * 1000 / n # ms per op backward + + s_in = tuple(x.shape) if isinstance(x, torch.Tensor) else 'list' + s_out = tuple(y.shape) if isinstance(y, torch.Tensor) else 'list' + p = sum(list(x.numel() for x in m.parameters())) if isinstance(m, nn.Module) else 0 # parameters + print(f'{p:12}{flops:12.4g}{dtf:16.4g}{dtb:16.4g}{str(s_in):>24s}{str(s_out):>24s}') + + +def is_parallel(model): + return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) + + +def intersect_dicts(da, db, exclude=()): + # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values + return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape} + + +def initialize_weights(model): + for m in model.modules(): + t = type(m) + if t is nn.Conv2d: + pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') + elif t is nn.BatchNorm2d: + m.eps = 1e-3 + m.momentum = 0.03 + elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]: + m.inplace = True + + +def find_modules(model, mclass=nn.Conv2d): + # Finds layer indices matching module class 'mclass' + return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)] + + +def sparsity(model): + # Return global model sparsity + a, b = 0., 0. + for p in model.parameters(): + a += p.numel() + b += (p == 0).sum() + return b / a + + +def prune(model, amount=0.3): + # Prune model to requested global sparsity + import torch.nn.utils.prune as prune + print('Pruning model... ', end='') + for name, m in model.named_modules(): + if isinstance(m, nn.Conv2d): + prune.l1_unstructured(m, name='weight', amount=amount) # prune + prune.remove(m, 'weight') # make permanent + print(' %.3g global sparsity' % sparsity(model)) + + +def fuse_conv_and_bn(conv, bn): + # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/ + fusedconv = nn.Conv2d(conv.in_channels, + conv.out_channels, + kernel_size=conv.kernel_size, + stride=conv.stride, + padding=conv.padding, + groups=conv.groups, + bias=True).requires_grad_(False).to(conv.weight.device) + + # prepare filters + w_conv = conv.weight.clone().view(conv.out_channels, -1) + w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) + fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape)) + + # prepare spatial bias + b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias + b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) + fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) + + return fusedconv + + +def model_info(model, verbose=False, img_size=640): + # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320] + n_p = sum(x.numel() for x in model.parameters()) # number parameters + n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients + if verbose: + print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma')) + for i, (name, p) in enumerate(model.named_parameters()): + name = name.replace('module_list.', '') + print('%5g %40s %9s %12g %20s %10.3g %10.3g' % + (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) + + try: # FLOPS + from thop import profile + stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32 + img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device) # input + flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2 # stride GFLOPS + img_size = img_size if isinstance(img_size, list) else [img_size, img_size] # expand if int/float + fs = ', %.1f GFLOPS' % (flops * img_size[0] / stride * img_size[1] / stride) # 640x640 GFLOPS + except (ImportError, Exception): + fs = '' + + logger.info(f"Model Summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}") + + +def load_classifier(name='resnet101', n=2): + # Loads a pretrained model reshaped to n-class output + model = torchvision.models.__dict__[name](pretrained=True) + + # ResNet model properties + # input_size = [3, 224, 224] + # input_space = 'RGB' + # input_range = [0, 1] + # mean = [0.485, 0.456, 0.406] + # std = [0.229, 0.224, 0.225] + + # Reshape output to n classes + filters = model.fc.weight.shape[1] + model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True) + model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True) + model.fc.out_features = n + return model + + +def scale_img(img, ratio=1.0, same_shape=False, gs=32): # img(16,3,256,416) + # scales img(bs,3,y,x) by ratio constrained to gs-multiple + if ratio == 1.0: + return img + else: + h, w = img.shape[2:] + s = (int(h * ratio), int(w * ratio)) # new size + img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize + if not same_shape: # pad/crop img + h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)] + return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean + + +def copy_attr(a, b, include=(), exclude=()): + # Copy attributes from b to a, options to only include [...] and to exclude [...] + for k, v in b.