| | |
|---|
| | | from typing import Union |
|---|
| | | |
|---|
| | | import torch |
|---|
| | | from typeguard import check_argument_types |
|---|
| | | |
|---|
| | | from funasr.layers.abs_normalize import AbsNormalize |
|---|
| | | from funasr.losses.label_smoothing_loss import ( |
|---|
| | |
|---|
| | | preencoder: Optional[AbsPreEncoder] = None, |
|---|
| | | postencoder: Optional[AbsPostEncoder] = None, |
|---|
| | | ): |
|---|
| | | assert check_argument_types() |
|---|
| | | assert 0.0 <= ctc_weight <= 1.0, ctc_weight |
|---|
| | | assert 0.0 <= interctc_weight < 1.0, interctc_weight |
|---|
| | | |
|---|
| | |
|---|
| | | self.ctc = ctc |
|---|
| | | |
|---|
| | | self.extract_feats_in_collect_stats = extract_feats_in_collect_stats |
|---|
| | | 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 = int((text_lengths + 1).sum()) |
|---|
| | | loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device) |
|---|
| | | return loss, stats, weight |
|---|
| | | |
|---|
