mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-26 10:16:14 +00:00
Add attention rescoring
This commit is contained in:
parent
ca7dbb085e
commit
83c36ecc18
@ -21,7 +21,16 @@
|
|||||||
"""
|
"""
|
||||||
Usage:
|
Usage:
|
||||||
|
|
||||||
(1) ctc-decoding
|
(1) ctc-greedy-search
|
||||||
|
./zipformer/ctc_decode.py \
|
||||||
|
--epoch 30 \
|
||||||
|
--avg 15 \
|
||||||
|
--exp-dir ./zipformer/exp \
|
||||||
|
--use-ctc 1 \
|
||||||
|
--max-duration 600 \
|
||||||
|
--decoding-method ctc-greedy-search
|
||||||
|
|
||||||
|
(2) ctc-decoding
|
||||||
./zipformer/ctc_decode.py \
|
./zipformer/ctc_decode.py \
|
||||||
--epoch 30 \
|
--epoch 30 \
|
||||||
--avg 15 \
|
--avg 15 \
|
||||||
@ -30,7 +39,7 @@ Usage:
|
|||||||
--max-duration 600 \
|
--max-duration 600 \
|
||||||
--decoding-method ctc-decoding
|
--decoding-method ctc-decoding
|
||||||
|
|
||||||
(2) 1best
|
(3) 1best
|
||||||
./zipformer/ctc_decode.py \
|
./zipformer/ctc_decode.py \
|
||||||
--epoch 30 \
|
--epoch 30 \
|
||||||
--avg 15 \
|
--avg 15 \
|
||||||
@ -40,7 +49,7 @@ Usage:
|
|||||||
--hlg-scale 0.6 \
|
--hlg-scale 0.6 \
|
||||||
--decoding-method 1best
|
--decoding-method 1best
|
||||||
|
|
||||||
(3) nbest
|
(4) nbest
|
||||||
./zipformer/ctc_decode.py \
|
./zipformer/ctc_decode.py \
|
||||||
--epoch 30 \
|
--epoch 30 \
|
||||||
--avg 15 \
|
--avg 15 \
|
||||||
@ -50,7 +59,7 @@ Usage:
|
|||||||
--hlg-scale 0.6 \
|
--hlg-scale 0.6 \
|
||||||
--decoding-method nbest
|
--decoding-method nbest
|
||||||
|
|
||||||
(4) nbest-rescoring
|
(5) nbest-rescoring
|
||||||
./zipformer/ctc_decode.py \
|
./zipformer/ctc_decode.py \
|
||||||
--epoch 30 \
|
--epoch 30 \
|
||||||
--avg 15 \
|
--avg 15 \
|
||||||
@ -62,7 +71,7 @@ Usage:
|
|||||||
--lm-dir data/lm \
|
--lm-dir data/lm \
|
||||||
--decoding-method nbest-rescoring
|
--decoding-method nbest-rescoring
|
||||||
|
|
||||||
(5) whole-lattice-rescoring
|
(6) whole-lattice-rescoring
|
||||||
./zipformer/ctc_decode.py \
|
./zipformer/ctc_decode.py \
|
||||||
--epoch 30 \
|
--epoch 30 \
|
||||||
--avg 15 \
|
--avg 15 \
|
||||||
@ -73,6 +82,29 @@ Usage:
|
|||||||
--nbest-scale 1.0 \
|
--nbest-scale 1.0 \
|
||||||
--lm-dir data/lm \
|
--lm-dir data/lm \
|
||||||
--decoding-method whole-lattice-rescoring
|
--decoding-method whole-lattice-rescoring
|
||||||
|
|
||||||
|
(7) attention-decoder-rescoring-no-ngram
|
||||||
|
./zipformer/ctc_decode.py \
|
||||||
|
--epoch 30 \
|
||||||
|
--avg 15 \
|
||||||
|
--exp-dir ./zipformer/exp \
|
||||||
|
--use-ctc 1 \
|
||||||
|
--use-attention-decoder 1 \
|
||||||
|
--max-duration 100 \
|
||||||
|
--decoding-method attention-decoder-rescoring-no-ngram
|
||||||
|
|
||||||
|
(8) attention-decoder-rescoring-with-ngram
|
||||||
|
./zipformer/ctc_decode.py \
|
||||||
|
--epoch 30 \
|
||||||
|
--avg 15 \
|
||||||
|
--exp-dir ./zipformer/exp \
|
||||||
|
--use-ctc 1 \
|
||||||
|
--use-attention-decoder 1 \
|
||||||
|
--max-duration 100 \
|
||||||
|
--hlg-scale 0.6 \
|
||||||
|
--nbest-scale 1.0 \
|
||||||
|
--lm-dir data/lm \
|
||||||
|
--decoding-method attention-decoder-rescoring-with-ngram
