| | |
|---|
| | | self.enable_maas_finetune = enable_maas_finetune |
|---|
| | | self.freeze_encoder2 = freeze_encoder2 |
|---|
| | | self.encoder1_encoder2_joint_training = encoder1_encoder2_joint_training |
|---|
| | | self.length_normalized_loss = length_normalized_loss |
|---|
| | | |
|---|
| | | def forward( |
|---|
| | | self, |
|---|
| | |
|---|
| | | stats["loss2"] = torch.clone(loss2.detach()) |
|---|
| | | 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 + 1).sum().type_as(batch_size) |
|---|
| | | loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device) |
|---|
| | | return loss, stats, weight |
|---|
| | | |
|---|