__dict__.items(): + if (len(include) and k not in include) or k.startswith('_') or k in exclude: + continue + else: + setattr(a, k, v) + + +class ModelEMA: + """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models + Keep a moving average of everything in the model state_dict (parameters and buffers). + This is intended to allow functionality like + https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage + A smoothed version of the weights is necessary for some training schemes to perform well. + This class is sensitive where it is initialized in the sequence of model init, + GPU assignment and distributed training wrappers. + """ + + def __init__(self, model, decay=0.9999, updates=0): + # Create EMA + self.ema = deepcopy(model.module if is_parallel(model) else model).eval() # FP32 EMA + # if next(model.parameters()).device.type != 'cpu': + # self.ema.half() # FP16 EMA + self.updates = updates # number of EMA updates + self.decay = lambda x: decay * (1 - math.exp(-x / 2000)) # decay exponential ramp (to help early epochs) + for p in self.ema.parameters(): + p.requires_grad_(False) + + def update(self, model): + # Update EMA parameters + with torch.no_grad(): + self.updates += 1 + d = self.decay(self.updates) + + msd = model.module.state_dict() if is_parallel(model) else model.state_dict() # model state_dict + for k, v in self.ema.state_dict().items(): + if v.dtype.is_floating_point: + v *= d + v += (1. - d) * msd[k].detach() + + def update_attr(self, model, include=(), exclude=('process_group', 'reducer')): + # Update EMA attributes + copy_attr(self.ema, model, include, exclude) diff --git a/utils/wandb_logging/__init__.py b/utils/wandb_logging/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/utils/wandb_logging/__pycache__/__init__.cpython-38.pyc b/utils/wandb_logging/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000..3aa4e5b Binary files /dev/null and b/utils/wandb_logging/__pycache__/__init__.cpython-38.pyc differ diff --git a/utils/wandb_logging/__pycache__/wandb_utils.cpython-38.pyc b/utils/wandb_logging/__pycache__/wandb_utils.cpython-38.pyc new file mode 100644 index 0000000..6c70cd4 Binary files /dev/null and b/utils/wandb_logging/__pycache__/wandb_utils.cpython-38.pyc differ diff --git a/utils/wandb_logging/log_dataset.py b/utils/wandb_logging/log_dataset.py new file mode 100644 index 0000000..d7a521f --- /dev/null +++ b/utils/wandb_logging/log_dataset.py @@ -0,0 +1,24 @@ +import argparse + +import yaml + +from wandb_utils import WandbLogger + +WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' + + +def create_dataset_artifact(opt): + with open(opt.data) as f: + data = yaml.load(f, Loader=yaml.SafeLoader) # data dict + logger = WandbLogger(opt, '', None, data, job_type='Dataset Creation') + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path') + parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset') + parser.add_argument('--project', type=str, default='YOLOv5', help='name of W&B Project') + opt = parser.parse_args() + opt.resume = False # Explicitly disallow resume check for dataset upload job + + create_dataset_artifact(opt) diff --git a/utils/wandb_logging/wandb_utils.py b/utils/wandb_logging/wandb_utils.py new file mode 100644 index 0000000..d8f50ae --- /dev/null +++ b/utils/wandb_logging/wandb_utils.py @@ -0,0 +1,306 @@ +import json +import sys +from pathlib import Path + +import torch +import yaml +from tqdm import tqdm + +sys.path.append(str(Path(__file__).parent.parent.parent)) # add utils/ to path +from utils.datasets import LoadImagesAndLabels +from utils.datasets import img2label_paths +from utils.general import colorstr, xywh2xyxy, check_dataset + +try: + import wandb + from wandb import init, finish +except ImportError: + wandb = None + +WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' + + +def remove_prefix(from_string, prefix=WANDB_ARTIFACT_PREFIX): + return from_string[len(prefix):] + + +def check_wandb_config_file(data_config_file): + wandb_config = '_wandb.'