v3.0 Release (#725)

* initial commit

* remove yolov3-spp from test.py study

* update study --img range

* update mAP

* cleanup and speed updates

* update README plot

Dockerfile

 
 # Build and Push
 # t=ultralytics/yolov5:latest && sudo docker build -t $t . && sudo docker push $t
+# for v in {300..303}; do t=ultralytics/coco:v$v && sudo docker build -t $t . && sudo docker push $t; done
 
 # Pull and Run
 # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host $t
 # sudo docker commit 092b16b25c5b usr/resume && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco --entrypoint=sh usr/resume
 
 # Send weights to GCP
-# python -c "from utils.general import *; strip_optimizer('runs/exp0_*/weights/last.pt', 'tmp.pt')" && gsutil cp tmp.pt gs://*
+# python -c "from utils.general import *; strip_optimizer('runs/exp0_*/weights/best.pt', 'tmp.pt')" && gsutil cp tmp.pt gs://*.pt
 
 # Clean up
 # docker system prune -a --volumes

README.md

 
 This repository represents Ultralytics open-source research into future object detection methods, and incorporates our lessons learned and best practices evolved over training thousands of models on custom client datasets with our previous YOLO repository https://github.com/ultralytics/yolov3. **All code and models are under active development, and are subject to modification or deletion without notice.** Use at your own risk.
 
-<img src="https://user-images.githubusercontent.com/26833433/85340570-30360a80-b49b-11ea-87cf-bdf33d53ae15.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 8, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS.
+<img src="https://user-images.githubusercontent.com/26833433/90187293-6773ba00-dd6e-11ea-8f90-cd94afc0427f.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS. EfficientDet data from [google/automl](https://github.com/google/automl) at batch size 8.
 
+- **August 13, 2020**: [v3.0 release](https://github.com/ultralytics/yolov5/releases/tag/v3.0): nn.Hardswish() activations, data autodownload, native AMP.
 - **July 23, 2020**: [v2.0 release](https://github.com/ultralytics/yolov5/releases/tag/v2.0): improved model definition, training and mAP.
 - **June 22, 2020**: [PANet](https://arxiv.org/abs/1803.01534) updates: new heads, reduced parameters, improved speed and mAP [364fcfd](https://github.com/ultralytics/yolov5/commit/364fcfd7dba53f46edd4f04c037a039c0a287972).
 - **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
 
 | Model | AP<sup>val</sup> | AP<sup>test</sup> | AP<sub>50</sub> | Speed<sub>GPU</sub> | FPS<sub>GPU</sub> || params | FLOPS |
 |---------- |------ |------ |------ | -------- | ------| ------ |------  |  :------: |
-| [YOLOv5s](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 36.1     | 36.1     | 55.3     | **2.1ms** | **476** || 7.5M   | 13.2B
-| [YOLOv5m](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 43.5     | 43.5     | 62.5     | 3.0ms     | 333     || 21.8M  | 39.4B
-| [YOLOv5l](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 47.0     | 47.1     | 65.6     | 3.9ms     | 256     || 47.8M  | 88.1B
-| [YOLOv5x](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | **49.0** | **49.0** | **67.4** | 6.1ms     | 164     || 89.0M  | 166.4B
+| [YOLOv5s](https://github.com/ultralytics/yolov5/releases/tag/v3.0)    | 37.0     | 37.0     | 56.2     | **2.4ms** | **476** || 7.5M   | 13.2B
+| [YOLOv5m](https://github.com/ultralytics/yolov5/releases/tag/v3.0)    | 44.3     | 44.3     | 63.2     | 3.4ms     | 333     || 21.8M  | 39.4B
+| [YOLOv5l](https://github.com/ultralytics/yolov5/releases/tag/v3.0)    | 47.7     | 47.7     | 66.5     | 4.4ms     | 256     || 47.8M  | 88.1B
+| [YOLOv5x](https://github.com/ultralytics/yolov5/releases/tag/v3.0)    | **49.2** | **49.2** | **67.7** | 6.9ms     | 164     || 89.0M  | 166.4B
 | | | | | | || |
-| [YOLOv3-SPP](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J) | 45.6     | 45.5     | 65.2     | 4.5ms     | 222     || 63.0M  | 118.0B
-
+| [YOLOv5x](https://github.com/ultralytics/yolov5/releases/tag/v3.0) + TTA|**50.8**| **50.8** | **68.9** | 25.5ms    | 39      || 89.0M  | 354.3B
+| | | | | | || |
+| [YOLOv3-SPP](https://github.com/ultralytics/yolov5/releases/tag/v3.0) | 45.6     | 45.5     | 65.2     | 4.5ms     | 222     || 63.0M  | 118.0B
 