|
||||||
"""
|
"""
|
||||||
|
|
||||||
|
|
||||||
@ -87,9 +119,11 @@ import k2
|
|||||||
import sentencepiece as spm
|
import sentencepiece as spm
|
||||||
import torch
|
import torch
|
||||||
import torch.nn as nn
|
import torch.nn as nn
|
||||||
from asr_datamodule import GigaSpeechAsrDataModule
|
|
||||||
|
|
||||||
|
from asr_datamodule import GigaSpeechAsrDataModule
|
||||||
from gigaspeech_scoring import asr_text_post_processing
|
from gigaspeech_scoring import asr_text_post_processing
|
||||||
|
|
||||||
|
from lhotse import set_caching_enabled
|
||||||
from train import add_model_arguments, get_model, get_params
|
from train import add_model_arguments, get_model, get_params
|
||||||
|
|
||||||
from icefall.checkpoint import (
|
from icefall.checkpoint import (
|
||||||
@ -99,10 +133,13 @@ from icefall.checkpoint import (
|
|||||||
load_checkpoint,
|
load_checkpoint,
|
||||||
)
|
)
|
||||||
from icefall.decode import (
|
from icefall.decode import (
|
||||||
|
ctc_greedy_search,
|
||||||
get_lattice,
|
get_lattice,
|
||||||
nbest_decoding,
|
nbest_decoding,
|
||||||
nbest_oracle,
|
nbest_oracle,
|
||||||
one_best_decoding,
|
one_best_decoding,
|
||||||
|
rescore_with_attention_decoder_no_ngram,
|
||||||
|
rescore_with_attention_decoder_with_ngram,
|
||||||
rescore_with_n_best_list,
|
rescore_with_n_best_list,
|
||||||
rescore_with_whole_lattice,
|
rescore_with_whole_lattice,
|
||||||
)
|
)
|
||||||
@ -197,23 +234,30 @@ def get_parser():
|
|||||||
default="ctc-decoding",
|
default="ctc-decoding",
|
||||||
help="""Decoding method.
|
help="""Decoding method.
|
||||||
Supported values are:
|
Supported values are:
|
||||||
- (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
|
- (1) ctc-greedy-search. Use CTC greedy search. It uses a sentence piece
|
||||||
model, i.e., lang_dir/bpe.model, to convert word pieces to words.
|
model, i.e., lang_dir/bpe.model, to convert word pieces to words.
|
||||||
It needs neither a lexicon nor an n-gram LM.
|
It needs neither a lexicon nor an n-gram LM.
|
||||||
- (2) 1best. Extract the best path from the decoding lattice as the
|
- (2) ctc-decoding. Use CTC decoding. It uses a sentence piece
|
||||||
|
model, i.e., lang_dir/bpe.model, to convert word pieces to words.
|
||||||
|
It needs neither a lexicon nor an n-gram LM.
|
||||||
|
- (3) 1best. Extract the best path from the decoding lattice as the
|
||||||
decoding result.
|
decoding result.
|
||||||
- (3) nbest. Extract n paths from the decoding lattice; the path
|
- (4) nbest. Extract n paths from the decoding lattice; the path
|
||||||
with the highest score is the decoding result.
|
with the highest score is the decoding result.
|
||||||
- (4) nbest-rescoring. Extract n paths from the decoding lattice,
|
- (5) nbest-rescoring. Extract n paths from the decoding lattice,
|
||||||
rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
|
rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
|
||||||
the highest score is the decoding result.
|
the highest score is the decoding result.
