| 116 | return self.relu(x+y) |
| 117 | |
| 118 | class BasicEncoder(nn.Module): |
| 119 | def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0): |
| 120 | super(BasicEncoder, self).__init__() |
| 121 | self.norm_fn = norm_fn |
| 122 | |
| 123 | if self.norm_fn == 'group': |
| 124 | self.norm1 = nn.GroupNorm(num_groups=8, num_channels=64) |
| 125 | |
| 126 | elif self.norm_fn == 'batch': |
| 127 | self.norm1 = nn.BatchNorm2d(64) |
| 128 | |
| 129 | elif self.norm_fn == 'instance': |
| 130 | self.norm1 = nn.InstanceNorm2d(64) |
| 131 | |
| 132 | elif self.norm_fn == 'none': |
| 133 | self.norm1 = nn.Sequential() |
| 134 | |
| 135 | self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3) |
| 136 | self.relu1 = nn.ReLU(inplace=True) |
| 137 | |
| 138 | self.in_planes = 64 |
| 139 | self.layer1 = self._make_layer(64, stride=1) |
| 140 | self.layer2 = self._make_layer(96, stride=2) |
| 141 | self.layer3 = self._make_layer(128, stride=2) |
| 142 | |
| 143 | # output convolution |
| 144 | self.conv2 = nn.Conv2d(128, output_dim, kernel_size=1) |
| 145 | |
| 146 | self.dropout = None |
| 147 | if dropout > 0: |
| 148 | self.dropout = nn.Dropout2d(p=dropout) |
| 149 | |
| 150 | for m in self.modules(): |
| 151 | if isinstance(m, nn.Conv2d): |
| 152 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') |
| 153 | elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)): |
| 154 | if m.weight is not None: |
| 155 | nn.init.constant_(m.weight, 1) |
| 156 | if m.bias is not None: |
| 157 | nn.init.constant_(m.bias, 0) |
| 158 | |
| 159 | def _make_layer(self, dim, stride=1): |
| 160 | layer1 = ResidualBlock(self.in_planes, dim, self.norm_fn, stride=stride) |
| 161 | layer2 = ResidualBlock(dim, dim, self.norm_fn, stride=1) |
| 162 | layers = (layer1, layer2) |
| 163 | |
| 164 | self.in_planes = dim |
| 165 | return nn.Sequential(*layers) |
| 166 | |
| 167 | |
| 168 | def forward(self, x): |
| 169 | |
| 170 | # if input is list, combine batch dimension |
| 171 | is_list = isinstance(x, tuple) or isinstance(x, list) |
| 172 | if is_list: |
| 173 | batch_dim = x[0].shape[0] |
| 174 | x = torch.cat(x, dim=0) |
| 175 | |