 ** AP<sup>test</sup> denotes COCO [test-dev2017](http://cocodataset.org/#upload) server results, all other AP results in the table denote val2017 accuracy.  
-** All AP numbers are for single-model single-scale without ensemble or test-time augmentation. Reproduce by `python test.py --data coco.yaml --img 672 --conf 0.001`  
-** Speed<sub>GPU</sub> measures end-to-end time per image averaged over 5000 COCO val2017 images using a GCP [n1-standard-16](https://cloud.google.com/compute/docs/machine-types#n1_standard_machine_types) instance with one V100 GPU, and includes image preprocessing, PyTorch FP16 image inference at --batch-size 32 --img-size 640, postprocessing and NMS. Average NMS time included in this chart is 1-2ms/img.  Reproduce by `python test.py --data coco.yaml --img 640 --conf 0.1`  
+** All AP numbers are for single-model single-scale without ensemble or test-time augmentation. **Reproduce** by `python test.py --data coco.yaml --img 640 --conf 0.001`  
+** Speed<sub>GPU</sub> measures end-to-end time per image averaged over 5000 COCO val2017 images using a GCP [n1-standard-16](https://cloud.google.com/compute/docs/machine-types#n1_standard_machine_types) instance with one V100 GPU, and includes image preprocessing, PyTorch FP16 image inference at --batch-size 32 --img-size 640, postprocessing and NMS. Average NMS time included in this chart is 1-2ms/img.  **Reproduce** by `python test.py --data coco.yaml --img 640 --conf 0.1`  
 ** All checkpoints are trained to 300 epochs with default settings and hyperparameters (no autoaugmentation). 
-
+** Test Time Augmentation ([TTA](https://github.com/ultralytics/yolov5/issues/303)) runs at 3 image sizes. **Reproduce** by `python test.py --data coco.yaml --img 832 --augment` 
 
 ## Requirements
 
 
 ## Training
 
-Download [COCO](https://github.com/ultralytics/yolov5/blob/master/data/get_coco2017.sh) and run command below. Training times for YOLOv5s/m/l/x are 2/4/6/8 days on a single V100 (multi-GPU times faster). Use the largest `--batch-size` your GPU allows (batch sizes shown for 16 GB devices).
+Download [COCO](https://github.com/ultralytics/yolov5/blob/master/data/scripts/get_coco.sh) and run command below. Training times for YOLOv5s/m/l/x are 2/4/6/8 days on a single V100 (multi-GPU times faster). Use the largest `--batch-size` your GPU allows (batch sizes shown for 16 GB devices).
 ```bash
 $ python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size 64
-                                         yolov5m                                48
-                                         yolov5l                                32
+                                         yolov5m                                40
+                                         yolov5l                                24
                                          yolov5x                                16
 ```
 <img src="https://user-images.githubusercontent.com/26833433/84186698-c4d54d00-aa45-11ea-9bde-c632c1230ccd.png" width="900">

data/hyp.finetune.yaml

 hsv_s: 0.7  # image HSV-Saturation augmentation (fraction)
 hsv_v: 0.4  # image HSV-Value augmentation (fraction)
 degrees: 0.0  # image rotation (+/- deg)
-translate: 0.5  # image translation (+/- fraction)
+translate: 0.1  # image translation (+/- fraction)
 scale: 0.5  # image scale (+/- gain)
 shear: 0.0  # image shear (+/- deg)
 perspective: 0.0  # image perspective (+/- fraction), range 0-0.001

data/hyp.scratch.yaml

 hsv_s: 0.7  # image HSV-Saturation augmentation (fraction)
 hsv_v: 0.4  # image HSV-Value augmentation (fraction)
 degrees: 0.0  # image rotation (+/- deg)
-translate: 0.5  # image translation (+/- fraction)
+translate: 0.1  # image translation (+/- fraction)
 scale: 0.5  # image scale (+/- gain)
 shear: 0.0  # image shear (+/- deg)
 perspective: 0.0  # image perspective (+/- fraction), range 0-0.001

models/common.py

         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.LeakyReLU(0.1, inplace=True) if act else nn.Identity()
+        self.act = nn.Hardswish() if act else nn.Identity()
 
     def forward(self, x):
         return self.act(self.bn(self.conv(x)))

models/yolo.py

 
 logger = logging.getLogger(__name__)
 
