| | |
|---|
| | | if self.crit_attn_weight > 0: |
|---|
| | | self.attn_loss = torch.nn.L1Loss() |
|---|
| | | self.crit_attn_smooth = crit_attn_smooth |
|---|
| | | self.length_normalized_loss = length_normalized_loss |
|---|
| | | |
|---|
| | | def forward( |
|---|
| | | self, |
|---|
| | |
|---|
| | | |
|---|
| | | stats["loss"] = torch.clone(loss.detach()) |
|---|
| | | # force_gatherable: to-device and to-tensor if scalar for DataParallel |
|---|
| | | if self.length_normalized_loss: |
|---|
| | | batch_size = (text_lengths + self.predictor_bias).sum().type_as(batch_size) |
|---|
| | | loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device) |
|---|
| | | return loss, stats, weight |
|---|
| | | |
|---|
