mobilenetv2.py 9.5 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260
  1. from functools import partial
  2. from typing import Any, Callable, List, Optional
  3. import torch
  4. from torch import nn, Tensor
  5. from ..ops.misc import Conv2dNormActivation
  6. from ..transforms._presets import ImageClassification
  7. from ..utils import _log_api_usage_once
  8. from ._api import register_model, Weights, WeightsEnum
  9. from ._meta import _IMAGENET_CATEGORIES
  10. from ._utils import _make_divisible, _ovewrite_named_param, handle_legacy_interface
  11. __all__ = ["MobileNetV2", "MobileNet_V2_Weights", "mobilenet_v2"]
  12. # necessary for backwards compatibility
  13. class InvertedResidual(nn.Module):
  14. def __init__(
  15. self, inp: int, oup: int, stride: int, expand_ratio: int, norm_layer: Optional[Callable[..., nn.Module]] = None
  16. ) -> None:
  17. super().__init__()
  18. self.stride = stride
  19. if stride not in [1, 2]:
  20. raise ValueError(f"stride should be 1 or 2 instead of {stride}")
  21. if norm_layer is None:
  22. norm_layer = nn.BatchNorm2d
  23. hidden_dim = int(round(inp * expand_ratio))
  24. self.use_res_connect = self.stride == 1 and inp == oup
  25. layers: List[nn.Module] = []
  26. if expand_ratio != 1:
  27. # pw
  28. layers.append(
  29. Conv2dNormActivation(inp, hidden_dim, kernel_size=1, norm_layer=norm_layer, activation_layer=nn.ReLU6)
  30. )
  31. layers.extend(
  32. [
  33. # dw
  34. Conv2dNormActivation(
  35. hidden_dim,
  36. hidden_dim,
  37. stride=stride,
  38. groups=hidden_dim,
  39. norm_layer=norm_layer,
  40. activation_layer=nn.ReLU6,
  41. ),
  42. # pw-linear
  43. nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
  44. norm_layer(oup),
  45. ]
  46. )
  47. self.conv = nn.Sequential(*layers)
  48. self.out_channels = oup
  49. self._is_cn = stride > 1
  50. def forward(self, x: Tensor) -> Tensor:
  51. if self.use_res_connect:
  52. return x + self.conv(x)
  53. else:
  54. return self.conv(x)
  55. class MobileNetV2(nn.Module):
  56. def __init__(
  57. self,
  58. num_classes: int = 1000,
  59. width_mult: float = 1.0,
  60. inverted_residual_setting: Optional[List[List[int]]] = None,
  61. round_nearest: int = 8,
  62. block: Optional[Callable[..., nn.Module]] = None,
  63. norm_layer: Optional[Callable[..., nn.Module]] = None,
  64. dropout: float = 0.2,
  65. ) -> None:
  66. """
  67. MobileNet V2 main class
  68. Args:
  69. num_classes (int): Number of classes
  70. width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
  71. inverted_residual_setting: Network structure
  72. round_nearest (int): Round the number of channels in each layer to be a multiple of this number
  73. Set to 1 to turn off rounding
  74. block: Module specifying inverted residual building block for mobilenet
  75. norm_layer: Module specifying the normalization layer to use
  76. dropout (float): The droupout probability
  77. """
  78. super().__init__()
  79. _log_api_usage_once(self)
  80. if block is None:
  81. block = InvertedResidual
  82. if norm_layer is None:
  83. norm_layer = nn.BatchNorm2d
  84. input_channel = 32
  85. last_channel = 1280
  86. if inverted_residual_setting is None:
  87. inverted_residual_setting = [
  88. # t, c, n, s
  89. [1, 16, 1, 1],
  90. [6, 24, 2, 2],
  91. [6, 32, 3, 2],
  92. [6, 64, 4, 2],
  93. [6, 96, 3, 1],
  94. [6, 160, 3, 2],
  95. [6, 320, 1, 1],
  96. ]
  97. # only check the first element, assuming user knows t,c,n,s are required
  98. if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
  99. raise ValueError(
  100. f"inverted_residual_setting should be non-empty or a 4-element list, got {inverted_residual_setting}"
  101. )
  102. # building first layer
  103. input_channel = _make_divisible(input_channel * width_mult, round_nearest)
  104. self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
  105. features: List[nn.Module] = [
  106. Conv2dNormActivation(3, input_channel, stride=2, norm_layer=norm_layer, activation_layer=nn.ReLU6)
  107. ]
  108. # building inverted residual blocks
  109. for t, c, n, s in inverted_residual_setting:
  110. output_channel = _make_divisible(c * width_mult, round_nearest)
  111. for i in range(n):
  112. stride = s if i == 0 else 1
  113. features.append(block(input_channel, output_channel, stride, expand_ratio=t, norm_layer=norm_layer))
  114. input_channel = output_channel
  115. # building last several layers
  116. features.append(
  117. Conv2dNormActivation(