+
 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.stride = None  # strides computed during build
         self.nc = nc  # number of classes
         self.no = nc + 5  # number of outputs per anchor
         self.nl = len(anchors)  # number of detection layers
         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
-        self.export = False  # onnx export
 
     def forward(self, x):
         # x = x.copy()  # for profiling

test.py

              opt.verbose)
 
     elif opt.task == 'study':  # run over a range of settings and save/plot
-        for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt', 'yolov3-spp.pt']:
+        for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']:
             f = 'study_%s_%s.txt' % (Path(opt.data).stem, Path(weights).stem)  # filename to save to
-            x = list(range(352, 832, 64))  # x axis
+            x = list(range(320, 800, 64))  # x axis
             y = []  # y axis
             for i in x:  # img-size
                 print('\nRunning %s point %s...' % (f, i))
                 y.append(r + t)  # results and times
             np.savetxt(f, y, fmt='%10.4g')  # save
         os.system('zip -r study.zip study_*.txt')
-        # plot_study_txt(f, x)  # plot
+        # utils.general.plot_study_txt(f, x)  # plot

train.py

 
 logger = logging.getLogger(__name__)
 
+
 def train(hyp, opt, device, tb_writer=None):
     logger.info(f'Hyperparameters {hyp}')
     log_dir = Path(tb_writer.log_dir) if tb_writer else Path(opt.logdir) / 'evolve'  # logging directory
         start_epoch = ckpt['epoch'] + 1
         if epochs < start_epoch:
             logger.info('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' %
-                  (weights, ckpt['epoch'], epochs))
+                        (weights, ckpt['epoch'], epochs))
             epochs += ckpt['epoch']  # finetune additional epochs
 
         del ckpt, state_dict
     parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
     parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify')
     parser.add_argument('--logdir', type=str, default='runs/', help='logging directory')
-    parser.add_argument('--workers', type=int, default=8, help='maximum number of workers for dataloader')
+    parser.add_argument('--workers', type=int, default=8, help='maximum number of dataloader workers')
     opt = parser.parse_args()
 
     # Set DDP variables
         if last and not opt.weights:
             logger.info(f'Resuming training from {last}')
         opt.weights = last if opt.resume and not opt.weights else opt.weights
-    if opt.global_rank in [-1,0]:
+    if opt.global_rank in [-1, 0]:
         check_git_status()
 
     opt.hyp = opt.hyp or ('data/hyp.finetune.yaml' if opt.weights else 'data/hyp.scratch.yaml')

utils/activations.py

         return x * torch.sigmoid(x)
 
 
-class HardSwish(nn.Module):
-    @staticmethod
-    def forward(x):
-        return x * F.hardtanh(x + 3, 0., 6., True) / 6.
-
-
 class MemoryEfficientSwish(nn.Module):
     class F(torch.autograd.Function):
         @staticmethod

utils/datasets.py

 
     labels4 = []
     s = self.img_size
-    yc, xc = s, s  # mosaic center x, y
+    yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border]  # mosaic center x, y
     indices = [index] + [random.randint(0, len(self.labels) - 1) for _ in range(3)]  # 3 additional image indices
     for i, index in enumerate(indices):
         # Load image
     return img, targets
 
 
-def box_candidates(box1, box2, wh_thr=2, ar_thr=20, area_thr=0.2):  # box1(4,n), box2(4,n)
+def box_candidates(box1, box2, wh_thr=2, ar_thr=20, area_thr=0.1):  # 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]

utils/general.py

     ax = ax.ravel()
 
     fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True)
-    for f in ['coco_study/study_coco_yolov5%s.txt' % x for x in ['s', 'm', 'l', 'x']]:
+    for f in ['study/study_coco_yolov5%s.txt' % x for x in ['s', 'm', 'l', 'x']]:
         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)']
         ax2.plot(y[6, :j], y[3, :j] * 1E2, '.-', linewidth=2, markersize=8,
                  label=Path(f).stem.replace('study_coco_', '').replace('yolo', 'YOLO'))
 
-    ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [33.8, 39.6, 43.0, 47.5, 49.4, 50.7],
+    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()
     ax2.set_ylabel('COCO AP val')
     ax2.legend(loc='lower right')
     plt.savefig('study_mAP_latency.png', dpi=300)
-    plt.savefig(f.replace('.txt', '.png'), dpi=200)
+    plt.savefig(f.replace('.txt', '.png'), dpi=300)
 
 
 def plot_labels(labels, save_dir=''):
     s = ['GIoU', 'Objectness', 'Classification', 'Precision', 'Recall',
          'val GIoU', 'val Objectness', 'val Classification', 'mAP@0.5', 'mAP@0.5:0.95']
     if bucket:
-        os.system('rm -rf storage.googleapis.com')
-        files = ['https://storage.googleapis.com/%s/results%g.txt' % (bucket, x) for x in id]
+        # os.system('rm -rf storage.googleapis.com')
+        # 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 = glob.glob(str(Path(save_dir) / 'results*.txt')) + glob.glob('../../Downloads/results*.txt')
     for fi, f in enumerate(files):
                 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:
-            print('Warning: Plotting error for %s, skipping file' % f)
+        except Exception as e:
+            print('Warning: Plotting error for %s; %s' % (f, e))
 
     fig.tight_layout()
     ax[1].legend()