| | |
|---|
| | | stats["loss_pre"] = loss_pre.detach().cpu() if loss_pre is not None else None |
|---|
| | | |
|---|
| | | stats["loss"] = torch.clone(loss.detach()) |
|---|
| | | stats["batch_size"] = batch_size |
|---|
| | | |
|---|
| | | # force_gatherable: to-device and to-tensor if scalar for DataParallel |
|---|
| | | if self.length_normalized_loss: |
|---|
| | |
|---|
| | | speech = speech.to(device=kwargs["device"]) |
|---|
| | | speech_lengths = speech_lengths.to(device=kwargs["device"]) |
|---|
| | | # Encoder |
|---|
| | | if kwargs.get("fp16", False): |
|---|
| | | speech = speech.half() |
|---|
| | | encoder_out, encoder_out_lens = self.encode(speech, speech_lengths) |
|---|
| | | if isinstance(encoder_out, tuple): |
|---|
| | | encoder_out = encoder_out[0] |
|---|