  118. input_channel, self.last_channel, kernel_size=1, norm_layer=norm_layer, activation_layer=nn.ReLU6
  119. )
  120. )
  121. # make it nn.Sequential
  122. self.features = nn.Sequential(*features)
  123. # building classifier
  124. self.classifier = nn.Sequential(
  125. nn.Dropout(p=dropout),
  126. nn.Linear(self.last_channel, num_classes),
  127. )
  128. # weight initialization
  129. for m in self.modules():
  130. if isinstance(m, nn.Conv2d):
  131. nn.init.kaiming_normal_(m.weight, mode="fan_out")
  132. if m.bias is not None:
  133. nn.init.zeros_(m.bias)
  134. elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
  135. nn.init.ones_(m.weight)
  136. nn.init.zeros_(m.bias)
  137. elif isinstance(m, nn.Linear):
  138. nn.init.normal_(m.weight, 0, 0.01)
  139. nn.init.zeros_(m.bias)
  140. def _forward_impl(self, x: Tensor) -> Tensor:
  141. # This exists since TorchScript doesn't support inheritance, so the superclass method
  142. # (this one) needs to have a name other than `forward` that can be accessed in a subclass
  143. x = self.features(x)
  144. # Cannot use "squeeze" as batch-size can be 1
  145. x = nn.functional.adaptive_avg_pool2d(x, (1, 1))
  146. x = torch.flatten(x, 1)
  147. x = self.classifier(x)
  148. return x
  149. def forward(self, x: Tensor) -> Tensor:
  150. return self._forward_impl(x)
  151. _COMMON_META = {
  152. "num_params": 3504872,
  153. "min_size": (1, 1),
  154. "categories": _IMAGENET_CATEGORIES,
  155. }
  156. class MobileNet_V2_Weights(WeightsEnum):
  157. IMAGENET1K_V1 = Weights(
  158. url="https://download.pytorch.org/models/mobilenet_v2-b0353104.pth",
  159. transforms=partial(ImageClassification, crop_size=224),
  160. meta={
  161. **_COMMON_META,
  162. "recipe": "https://github.com/pytorch/vision/tree/main/references/classification#mobilenetv2",
  163. "_metrics": {
  164. "ImageNet-1K": {
  165. "acc@1": 71.878,
  166. "acc@5": 90.286,
  167. }
  168. },
  169. "_ops": 0.301,
  170. "_file_size": 13.555,
  171. "_docs": """These weights reproduce closely the results of the paper using a simple training recipe.""",
  172. },
  173. )
  174. IMAGENET1K_V2 = Weights(
  175. url="https://download.pytorch.org/models/mobilenet_v2-7ebf99e0.pth",
  176. transforms=partial(ImageClassification, crop_size=224, resize_size=232),
  177. meta={
  178. **_COMMON_META,
  179. "recipe": "https://github.com/pytorch/vision/issues/3995#new-recipe-with-reg-tuning",
  180. "_metrics": {
  181. "ImageNet-1K": {
  182. "acc@1": 72.154,
  183. "acc@5": 90.822,
  184. }
  185. },
  186. "_ops": 0.301,
  187. "_file_size": 13.598,
  188. "_docs": """
  189. These weights improve upon the results of the original paper by using a modified version of TorchVision's
  190. `new training recipe
  191. <https://pytorch.org/blog/how-to-train-state-of-the-art-models-using-torchvision-latest-primitives/>`_.
  192. """,
  193. },
  194. )
  195. DEFAULT = IMAGENET1K_V2
  196. @register_model()
  197. @handle_legacy_interface(weights=("pretrained", MobileNet_V2_Weights.IMAGENET1K_V1))
  198. def mobilenet_v2(
  199. *, weights: Optional[MobileNet_V2_Weights] = None, progress: bool = True, **kwargs: Any
  200. ) -> MobileNetV2:
  201. """MobileNetV2 architecture from the `MobileNetV2: Inverted Residuals and Linear
  202. Bottlenecks <https://arxiv.org/abs/1801.04381>`_ paper.
  203. Args:
  204. weights (:class:`~torchvision.models.MobileNet_V2_Weights`, optional): The
  205. pretrained weights to use. See
  206. :class:`~torchvision.models.MobileNet_V2_Weights` below for
  207. more details, and possible values. By default, no pre-trained
  208. weights are used.
  209. progress (bool, optional): If True, displays a progress bar of the
  210. download to stderr. Default is True.
  211. **kwargs: parameters passed to the ``torchvision.models.mobilenetv2.MobileNetV2``
  212. base class. Please refer to the `source code
  213. <https://github.com/pytorch/vision/blob/main/torchvision/models/mobilenetv2.py>`_
  214. for more details about this class.
  215. .. autoclass:: torchvision.models.MobileNet_V2_Weights
  216. :members:
  217. """
  218. weights = MobileNet_V2_Weights.verify(weights)
  219. if weights is not None:
  220. _ovewrite_named_param(kwargs, "num_classes", len(weights.meta["categories"]))
  221. model = MobileNetV2(**kwargs)
  222. if weights is not None:
  223. model.load_state_dict(weights.get_state_dict(progress=progress, check_hash=True))
  224. return model