|
||||||
- (5) whole-lattice-rescoring. Rescore the decoding lattice with an
|
- (6) whole-lattice-rescoring. Rescore the decoding lattice with an
|
||||||
n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
|
n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
|
||||||
is the decoding result.
|
is the decoding result.
|
||||||
you have trained an RNN LM using ./rnn_lm/train.py
|
you have trained an RNN LM using ./rnn_lm/train.py
|
||||||
- (6) nbest-oracle. Its WER is the lower bound of any n-best
|
- (7) nbest-oracle. Its WER is the lower bound of any n-best
|
||||||
rescoring method can achieve. Useful for debugging n-best
|
rescoring method can achieve. Useful for debugging n-best
|
||||||
rescoring method.
|
rescoring method.
|
||||||
|
- (8) attention-decoder-rescoring-no-ngram. Extract n paths from the decoding
|
||||||
|
lattice, rescore them with the attention decoder.
|
||||||
|
- (9) attention-decoder-rescoring-with-ngram. Extract n paths from the LM
|
||||||
|
rescored lattice, rescore them with the attention decoder.
|
||||||
""",
|
""",
|
||||||
)
|
)
|
||||||
|
|
||||||
@ -256,6 +300,13 @@ def get_parser():
|
|||||||
""",
|
""",
|
||||||
)
|
)
|
||||||
|
|
||||||
|
parser.add_argument(
|
||||||
|
"--skip-scoring",
|
||||||
|
type=str2bool,
|
||||||
|
default=False,
|
||||||
|
help="""Skip scoring, but still save the ASR output (for eval sets).""",
|
||||||
|
)
|
||||||
|
|
||||||
add_model_arguments(parser)
|
add_model_arguments(parser)
|
||||||
|
|
||||||
return parser
|
return parser
|
||||||
@ -276,17 +327,6 @@ def get_decoding_params() -> AttributeDict:
|
|||||||
return params
|
return params
|
||||||
|
|
||||||
|
|
||||||
def post_processing(
|
|
||||||
results: List[Tuple[str, List[str], List[str]]],
|
|
||||||
) -> List[Tuple[str, List[str], List[str]]]:
|
|
||||||
new_results = []
|
|
||||||
for key, ref, hyp in results:
|
|
||||||
new_ref = asr_text_post_processing(" ".join(ref)).split()
|
|
||||||
new_hyp = asr_text_post_processing(" ".join(hyp)).split()
|
|
||||||
new_results.append((key, new_ref, new_hyp))
|
|
||||||
return new_results
|
|
||||||
|
|
||||||
|
|
||||||
def decode_one_batch(
|
def decode_one_batch(
|
||||||
params: AttributeDict,
|
params: AttributeDict,
|
||||||
model: nn.Module,
|
model: nn.Module,
|
||||||
@ -365,6 +405,15 @@ def decode_one_batch(
|
|||||||
encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
|
encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
|
||||||
ctc_output = model.ctc_output(encoder_out) # (N, T, C)
|
ctc_output = model.ctc_output(encoder_out) # (N, T, C)
|
||||||
|
|
||||||
|
if params.decoding_method == "ctc-greedy-search":
|
||||||
|
hyps = ctc_greedy_search(ctc_output, encoder_out_lens)