.join(data_config_file.rsplit('.', 1)) # updated data.yaml path + if Path(wandb_config).is_file(): + return wandb_config + return data_config_file + + +def get_run_info(run_path): + run_path = Path(remove_prefix(run_path, WANDB_ARTIFACT_PREFIX)) + run_id = run_path.stem + project = run_path.parent.stem + model_artifact_name = 'run_' + run_id + '_model' + return run_id, project, model_artifact_name + + +def check_wandb_resume(opt): + process_wandb_config_ddp_mode(opt) if opt.global_rank not in [-1, 0] else None + if isinstance(opt.resume, str): + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + if opt.global_rank not in [-1, 0]: # For resuming DDP runs + run_id, project, model_artifact_name = get_run_info(opt.resume) + api = wandb.Api() + artifact = api.artifact(project + '/' + model_artifact_name + ':latest') + modeldir = artifact.download() + opt.weights = str(Path(modeldir) / "last.pt") + return True + return None + + +def process_wandb_config_ddp_mode(opt): + with open(opt.data) as f: + data_dict = yaml.load(f, Loader=yaml.SafeLoader) # data dict + train_dir, val_dir = None, None + if isinstance(data_dict['train'], str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX): + api = wandb.Api() + train_artifact = api.artifact(remove_prefix(data_dict['train']) + ':' + opt.artifact_alias) + train_dir = train_artifact.download() + train_path = Path(train_dir) / 'data/images/' + data_dict['train'] = str(train_path) + + if isinstance(data_dict['val'], str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX): + api = wandb.Api() + val_artifact = api.artifact(remove_prefix(data_dict['val']) + ':' + opt.artifact_alias) + val_dir = val_artifact.download() + val_path = Path(val_dir) / 'data/images/' + data_dict['val'] = str(val_path) + if train_dir or val_dir: + ddp_data_path = str(Path(val_dir) / 'wandb_local_data.yaml') + with open(ddp_data_path, 'w') as f: + yaml.dump(data_dict, f) + opt.data = ddp_data_path + + +class WandbLogger(): + def __init__(self, opt, name, run_id, data_dict, job_type='Training'): + # Pre-training routine -- + self.job_type = job_type + self.wandb, self.wandb_run, self.data_dict = wandb, None if not wandb else wandb.run, data_dict + # It's more elegant to stick to 1 wandb.init call, but useful config data is overwritten in the WandbLogger's wandb.init call + if isinstance(opt.resume, str): # checks resume from artifact + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + run_id, project, model_artifact_name = get_run_info(opt.resume) + model_artifact_name = WANDB_ARTIFACT_PREFIX + model_artifact_name + assert wandb, 'install wandb to resume wandb runs' + # Resume wandb-artifact:// runs here| workaround for not overwriting wandb.config + self.wandb_run = wandb.init(id=run_id, project=project, resume='allow') + opt.resume = model_artifact_name + elif self.wandb: + self.wandb_run = wandb.init(config=opt, + resume="allow", + project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem, + name=name, + job_type=job_type, + id=run_id) if not wandb.run else wandb.run + if self.wandb_run: + if self.job_type == 'Training': + if not opt.resume: + wandb_data_dict = self.check_and_upload_dataset(opt) if opt.upload_dataset else data_dict + # Info useful for resuming from artifacts + self.wandb_run.config.opt = vars(opt) + self.wandb_run.config.data_dict = wandb_data_dict + self.data_dict = self.setup_training(opt, data_dict) + if self.job_type == 'Dataset Creation': + self.data_dict = self.check_and_upload_dataset(opt) + else: + prefix = colorstr('wandb: ') + print(f"{prefix}Install Weights & Biases for YOLOv5 logging with 'pip install wandb' (recommended)") + + def check_and_upload_dataset(self, opt): + assert wandb, 'Install wandb to upload dataset' + check_dataset(self.data_dict) + config_path = self.log_dataset_artifact(opt.data, + opt.single_cls, + 'YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem) + print("Created dataset config file ", config_path) + with open(config_path) as f: + wandb_data_dict = yaml.load(f, Loader=yaml.SafeLoader) + return wandb_data_dict + + def setup_training(self, opt, data_dict): + self.log_dict, self.current_epoch, self.log_imgs = {}, 0, 16 # Logging Constants + self.bbox_interval = opt.bbox_interval + if isinstance(opt.resume, str): + modeldir, _ = self.download_model_artifact(opt) + if modeldir: + self.weights = Path(modeldir) / "last.pt" + config = self.wandb_run.config + opt.weights, opt.save_period, opt.batch_size, opt.bbox_interval, opt.epochs, opt.hyp = str( + self.weights), config.save_period, config.total_batch_size, config.bbox_interval, config.epochs, \ + config.opt['hyp'] + data_dict = dict(self.wandb_run.config.data_dict) # eliminates the need for config file to resume + if 'val_artifact' not in self.__dict__: # If --upload_dataset is set, use the existing artifact, don't download + self.train_artifact_path, self.train_artifact = self.download_dataset_artifact(data_dict.get('train'), + opt.artifact_alias) + self.val_artifact_path, self.val_artifact = self.download_dataset_artifact(data_dict.get('val'), + opt.artifact_alias) + self.result_artifact, self.result_table, self.val_table, self.weights = None, None, None, None + if self.train_artifact_path is not None: + train_path = Path(self.train_artifact_path) / 'data/images/' + data_dict['train'] = str(train_path) + if self.val_artifact_path is not None: + val_path = Path(self.val_artifact_path) / 'data/images/' + data_dict['val'] = str(val_path) + self.val_table = self.val_artifact.get("val") + self.map_val_table_path() + if self.val_artifact is not None: + self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation") + self.result_table = wandb.Table(["epoch", "id", "prediction", "avg_confidence"]) + if opt.bbox_interval == -1: + self.bbox_interval = opt.bbox_interval = (opt.epochs // 10) if opt.epochs > 10 else 1 + return data_dict + + def download_dataset_artifact(self, path, alias): + if isinstance(path, str) and path.startswith(WANDB_ARTIFACT_PREFIX): + dataset_artifact = wandb.use_artifact(remove_prefix(path, WANDB_ARTIFACT_PREFIX) + ":" + alias) + assert dataset_artifact is not None, "'Error: W&B dataset artifact doesn\'t exist'" + datadir = dataset_artifact.download() + return datadir, dataset_artifact + return None, None + + def download_model_artifact(self, opt): + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + model_artifact = wandb.use_artifact(remove_prefix(opt.resume, WANDB_ARTIFACT_PREFIX) + ":latest") + assert model_artifact is not None, 'Error: W&B model artifact doesn\'t exist' + modeldir = model_artifact.download() + epochs_trained = model_artifact.metadata.get('epochs_trained') + total_epochs = model_artifact.metadata.get('total_epochs') + assert epochs_trained < total_epochs, 'training to %g epochs is finished, nothing to resume.' % ( + total_epochs) + return modeldir, model_artifact + return None, None + + def log_model(self, path, opt, epoch, fitness_score, best_model=False): + model_artifact = wandb.Artifact('run_' + wandb.run.id + '_model', type='model', metadata={ + 'original_url': str(path), + 'epochs_trained': epoch + 1, + 'save period': opt.save_period, + 'project': opt.project, + 'total_epochs': opt.epochs, + 'fitness_score': fitness_score + }) + model_artifact.add_file(str(path / 'last.pt'), name='last.pt') + wandb.log_artifact(model_artifact, + aliases=['latest', 'epoch ' + str(self.current_epoch), 'best' if best_model else '']) + print("Saving model artifact on epoch ", epoch + 1) + + def log_dataset_artifact(self, data_file, single_cls, project, overwrite_config=False): + with open(data_file) as f: + data = yaml.load(f, Loader=yaml.SafeLoader) # data dict + nc, names = (1, ['item']) if single_cls else (int(data['nc']), data['names']) + names = {k: v for k, v in enumerate(names)} # to index dictionary + self.train_artifact = self.create_dataset_table(LoadImagesAndLabels( + data['train']), names, name='train') if data.get('train') else None + self.val_artifact = self.create_dataset_table(LoadImagesAndLabels( + data['val']), names, name='val') if data.get('val') else None + if data.get('train'): + data['train'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'train') + if data.get('val'): + data['val'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'val') + path = data_file if overwrite_config else '_wandb.'