| 72 | return out |
| 73 | |
| 74 | class ShuffleNetV2(nn.Module): |
| 75 | def __init__(self, stages_repeats, stages_out_channels, num_classes=1000, inverted_residual=InvertedResidual): |
| 76 | super(ShuffleNetV2, self).__init__() |
| 77 | |
| 78 | if len(stages_repeats) != 3: |
| 79 | raise ValueError("expected stages_repeats as list of 3 positive ints") |
| 80 | if len(stages_out_channels) != 5: |
| 81 | raise ValueError("expected stages_out_channels as list of 5 positive ints") |
| 82 | self._stage_out_channels = stages_out_channels |
| 83 | |
| 84 | # input RGB image |
| 85 | input_channels = 3 |
| 86 | output_channels = self._stage_out_channels[0] |
| 87 | |
| 88 | self.conv1 = nn.Sequential( |
| 89 | nn.Conv2d(input_channels, output_channels, kernel_size=3, stride=2, padding=1, bias=False), |
| 90 | nn.BatchNorm2d(output_channels), |
| 91 | nn.ReLU(inplace=True) # inplace=True表示直接进行覆盖 |
| 92 | ) |
| 93 | input_channels = output_channels |
| 94 | |
| 95 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) |
| 96 | |
| 97 | self.stage2: nn.Sequential |
| 98 | self.stage3: nn.Sequential |
| 99 | self.stage4: nn.Sequential |
| 100 | |
| 101 | stage_names = ["stage{}".format(i) for i in [2, 3, 4]] |
| 102 | for name, repeats, output_channels in zip(stage_names, stages_repeats, self._stage_out_channels[1:]): |
| 103 | seq = [inverted_residual(input_channels, output_channels, 2)] |
| 104 | for i in range(repeats - 1): |
| 105 | seq.append(inverted_residual(output_channels, output_channels, 1)) |
| 106 | setattr(self, name, nn.Sequential(*seq)) |
| 107 | input_channels = output_channels |
| 108 | |
| 109 | output_channels = self._stage_out_channels[-1] |
| 110 | self.conv5 = nn.Sequential( |
| 111 | nn.Conv2d(input_channels, output_channels, kernel_size=1, stride=2, padding=0, bias=False), |
| 112 | nn.BatchNorm2d(output_channels), |
| 113 | nn.ReLU(inplace=True) |
| 114 | ) |
| 115 | |
| 116 | self.fc = nn.Linear(output_channels, num_classes) |
| 117 | |
| 118 | def _forward_impl(self, x): |
| 119 | x = self.conv1(x) |
| 120 | x = self.maxpool(x) |
| 121 | x = self.stage2(x) |
| 122 | x = self.stage3(x) |
| 123 | x = self.stage4(x) |
| 124 | x = self.conv5(x) |
| 125 | x = x.mean([2, 3]) |
| 126 | x = self.fc(x) |
| 127 | return x |
| 128 | |
| 129 | def forward(self, x): |
| 130 | return self._forward_impl(x) |
| 131 |
no outgoing calls
no test coverage detected