|
||||||
|
# hyps is a list of str, e.g., ['xxx yyy zzz', ...]
|
||||||
|
hyps = bpe_model.decode(hyps)
|
||||||
|
# hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
|
||||||
|
hyps = [s.split() for s in hyps]
|
||||||
|
key = "ctc-greedy-search"
|
||||||
|
return {key: hyps}
|
||||||
|
|
||||||
supervision_segments = torch.stack(
|
supervision_segments = torch.stack(
|
||||||
(
|
(
|
||||||
supervisions["sequence_idx"],
|
supervisions["sequence_idx"],
|
||||||
@ -417,7 +466,27 @@ def decode_one_batch(
|
|||||||
# hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
|
# hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
|
||||||
hyps = [s.split() for s in hyps]
|
hyps = [s.split() for s in hyps]
|
||||||
key = "ctc-decoding"
|
key = "ctc-decoding"
|
||||||
return {key: hyps}
|
return {key: hyps} # note: returns words
|
||||||
|
|
||||||
|
if params.decoding_method == "attention-decoder-rescoring-no-ngram":
|
||||||
|
best_path_dict = rescore_with_attention_decoder_no_ngram(
|
||||||
|
lattice=lattice,
|
||||||
|
num_paths=params.num_paths,
|
||||||
|
attention_decoder=model.attention_decoder,
|
||||||
|
encoder_out=encoder_out,
|
||||||
|
encoder_out_lens=encoder_out_lens,
|
||||||
|
nbest_scale=params.nbest_scale,
|
||||||
|
)
|
||||||
|
ans = dict()
|
||||||
|
for a_scale_str, best_path in best_path_dict.items():
|
||||||
|
# token_ids is a lit-of-list of IDs
|
||||||
|
token_ids = get_texts(best_path)
|
||||||
|
# hyps is a list of str, e.g., ['xxx yyy zzz', ...]
|
||||||
|
hyps = bpe_model.decode(token_ids)
|
||||||
|
# hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
|
||||||
|
hyps = [s.split() for s in hyps]
|
||||||
|
ans[a_scale_str] = hyps
|
||||||
|
return ans
|
||||||
|
|
||||||
if params.decoding_method == "nbest-oracle":
|
if params.decoding_method == "nbest-oracle":
|
||||||
# Note: You can also pass rescored lattices to it.
|
# Note: You can also pass rescored lattices to it.
|
||||||
@ -434,7 +503,7 @@ def decode_one_batch(
|
|||||||
)
|
)
|
||||||
hyps = get_texts(best_path)
|
hyps = get_texts(best_path)
|
||||||
hyps = [[word_table[i] for i in ids] for ids in hyps]
|
hyps = [[word_table[i] for i in ids] for ids in hyps]
|
||||||
key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa
|
key = f"oracle_{params.num_paths}_nbest-scale-{params.nbest_scale}" # noqa
|
||||||
return {key: hyps}
|
return {key: hyps}
|
||||||
|
|
||||||
if params.decoding_method in ["1best", "nbest"]:
|
if params.decoding_method in ["1best", "nbest"]:
|
||||||
@ -442,7 +511,7 @@ def decode_one_batch(
|
|||||||
best_path = one_best_decoding(
|
best_path = one_best_decoding(
|
||||||
lattice=lattice, use_double_scores=params.use_double_scores
|
lattice=lattice, use_double_scores=params.use_double_scores
|
||||||
)
|
)
|
||||||
key = "no_rescore"
|
key = "no-rescore"
|
||||||
else:
|
else:
|
||||||
best_path = nbest_decoding(
|
best_path = nbest_decoding(
|
||||||
lattice=lattice,
|
lattice=lattice,
|
||||||
@ -450,15 +519,16 @@ def decode_one_batch(
|
|||||||
use_double_scores=params.use_double_scores,
|
use_double_scores=params.use_double_scores,
|
||||||
nbest_scale=params.nbest_scale,
|
nbest_scale=params.nbest_scale,
|
||||||
)
|
)
|
||||||
key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa
|
key = f"no-rescore_nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa
|
||||||
|
|
||||||
hyps = get_texts(best_path)
|
hyps = get_texts(best_path)
