mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Remove some unhelpful or unused options in decode.py, setting equivalent to --left-context=0

Browse Source 
for padding.  Restore default of causal training.
This commit is contained in:
Daniel Povey 2023-02-13 12:58:33 +08:00
parent a5fb97d298
commit 686e7e8828
2 changed files with 20 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
egs/librispeech/ASR/pruned_transducer_stateless7/decode.py
View File

@ -378,20 +378,23 @@ def decode_one_batch(
supervisions = batch["supervisions"]
feature_lens = supervisions["num_frames"].to(device)
feature_lens += params.left_context
feature = torch.nn.functional.pad(
feature,
pad=(0, 0, 0, params.left_context),
value=LOG_EPS,
)
# this seems to cause insertions at the end of the utterance if used with zipformer.
#feature_lens += params.left_context
#feature = torch.nn.functional.pad(
# feature,
# pad=(0, 0, 0, params.left_context),
# value=LOG_EPS,
#)
if params.simulate_streaming:
# the chunk size and left context are now stored with the model.
# TODO: implement streaming_forward.
encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
x=feature,
x_lens=feature_lens,
chunk_size=params.decode_chunk_size,
left_context=params.left_context,
simulate_streaming=True,
#chunk_size=params.decode_chunk_size,
#left_context=params.left_context,
#simulate_streaming=True,
)
else:
encoder_out, encoder_out_lens = model.encoder(
@ -666,9 +669,11 @@ def main():
else:
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
if params.simulate_streaming:
params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
params.suffix += f"-left-context-{params.left_context}"
# TODO: may still want to add something here? for now I am just
# moving the decoding directories around after decoding.
#if params.simulate_streaming:
#params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
#params.suffix += f"-left-context-{params.left_context}"
if "fast_beam_search" in params.decoding_method:
params.suffix += f"-beam-{params.beam}"

6
egs/librispeech/ASR/pruned_transducer_stateless7/train.py
View File

@ -230,15 +230,15 @@ def add_model_arguments(parser: argparse.ArgumentParser):
parser.add_argument(
"--causal",
type=str2bool,
default=False,
default=True,
help="If True, use causal version of model.",
)
parser.add_argument(
"--chunk-size",
type=str,
default="-1", # "16,32,64,-1",
help="Chunk sizes will be chosen randomly from this list during training. "
default="16,32,64,-1",
help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
" Must be just -1 if --causal=False"
)