mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-09 01:52:41 +00:00
c970df512b
* initial commit * update readme * Update README.md * change bool to str2bool for arg parser * run validation only at the end of epoch * black format * black format
718 lines
24 KiB
Python
718 lines
24 KiB
Python
import argparse
|
|
import logging
|
|
import math
|
|
from collections import defaultdict
|
|
from pathlib import Path
|
|
from typing import Dict, List, Optional, Tuple
|
|
import pprint
|
|
import k2
|
|
import sentencepiece as spm
|
|
import torch
|
|
import torch.nn as nn
|
|
from asr_datamodule import LibriSpeechAsrDataModule
|
|
from beam_search import (
|
|
beam_search,
|
|
fast_beam_search_nbest,
|
|
fast_beam_search_nbest_LG,
|
|
fast_beam_search_nbest_oracle,
|
|
fast_beam_search_one_best,
|
|
greedy_search,
|
|
greedy_search_batch,
|
|
modified_beam_search,
|
|
)
|
|
from train import add_model_arguments, get_params, get_transducer_model
|
|
|
|
from icefall.checkpoint import (
|
|
average_checkpoints,
|
|
average_checkpoints_with_averaged_model,
|
|
find_checkpoints,
|
|
load_checkpoint,
|
|
)
|
|
from icefall.lexicon import Lexicon
|
|
from icefall.utils import (
|
|
AttributeDict,
|
|
setup_logger,
|
|
store_transcripts,
|
|
str2bool,
|
|
write_error_stats,
|
|
)
|
|
|
|
LOG_EPS = math.log(1e-10)
|
|
|
|
|
|
def get_parser():
|
|
parser = argparse.ArgumentParser(
|
|
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--exp-dir",
|
|
type=str,
|
|
default="tiny_transducer_ctc/exp",
|
|
help="The experiment dir",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--epoch",
|
|
type=int,
|
|
default=30,
|
|
help="""It specifies the checkpoint to use for decoding.
|
|
Note: Epoch counts from 1.
|
|
You can specify --avg to use more checkpoints for model averaging.""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--iter",
|
|
type=int,
|
|
default=0,
|
|
help="""If positive, --epoch is ignored and it
|
|
will use the checkpoint exp_dir/checkpoint-iter.pt.
|
|
You can specify --avg to use more checkpoints for model averaging.
|
|
""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--avg",
|
|
type=int,
|
|
default=1,
|
|
help="Number of checkpoints to average. Automatically select "
|
|
"consecutive checkpoints before the checkpoint specified by "
|
|
"'--epoch' and '--iter'",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--use-averaged-model",
|
|
type=str2bool,
|
|
default=True,
|
|
help="Whether to load averaged model. Currently it only supports "
|
|
"using --epoch. If True, it would decode with the averaged model "
|
|
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
|
|
"Actually only the models with epoch number of `epoch-avg` and "
|
|
"`epoch` are loaded for averaging. ",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--lang-dir",
|
|
type=Path,
|
|
default="data/lang_bpe_500",
|
|
help="The lang dir containing word table and LG graph",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--decoding-method",
|
|
type=str,
|
|
default="fast_beam_search",
|
|
help="""Possible values are:
|
|
- greedy_search
|
|
- beam_search
|
|
- modified_beam_search
|
|
- fast_beam_search
|
|
- fast_beam_search_LG
|
|
- fast_beam_search_nbest
|
|
- fast_beam_search_nbest_oracle
|
|
- fast_beam_search_nbest_LG
|
|
""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--beam-size",
|
|
type=int,
|
|
default=4,
|
|
help="""An integer indicating how many candidates we will keep for each
|
|
frame. Used only when --decoding-method is beam_search or
|
|
modified_beam_search.""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--beam",
|
|
type=float,
|
|
default=20.0,
|
|
help="""A floating point value to calculate the cutoff score during beam
|
|
search (i.e., `cutoff = max-score - beam`), which is the same as the
|
|
`beam` in Kaldi.