|
||||||
hyps = [[word_table[i] for i in ids] for ids in hyps]
|
hyps = [[word_table[i] for i in ids] for ids in hyps]
|
||||||
return {key: hyps}
|
return {key: hyps} # note: returns BPE tokens
|
||||||
|
|
||||||
assert params.decoding_method in [
|
assert params.decoding_method in [
|
||||||
"nbest-rescoring",
|
"nbest-rescoring",
|
||||||
"whole-lattice-rescoring",
|
"whole-lattice-rescoring",
|
||||||
|
"attention-decoder-rescoring-with-ngram",
|
||||||
]
|
]
|
||||||
|
|
||||||
lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
|
lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
|
||||||
@ -479,6 +549,21 @@ def decode_one_batch(
|
|||||||
G_with_epsilon_loops=G,
|
G_with_epsilon_loops=G,
|
||||||
lm_scale_list=lm_scale_list,
|
lm_scale_list=lm_scale_list,
|
||||||
)
|
)
|
||||||
|
elif params.decoding_method == "attention-decoder-rescoring-with-ngram":
|
||||||
|
# lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
|
||||||
|
rescored_lattice = rescore_with_whole_lattice(
|
||||||
|
lattice=lattice,
|
||||||
|
G_with_epsilon_loops=G,
|
||||||
|
lm_scale_list=None,
|
||||||
|
)
|
||||||
|
best_path_dict = rescore_with_attention_decoder_with_ngram(
|
||||||
|
lattice=rescored_lattice,
|
||||||
|
num_paths=params.num_paths,
|
||||||
|
attention_decoder=model.attention_decoder,
|
||||||
|
encoder_out=encoder_out,
|
||||||
|
encoder_out_lens=encoder_out_lens,
|
||||||
|
nbest_scale=params.nbest_scale,
|
||||||
|
)
|
||||||
else:
|
else:
|
||||||
assert False, f"Unsupported decoding method: {params.decoding_method}"
|
assert False, f"Unsupported decoding method: {params.decoding_method}"
|
||||||
|
|
||||||
@ -572,39 +657,64 @@ def decode_dataset(
|
|||||||
return results
|
return results
|
||||||
|
|
||||||
|
|
||||||
def save_results(
|
def save_asr_output(
|
||||||
params: AttributeDict,
|
params: AttributeDict,
|
||||||
test_set_name: str,
|
test_set_name: str,
|
||||||
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
|
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
|
||||||
):
|
):
|
||||||
test_set_wers = dict()
|
"""
|
||||||
|
Save text produced by ASR.
|
||||||
|
"""
|
||||||
for key, results in results_dict.items():
|
for key, results in results_dict.items():
|
||||||
recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
|
recogs_filename = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
|
||||||
results = post_processing(results)
|
results = post_processing(results)
|
||||||
results = sorted(results)
|
results = sorted(results)
|
||||||
store_transcripts(filename=recog_path, texts=results)
|
store_transcripts(filename=recogs_filename, texts=results)
|
||||||
logging.info(f"The transcripts are stored in {recog_path}")
|
|
||||||
|
|
||||||
|
logging.info(f"The transcripts are stored in {recogs_filename}")
|
||||||
|
|
||||||
|
|
||||||
|
def save_wer_results(
|
||||||
|
params: AttributeDict,
|
||||||
|
test_set_name: str,
|
||||||
|
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
|
||||||
|
):
|
||||||
|
if params.decoding_method in (
|
||||||
|
"attention-decoder-rescoring-with-ngram",
|
||||||
|
"whole-lattice-rescoring",
|
||||||
|
):
|
||||||
|
# Set it to False since there are too many logs.
|
||||||
|
enable_log = False
|
||||||
|
else:
|
||||||
|
enable_log = True
|
||||||
|
|
||||||
|
test_set_wers = dict()
|
||||||
|
for key, results in results_dict.items():
|
||||||
|
results = post_processing(results)
