Nelze vybrat více než 25 témat Téma musí začínat písmenem nebo číslem, může obsahovat pomlčky („-“) a může být dlouhé až 35 znaků.

174 lines
7.8KB

  1. """Export a YOLOv5 *.pt model to TorchScript, ONNX, CoreML formats
  2. Usage:
  3. $ python path/to/export.py --weights yolov5s.pt --img 640 --batch 1
  4. """
  5. import argparse
  6. import sys
  7. import time
  8. from pathlib import Path
  9. import torch
  10. import torch.nn as nn
  11. from torch.utils.mobile_optimizer import optimize_for_mobile
  12. FILE = Path(__file__).absolute()
  13. sys.path.append(FILE.parents[0].as_posix()) # add yolov5/ to path
  14. from models.common import Conv
  15. from models.yolo import Detect
  16. from models.experimental import attempt_load
  17. from utils.activations import Hardswish, SiLU
  18. from utils.general import colorstr, check_img_size, check_requirements, file_size, set_logging
  19. from utils.torch_utils import select_device
  20. def run(weights='./yolov5s.pt', # weights path
  21. img_size=(640, 640), # image (height, width)
  22. batch_size=1, # batch size
  23. device='cpu', # cuda device, i.e. 0 or 0,1,2,3 or cpu
  24. include=('torchscript', 'onnx', 'coreml'), # include formats
  25. half=False, # FP16 half-precision export
  26. inplace=False, # set YOLOv5 Detect() inplace=True
  27. train=False, # model.train() mode
  28. optimize=False, # TorchScript: optimize for mobile
  29. dynamic=False, # ONNX: dynamic axes
  30. simplify=False, # ONNX: simplify model
  31. opset_version=12, # ONNX: opset version
  32. ):
  33. t = time.time()
  34. include = [x.lower() for x in include]
  35. img_size *= 2 if len(img_size) == 1 else 1 # expand
  36. # Load PyTorch model
  37. device = select_device(device)
  38. assert not (device.type == 'cpu' and half), '--half only compatible with GPU export, i.e. use --device 0'
  39. model = attempt_load(weights, map_location=device) # load FP32 model
  40. labels = model.names
  41. # Input
  42. gs = int(max(model.stride)) # grid size (max stride)
  43. img_size = [check_img_size(x, gs) for x in img_size] # verify img_size are gs-multiples
  44. img = torch.zeros(batch_size, 3, *img_size).to(device) # image size(1,3,320,192) iDetection
  45. # Update model
  46. if half:
  47. img, model = img.half(), model.half() # to FP16
  48. model.train() if train else model.eval() # training mode = no Detect() layer grid construction
  49. for k, m in model.named_modules():
  50. m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility
  51. if isinstance(m, Conv): # assign export-friendly activations
  52. if isinstance(m.act, nn.Hardswish):
  53. m.act = Hardswish()
  54. elif isinstance(m.act, nn.SiLU):
  55. m.act = SiLU()
  56. elif isinstance(m, Detect):
  57. m.inplace = inplace
  58. m.onnx_dynamic = dynamic
  59. # m.forward = m.forward_export # assign forward (optional)
  60. for _ in range(2):
  61. y = model(img) # dry runs
  62. print(f"\n{colorstr('PyTorch:')} starting from {weights} ({file_size(weights):.1f} MB)")
  63. # TorchScript export -----------------------------------------------------------------------------------------------
  64. if 'torchscript' in include or 'coreml' in include:
  65. prefix = colorstr('TorchScript:')
  66. try:
  67. print(f'\n{prefix} starting export with torch {torch.__version__}...')
  68. f = weights.replace('.pt', '.torchscript.pt') # filename
  69. ts = torch.jit.trace(model, img, strict=False)
  70. (optimize_for_mobile(ts) if optimize else ts).save(f)
  71. print(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)')
  72. except Exception as e:
  73. print(f'{prefix} export failure: {e}')
  74. # ONNX export ------------------------------------------------------------------------------------------------------
  75. if 'onnx' in include:
  76. prefix = colorstr('ONNX:')
  77. try:
  78. import onnx
  79. print(f'{prefix} starting export with onnx {onnx.__version__}...')