|
|
Used only when --decoding-method is fast_beam_search, fast_beam_search_LG,
|
|
fast_beam_search_nbest, fast_beam_search_nbest_LG,
|
|
and fast_beam_search_nbest_oracle
|
|
""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--ngram-lm-scale",
|
|
type=float,
|
|
default=0.1,
|
|
help="""
|
|
Used only when --decoding_method is fast_beam_search_LG or
|
|
fast_beam_search_nbest_LG.
|
|
It specifies the scale for n-gram LM scores.
|
|
""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--max-contexts",
|
|
type=int,
|
|
default=8,
|
|
help="""Used only when --decoding-method is
|
|
fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
|
|
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--max-states",
|
|
type=int,
|
|
default=64,
|
|
help="""Used only when --decoding-method is
|
|
fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
|
|
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--context-size",
|
|
type=int,
|
|
default=2,
|
|
help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
|
|
)
|
|
parser.add_argument(
|
|
"--max-sym-per-frame",
|
|
type=int,
|
|
default=1,
|
|
help="""Maximum number of symbols per frame.
|
|
Used only when --decoding_method is greedy_search""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--num-paths",
|
|
type=int,
|
|
default=100,
|
|
help="""Number of paths for nbest decoding.
|
|
Used only when the decoding method is fast_beam_search_nbest,
|
|
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--nbest-scale",
|
|
type=float,
|
|
default=0.5,
|
|
help="""Scale applied to lattice scores when computing nbest paths.
|
|
Used only when the decoding method is fast_beam_search_nbest,
|
|
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
|
|
)
|
|
|
|
add_model_arguments(parser)
|
|
|
|
return parser
|
|
|
|
|
|
def decode_one_batch(
|
|
params: AttributeDict,
|
|
model: nn.Module,
|
|
sp: spm.SentencePieceProcessor,
|
|
batch: dict,
|
|
word_table: Optional[k2.SymbolTable] = None,
|
|
decoding_graph: Optional[k2.Fsa] = None,
|
|
) -> Dict[str, List[List[str]]]:
|
|
"""Decode one batch and return the result in a dict. The dict has the
|
|
following format:
|
|
|
|
- key: It indicates the setting used for decoding. For example,
|
|
if greedy_search is used, it would be "greedy_search"
|
|
If beam search with a beam size of 7 is used, it would be
|
|
"beam_7"
|
|
- value: It contains the decoding result. `len(value)` equals to
|
|
batch size. `value[i]` is the decoding result for the i-th
|
|
utterance in the given batch.
|
|
Args:
|
|
params:
|
|
It's the return value of :func:`get_params`.
|
|
model:
|
|
The neural model.
|
|
sp:
|
|
The BPE model.
|
|
batch:
|
|
It is the return value from iterating
|
|
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
|
|
for the format of the `batch`.
|
|
word_table:
|
|
The word symbol table.
|
|
decoding_graph:
|
|
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
|
|
only when --decoding_method is fast_beam_search, fast_beam_search_LG,
|
|
fast_beam_search_nbest,
|
|
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
|
|
Returns:
|
|
Return the decoding result. See above description for the format of
|
|
the returned dict.
