| 193 | |
| 194 | |
| 195 | class SmallEncoder(nn.Module): |
| 196 | def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0): |
| 197 | super(SmallEncoder, self).__init__() |
| 198 | self.norm_fn = norm_fn |
| 199 | |
| 200 | if self.norm_fn == 'group': |
| 201 | self.norm1 = nn.GroupNorm(num_groups=8, num_channels=32) |
| 202 | |
| 203 | elif self.norm_fn == 'batch': |
| 204 | self.norm1 = nn.BatchNorm2d(32) |
| 205 | |
| 206 | elif self.norm_fn == 'instance': |
| 207 | self.norm1 = nn.InstanceNorm2d(32) |
| 208 | |
| 209 | elif self.norm_fn == 'none': |
| 210 | self.norm1 = nn.Sequential() |
| 211 | |
| 212 | self.conv1 = nn.Conv2d(3, 32, kernel_size=7, stride=2, padding=3) |
| 213 | self.relu1 = nn.ReLU(inplace=True) |
| 214 | |
| 215 | self.in_planes = 32 |
| 216 | self.layer1 = self._make_layer(32, stride=1) |
| 217 | self.layer2 = self._make_layer(64, stride=2) |
| 218 | self.layer3 = self._make_layer(96, stride=2) |
| 219 | |
| 220 | self.dropout = None |
| 221 | if dropout > 0: |
| 222 | self.dropout = nn.Dropout2d(p=dropout) |
| 223 | |
| 224 | self.conv2 = nn.Conv2d(96, output_dim, kernel_size=1) |
| 225 | |
| 226 | for m in self.modules(): |
| 227 | if isinstance(m, nn.Conv2d): |
| 228 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') |
| 229 | elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)): |
| 230 | if m.weight is not None: |
| 231 | nn.init.constant_(m.weight, 1) |
| 232 | if m.bias is not None: |
| 233 | nn.init.constant_(m.bias, 0) |
| 234 | |
| 235 | def _make_layer(self, dim, stride=1): |
| 236 | layer1 = BottleneckBlock(self.in_planes, dim, self.norm_fn, stride=stride) |
| 237 | layer2 = BottleneckBlock(dim, dim, self.norm_fn, stride=1) |
| 238 | layers = (layer1, layer2) |
| 239 | |
| 240 | self.in_planes = dim |
| 241 | return nn.Sequential(*layers) |
| 242 | |
| 243 | |
| 244 | def forward(self, x): |
| 245 | |
| 246 | # if input is list, combine batch dimension |
| 247 | is_list = isinstance(x, tuple) or isinstance(x, list) |
| 248 | if is_list: |
| 249 | batch_dim = x[0].shape[0] |
| 250 | x = torch.cat(x, dim=0) |
| 251 | |
| 252 | x = self.conv1(x) |