  80. f = weights.replace('.pt', '.onnx') # filename
  81. torch.onnx.export(model, img, f, verbose=False, opset_version=opset_version,
  82. training=torch.onnx.TrainingMode.TRAINING if train else torch.onnx.TrainingMode.EVAL,
  83. do_constant_folding=not train,
  84. input_names=['images'],
  85. output_names=['output'],
  86. dynamic_axes={'images': {0: 'batch', 2: 'height', 3: 'width'}, # shape(1,3,640,640)
  87. 'output': {0: 'batch', 1: 'anchors'} # shape(1,25200,85)
  88. } if dynamic else None)
  89. # Checks
  90. model_onnx = onnx.load(f) # load onnx model
  91. onnx.checker.check_model(model_onnx) # check onnx model
  92. # print(onnx.helper.printable_graph(model_onnx.graph)) # print
  93. # Simplify
  94. if simplify:
  95. try:
  96. check_requirements(['onnx-simplifier'])
  97. import onnxsim
  98. print(f'{prefix} simplifying with onnx-simplifier {onnxsim.__version__}...')
  99. model_onnx, check = onnxsim.simplify(
  100. model_onnx,
  101. dynamic_input_shape=dynamic,
  102. input_shapes={'images': list(img.shape)} if dynamic else None)
  103. assert check, 'assert check failed'
  104. onnx.save(model_onnx, f)
  105. except Exception as e:
  106. print(f'{prefix} simplifier failure: {e}')
  107. print(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)')
  108. except Exception as e:
  109. print(f'{prefix} export failure: {e}')
  110. # CoreML export ----------------------------------------------------------------------------------------------------
  111. if 'coreml' in include:
  112. prefix = colorstr('CoreML:')
  113. try:
  114. import coremltools as ct
  115. print(f'{prefix} starting export with coremltools {ct.__version__}...')
  116. assert train, 'CoreML exports should be placed in model.train() mode with `python export.py --train`'
  117. model = ct.convert(ts, inputs=[ct.ImageType('image', shape=img.shape, scale=1 / 255.0, bias=[0, 0, 0])])
  118. f = weights.replace('.pt', '.mlmodel') # filename
  119. model.save(f)
  120. print(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)')
  121. except Exception as e:
  122. print(f'{prefix} export failure: {e}')
  123. # Finish
  124. print(f'\nExport complete ({time.time() - t:.2f}s). Visualize with https://github.com/lutzroeder/netron.')
  125. def parse_opt():
  126. parser = argparse.ArgumentParser()
  127. parser.add_argument('--weights', type=str, default='./yolov5s.pt', help='weights path')
  128. parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image (height, width)')
  129. parser.add_argument('--batch-size', type=int, default=1, help='batch size')
  130. parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
  131. parser.add_argument('--include', nargs='+', default=['torchscript', 'onnx', 'coreml'], help='include formats')
  132. parser.add_argument('--half', action='store_true', help='FP16 half-precision export')
  133. parser.add_argument('--inplace', action='store_true', help='set YOLOv5 Detect() inplace=True')
  134. parser.add_argument('--train', action='store_true', help='model.train() mode')
  135. parser.add_argument('--optimize', action='store_true', help='TorchScript: optimize for mobile')
  136. parser.add_argument('--dynamic', action='store_true', help='ONNX: dynamic axes')
  137. parser.add_argument('--simplify', action='store_true', help='ONNX: simplify model')
  138. parser.add_argument('--opset-version', type=int, default=12, help='ONNX: opset version')
  139. opt = parser.parse_args()
  140. return opt
  141. def main(opt):
  142. set_logging()
  143. print(colorstr('export: ') + ', '.join(f'{k}={v}' for k, v in vars(opt).items()))
  144. run(**vars(opt))
  145. if __name__ == "__main__":
  146. opt = parse_opt()
  147. main(opt)