|
||||||
# The following prints out WERs, per-word error statistics and aligned
|
# The following prints out WERs, per-word error statistics and aligned
|
||||||
# ref/hyp pairs.
|
# ref/hyp pairs.
|
||||||
errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
|
errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
|
||||||
with open(errs_filename, "w") as f:
|
with open(errs_filename, "w", encoding="utf8") as fd:
|
||||||
wer = write_error_stats(f, f"{test_set_name}-{key}", results)
|
wer = write_error_stats(
|
||||||
|
fd, f"{test_set_name}_{key}", results, enable_log=enable_log
|
||||||
|
)
|
||||||
test_set_wers[key] = wer
|
test_set_wers[key] = wer
|
||||||
|
|
||||||
logging.info("Wrote detailed error stats to {}".format(errs_filename))
|
logging.info(f"Wrote detailed error stats to {errs_filename}")
|
||||||
|
|
||||||
test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
|
test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
|
||||||
errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
|
|
||||||
with open(errs_info, "w") as f:
|
|
||||||
print("settings\tWER", file=f)
|
|
||||||
for key, val in test_set_wers:
|
|
||||||
print("{}\t{}".format(key, val), file=f)
|
|
||||||
|
|
||||||
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
|
wer_filename = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
|
||||||
note = "\tbest for {}".format(test_set_name)
|
|
||||||
|
with open(wer_filename, "w", encoding="utf8") as fd:
|
||||||
|
print("settings\tWER", file=fd)
|
||||||
|
for key, val in test_set_wers:
|
||||||
|
print(f"{key}\t{val}", file=fd)
|
||||||
|
|
||||||
|
s = f"\nFor {test_set_name}, WER of different settings are:\n"
|
||||||
|
note = f"\tbest for {test_set_name}"
|
||||||
for key, val in test_set_wers:
|
for key, val in test_set_wers:
|
||||||
s += "{}\t{}{}\n".format(key, val, note)
|
s += f"{key}\t{val}{note}\n"
|
||||||
note = ""
|
note = ""
|
||||||
logging.info(s)
|
logging.info(s)
|
||||||
|
|
||||||
@ -623,20 +733,26 @@ def main():
|
|||||||
params.update(get_decoding_params())
|
params.update(get_decoding_params())
|
||||||
params.update(vars(args))
|
params.update(vars(args))
|
||||||
|
|
||||||
|
# enable AudioCache
|
||||||
|
set_caching_enabled(True) # lhotse
|
||||||
|
|
||||||
assert params.decoding_method in (
|
assert params.decoding_method in (
|
||||||
|
"ctc-greedy-search",
|
||||||
"ctc-decoding",
|
"ctc-decoding",
|
||||||
"1best",
|
"1best",
|
||||||
"nbest",
|
"nbest",
|
||||||
"nbest-rescoring",
|
"nbest-rescoring",
|
||||||
"whole-lattice-rescoring",
|
"whole-lattice-rescoring",
|
||||||
"nbest-oracle",
|
"nbest-oracle",
|
||||||
|
"attention-decoder-rescoring-no-ngram",
|
||||||
|
"attention-decoder-rescoring-with-ngram",
|
||||||
)
|
)
|
||||||
params.res_dir = params.exp_dir / params.decoding_method
|
params.res_dir = params.exp_dir / params.decoding_method
|
||||||
|
|
||||||
if params.iter > 0:
|
if params.iter > 0:
|
||||||
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
|
params.suffix = f"iter-{params.iter}_avg-{params.avg}"
|
||||||
else:
|
else:
|
||||||
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
|
params.suffix = f"epoch-{params.epoch}_avg-{params.avg}"
|
||||||
|
|
||||||
if params.causal:
|
if params.causal:
|
||||||
assert (
|
assert (
|
||||||
@ -645,11 +761,11 @@ def main():
|
|||||||
assert (
|
assert (
|
||||||
"," not in params.left_context_frames
|
"," not in params.left_context_frames
|
||||||
), "left_context_frames should be one value in decoding."
|
), "left_context_frames should be one value in decoding."
