(self, features)
| 150 | self.avgpool = nn.AdaptiveAvgPool2d((None, 1)) |
| 151 | |
| 152 | def forward(self, features): |
| 153 | batch_size = features.size(0) |
| 154 | num_steps = features.size(3) |
| 155 | features = self.avgpool(features.permute(0, 1, 3, 2)).squeeze(3) |
| 156 | hidden = (torch.FloatTensor(batch_size, self.hidden_size).zero_().to(device), |
| 157 | torch.FloatTensor(batch_size, self.hidden_size).zero_().to(device)) |
| 158 | for i in range(num_steps): |
| 159 | hidden = self.sequence_cell(features[:, :, i], hidden) |
| 160 | holistic_feature = hidden |
| 161 | |
| 162 | return holistic_feature |
| 163 | |
| 164 | |
| 165 | class LSTMDecoder(nn.Module): |
nothing calls this directly
no outgoing calls
no test coverage detected