Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[BugFix] Force _get_batch_preds() to keep logits in decoder timestamps generator #6499

Merged
merged 5 commits into from
Apr 28, 2023
Merged

[BugFix] Force _get_batch_preds() to keep logits in decoder timestamps generator #6499

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
229a79e
[BugFix] _get_batch_preds() is forced to keep logits in decoder time…
tango4j Apr 26, 2023
ecee173
Merge branch 'main' into fix/get_batch_preds
tango4j Apr 27, 2023
046ad90
Ingnore keep_logits boolean in FrameASRBatchLogits
tango4j Apr 27, 2023
a1a2f42
Merge branches 'fix/get_batch_preds' and 'fix/get_batch_preds' of htt…
tango4j Apr 27, 2023
834bf48
Merge branch 'main' into fix/get_batch_preds
jbalam-nv Apr 27, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 6 additions & 4 deletions nemo/collections/asr/parts/utils/decoder_timestamps_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -232,7 +232,7 @@ def get_wer_feat_logit(audio_file_path, asr, frame_len, tokens_per_chunk, delay,
return hyp, tokens, log_prob


class FrameBatchASR_Logits(FrameBatchASR):
class FrameBatchASRLogits(FrameBatchASR):
"""
A class for streaming frame-based ASR.
Inherits from FrameBatchASR and adds new capability of returning the logit output.
Expand Down Expand Up @@ -260,10 +260,9 @@ def read_audio_file_and_return(self, audio_filepath: str, delay: float, model_st
self.set_frame_reader(frame_reader)

@torch.no_grad()
def _get_batch_preds(self):
def _get_batch_preds(self, keep_logits):
device = self.asr_model.device
for batch in iter(self.data_loader):

feat_signal, feat_signal_len = batch
feat_signal, feat_signal_len = feat_signal.to(device), feat_signal_len.to(device)
log_probs, encoded_len, predictions = self.asr_model(
Expand All @@ -272,9 +271,12 @@ def _get_batch_preds(self):
preds = torch.unbind(predictions)
for pred in preds:
self.all_preds.append(pred.cpu().numpy())
# Always keep logits in FrameBatchASRLogits
_ = keep_logits
log_probs_tup = torch.unbind(log_probs)
for log_prob in log_probs_tup:
self.all_logprobs.append(log_prob)
del log_probs, log_probs_tup
del encoded_len
del predictions

Expand Down Expand Up @@ -635,7 +637,7 @@ def run_ASR_BPE_CTC(self, asr_model: Type[EncDecCTCModelBPE]) -> Tuple[Dict, Dic
log_prediction=asr_model._cfg.get("log_prediction", False),
)

frame_asr = FrameBatchASR_Logits(
frame_asr = FrameBatchASRLogits(
asr_model=asr_model,
frame_len=self.chunk_len_in_sec,
total_buffer=self.total_buffer_in_secs,
Expand Down
Loading