.join(data_file.rsplit('.', 1)) # updated data.yaml path + data.pop('download', None) + with open(path, 'w') as f: + yaml.dump(data, f) + + if self.job_type == 'Training': # builds correct artifact pipeline graph + self.wandb_run.use_artifact(self.val_artifact) + self.wandb_run.use_artifact(self.train_artifact) + self.val_artifact.wait() + self.val_table = self.val_artifact.get('val') + self.map_val_table_path() + else: + self.wandb_run.log_artifact(self.train_artifact) + self.wandb_run.log_artifact(self.val_artifact) + return path + + def map_val_table_path(self): + self.val_table_map = {} + print("Mapping dataset") + for i, data in enumerate(tqdm(self.val_table.data)): + self.val_table_map[data[3]] = data[0] + + def create_dataset_table(self, dataset, class_to_id, name='dataset'): + # TODO: Explore multiprocessing to slpit this loop parallely| This is essential for speeding up the the logging + artifact = wandb.Artifact(name=name, type="dataset") + img_files = tqdm([dataset.path]) if isinstance(dataset.path, str) and Path(dataset.path).is_dir() else None + img_files = tqdm(dataset.img_files) if not img_files else img_files + for img_file in img_files: + if Path(img_file).is_dir(): + artifact.add_dir(img_file, name='data/images') + labels_path = 'labels'.join(dataset.path.rsplit('images', 1)) + artifact.add_dir(labels_path, name='data/labels') + else: + artifact.add_file(img_file, name='data/images/' + Path(img_file).name) + label_file = Path(img2label_paths([img_file])[0]) + artifact.add_file(str(label_file), + name='data/labels/' + label_file.name) if label_file.exists() else None + table = wandb.Table(columns=["id", "train_image", "Classes", "name"]) + class_set = wandb.Classes([{'id': id, 'name': name} for id, name in class_to_id.items()]) + for si, (img, labels, paths, shapes) in enumerate(tqdm(dataset)): + height, width = shapes[0] + labels[:, 2:] = (xywh2xyxy(labels[:, 2:].view(-1, 4))) * torch.Tensor([width, height, width, height]) + box_data, img_classes = [], {} + for cls, *xyxy in labels[:, 1:].tolist(): + cls = int(cls) + box_data.append({"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]}, + "class_id": cls, + "box_caption": "%s" % (class_to_id[cls]), + "scores": {"acc": 1}, + "domain": "pixel"}) + img_classes[cls] = class_to_id[cls] + boxes = {"ground_truth": {"box_data": box_data, "class_labels": class_to_id}} # inference-space + table.add_data(si, wandb.Image(paths, classes=class_set, boxes=boxes), json.dumps(img_classes), + Path(paths).name) + artifact.add(table, name) + return artifact + + def log_training_progress(self, predn, path, names): + if self.val_table and self.result_table: + class_set = wandb.Classes([{'id': id, 'name': name} for id, name in names.items()]) + box_data = [] + total_conf = 0 + for *xyxy, conf, cls in predn.tolist(): + if conf >= 0.25: + box_data.append( + {"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]}, + "class_id": int(cls), + "box_caption": "%s %.3f" % (names[cls], conf), + "scores": {"class_score": conf}, + "domain": "pixel"}) + total_conf = total_conf + conf + boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space + id = self.val_table_map[Path(path).name] + self.result_table.add_data(self.current_epoch, + id, + wandb.Image(self.val_table.data[id][1], boxes=boxes, classes=class_set), + total_conf / max(1, len(box_data)) + ) + + def log(self, log_dict): + if self.wandb_run: + for key, value in log_dict.items(): + self.log_dict[key] = value + + def end_epoch(self, best_result=False): + if self.wandb_run: + wandb.log(self.log_dict) + self.log_dict = {} + if self.result_artifact: + train_results = wandb.JoinedTable(self.val_table, self.result_table, "id") + self.result_artifact.add(train_results, 'result') + wandb.log_artifact(self.result_artifact, aliases=['latest', 'epoch ' + str(self.current_epoch), + ('best' if best_result else '')]) + self.result_table = wandb.Table(["epoch", "id", "prediction", "avg_confidence"]) + self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation") + + def finish_run(self): + if self.wandb_run: + if self.log_dict: + wandb.log(self.log_dict) + wandb.run.finish()