|
||||||
params.suffix += f"-chunk-{params.chunk_size}"
|
params.suffix += f"_chunk-{params.chunk_size}"
|
||||||
params.suffix += f"-left-context-{params.left_context_frames}"
|
params.suffix += f"_left-context-{params.left_context_frames}"
|
||||||
|
|
||||||
if params.use_averaged_model:
|
if params.use_averaged_model:
|
||||||
params.suffix += "-use-averaged-model"
|
params.suffix += "_use-averaged-model"
|
||||||
|
|
||||||
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
|
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
|
||||||
logging.info("Decoding started")
|
logging.info("Decoding started")
|
||||||
@ -668,8 +784,14 @@ def main():
|
|||||||
params.vocab_size = num_classes
|
params.vocab_size = num_classes
|
||||||
# <blk> and <unk> are defined in local/train_bpe_model.py
|
# <blk> and <unk> are defined in local/train_bpe_model.py
|
||||||
params.blank_id = 0
|
params.blank_id = 0
|
||||||
|
params.eos_id = 1
|
||||||
|
params.sos_id = 1
|
||||||
|
|
||||||
if params.decoding_method == "ctc-decoding":
|
if params.decoding_method in [
|
||||||
|
"ctc-greedy-search",
|
||||||
|
"ctc-decoding",
|
||||||
|
"attention-decoder-rescoring-no-ngram",
|
||||||
|
]:
|
||||||
HLG = None
|
HLG = None
|
||||||
H = k2.ctc_topo(
|
H = k2.ctc_topo(
|
||||||
max_token=max_token_id,
|
max_token=max_token_id,
|
||||||
@ -693,6 +815,7 @@ def main():
|
|||||||
if params.decoding_method in (
|
if params.decoding_method in (
|
||||||
"nbest-rescoring",
|
"nbest-rescoring",
|
||||||
"whole-lattice-rescoring",
|
"whole-lattice-rescoring",
|
||||||
|
"attention-decoder-rescoring-with-ngram",
|
||||||
):
|
):
|
||||||
if not (params.lm_dir / "G_4_gram.pt").is_file():
|
if not (params.lm_dir / "G_4_gram.pt").is_file():
|
||||||
logging.info("Loading G_4_gram.fst.txt")
|
logging.info("Loading G_4_gram.fst.txt")
|
||||||
@ -724,7 +847,10 @@ def main():
|
|||||||
d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
|
d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
|
||||||
G = k2.Fsa.from_dict(d)
|
G = k2.Fsa.from_dict(d)
|
||||||
|
|
||||||
if params.decoding_method == "whole-lattice-rescoring":
|
if params.decoding_method in [
|
||||||
|
"whole-lattice-rescoring",
|
||||||
|
"attention-decoder-rescoring-with-ngram",
|
||||||
|
]:
|
||||||
# Add epsilon self-loops to G as we will compose
|
# Add epsilon self-loops to G as we will compose
|
||||||
# it with the whole lattice later
|
# it with the whole lattice later
|
||||||
G = k2.add_epsilon_self_loops(G)
|
G = k2.add_epsilon_self_loops(G)
|
||||||
@ -825,6 +951,7 @@ def main():
|
|||||||
|
|
||||||
# we need cut ids to display recognition results.
|
# we need cut ids to display recognition results.
|
||||||
args.return_cuts = True
|
args.return_cuts = True
|
||||||
|
|
||||||
gigaspeech = GigaSpeechAsrDataModule(args)
|
gigaspeech = GigaSpeechAsrDataModule(args)
|
||||||
|
|
||||||
test_cuts = gigaspeech.test_cuts()
|
test_cuts = gigaspeech.test_cuts()
|
||||||
@ -832,9 +959,9 @@ def main():
|
|||||||
test_dl = gigaspeech.test_dataloaders(test_cuts)
|
test_dl = gigaspeech.test_dataloaders(test_cuts)
|
||||||
|
|
||||||
test_sets = ["test"]
|
test_sets = ["test"]
|
||||||
test_dl = [test_dl]
|
test_dls = [test_dl]
|
||||||
|
|
||||||
for test_set, test_dl in zip(test_sets, test_dl):
|
for test_set, test_dl in zip(test_sets, test_dls):
|
||||||
results_dict = decode_dataset(
|
results_dict = decode_dataset(
|
||||||
dl=test_dl,
|
dl=test_dl,
|
||||||
params=params,
|
params=params,
|
||||||
@ -846,12 +973,19 @@ def main():
|
|||||||
G=G,
|
G=G,
|
||||||
)
|
)
|
||||||
|
|
||||||
save_results(
|
save_asr_output(
|
||||||
params=params,
|
params=params,
|
||||||
test_set_name=test_set,
|
test_set_name=test_set,
|
||||||
results_dict=results_dict,
|
results_dict=results_dict,
|
||||||
)
|
)
|
||||||
|
|
||||||
|
if not params.skip_scoring:
|
||||||
|
save_wer_results(
|
||||||
|
params=params,
|
||||||
|
test_set_name=test_set,
|
||||||
|
results_dict=results_dict,
|
||||||
|
)
|
||||||
|
|
||||||
logging.info("Done!")
|
logging.info("Done!")
|
||||||
|
|
||||||
|
|
||||||
|
Loading…
x
Reference in New Issue
Block a user