(self, block, layers, num_classes=1000,
dcn=None, stage_with_dcn=(False, False, False, False))
| 181 | |
| 182 | class ResNet(nn.Module): |
| 183 | def __init__(self, block, layers, num_classes=1000, |
| 184 | dcn=None, stage_with_dcn=(False, False, False, False)): |
| 185 | self.dcn = dcn |
| 186 | self.stage_with_dcn = stage_with_dcn |
| 187 | self.inplanes = 64 |
| 188 | super(ResNet, self).__init__() |
| 189 | self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, |
| 190 | bias=False) |
| 191 | self.bn1 = BatchNorm2d(64) |
| 192 | self.relu = nn.ReLU(inplace=True) |
| 193 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) |
| 194 | self.layer1 = self._make_layer(block, 64, layers[0]) |
| 195 | self.layer2 = self._make_layer( |
| 196 | block, 128, layers[1], stride=2, dcn=dcn) |
| 197 | self.layer3 = self._make_layer( |
| 198 | block, 256, layers[2], stride=2, dcn=dcn) |
| 199 | self.layer4 = self._make_layer( |
| 200 | block, 512, layers[3], stride=2, dcn=dcn) |
| 201 | self.avgpool = nn.AvgPool2d(7, stride=1) |
| 202 | self.fc = nn.Linear(512 * block.expansion, num_classes) |
| 203 | |
| 204 | self.smooth = nn.Conv2d(2048, 256, kernel_size=1, stride=1, padding=1) |
| 205 | |
| 206 | for m in self.modules(): |
| 207 | if isinstance(m, nn.Conv2d): |
| 208 | n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels |
| 209 | m.weight.data.normal_(0, math.sqrt(2. / n)) |
| 210 | elif isinstance(m, BatchNorm2d): |
| 211 | m.weight.data.fill_(1) |
| 212 | m.bias.data.zero_() |
| 213 | if self.dcn is not None: |
| 214 | for m in self.modules(): |
| 215 | if isinstance(m, Bottleneck) or isinstance(m, BasicBlock): |
| 216 | if hasattr(m, 'conv2_offset'): |
| 217 | constant_init(m.conv2_offset, 0) |
| 218 | |
| 219 | def _make_layer(self, block, planes, blocks, stride=1, dcn=None): |
| 220 | downsample = None |
no test coverage detected