|
|
"""
|
|
device = next(model.parameters()).device
|
|
feature = batch["inputs"]
|
|
assert feature.ndim == 3
|
|
|
|
feature = feature.to(device)
|
|
# at entry, feature is (N, T, C)
|
|
|
|
supervisions = batch["supervisions"]
|
|
feature_lens = supervisions["num_frames"].to(device)
|
|
|
|
encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
|
|
|
|
hyps = []
|
|
|
|
if (
|
|
params.decoding_method == "fast_beam_search"
|
|
or params.decoding_method == "fast_beam_search_LG"
|
|
):
|
|
hyp_tokens = fast_beam_search_one_best(
|
|
model=model,
|
|
decoding_graph=decoding_graph,
|
|
encoder_out=encoder_out,
|
|
encoder_out_lens=encoder_out_lens,
|
|
beam=params.beam,
|
|
max_contexts=params.max_contexts,
|
|
max_states=params.max_states,
|
|
)
|
|
|
|
if params.decoding_method == "fast_beam_search":
|
|
for hyp in sp.decode(hyp_tokens):
|
|
hyps.append(hyp.split())
|
|
else:
|
|
for hyp in hyp_tokens:
|
|
hyps.append([word_table[i] for i in hyp])
|
|
elif params.decoding_method == "fast_beam_search_nbest_LG":
|
|
hyp_tokens = fast_beam_search_nbest_LG(
|
|
model=model,
|
|
decoding_graph=decoding_graph,
|
|
encoder_out=encoder_out,
|
|
encoder_out_lens=encoder_out_lens,
|
|
beam=params.beam,
|
|
max_contexts=params.max_contexts,
|
|
max_states=params.max_states,
|
|
num_paths=params.num_paths,
|
|
nbest_scale=params.nbest_scale,
|
|
)
|
|
for hyp in hyp_tokens:
|
|
hyps.append([word_table[i] for i in hyp])
|
|
elif params.decoding_method == "fast_beam_search_nbest":
|
|
hyp_tokens = fast_beam_search_nbest(
|
|
model=model,
|
|
decoding_graph=decoding_graph,
|
|
encoder_out=encoder_out,
|
|
encoder_out_lens=encoder_out_lens,
|
|
beam=params.beam,
|
|
max_contexts=params.max_contexts,
|
|
max_states=params.max_states,
|
|
num_paths=params.num_paths,
|
|
nbest_scale=params.nbest_scale,
|
|
)
|
|
for hyp in sp.decode(hyp_tokens):
|
|
hyps.append(hyp.split())
|
|
elif params.decoding_method == "fast_beam_search_nbest_oracle":
|
|
hyp_tokens = fast_beam_search_nbest_oracle(
|
|
model=model,
|
|
decoding_graph=decoding_graph,
|
|
encoder_out=encoder_out,
|
|
encoder_out_lens=encoder_out_lens,
|
|
beam=params.beam,
|
|
max_contexts=params.max_contexts,
|
|
max_states=params.max_states,
|
|
num_paths=params.num_paths,
|
|
ref_texts=sp.encode(supervisions["text"]),
|
|
nbest_scale=params.nbest_scale,
|
|
)
|
|
for hyp in sp.decode(hyp_tokens):
|
|
hyps.append(hyp.split())
|
|
elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
|
|
hyp_tokens = greedy_search_batch(
|
|
model=model,
|
|
encoder_out=encoder_out,
|
|
encoder_out_lens=encoder_out_lens,
|
|
)
|
|
for hyp in sp.decode(hyp_tokens):
|
|
hyps.append(hyp.split())
|
|
elif params.decoding_method == "modified_beam_search":
|
|
hyp_tokens = modified_beam_search(
|
|
model=model,
|
|
encoder_out=encoder_out,
|
|
encoder_out_lens=encoder_out_lens,
|
|
beam=params.beam_size,
|
|
)
|
|
for hyp in sp.decode(hyp_tokens):
|
|
hyps.append(hyp.split())
|
|
else:
|
|
batch_size = encoder_out.size(0)
|
|
|
|
for i in range(batch_size):
|
|
# fmt: off
|
|
encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
|
|
# fmt: on
|
|
if params.decoding_method == "greedy_search":
|
|
hyp = greedy_search(
|
|
model=model,
|
|
encoder_out=encoder_out_i,
|
|
max_sym_per_frame=params.max_sym_per_frame,
|
|
)
|
|
elif params.decoding_method == "beam_search":
|
|
hyp = beam_search(
|
|
model=model,
|
|
encoder_out=encoder_out_i,
|
|
beam=params.beam_size,
|
|
)
|
|
else:
|
|
raise ValueError(
|
|
f"Unsupported decoding method: {params.decoding_method}"
|
|
)
|
|
hyps.append(sp.decode(hyp).split())
|
|
|
|
if params.decoding_method == "greedy_search":
|
|
return {"greedy_search": hyps}
|
|
elif "fast_beam_search" in params.decoding_method:
|
|
key = f"beam_{params.beam}_"
|
|
key += f"max_contexts_{params.max_contexts}_"
|
|
key += f"max_states_{params.max_states}"
|
|
if "nbest" in params.decoding_method:
|
|
key += f"_num_paths_{params.num_paths}_"
|
|
key += f"nbest_scale_{params.nbest_scale}"
|
|
if "LG" in params.decoding_method:
|
|
key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
|
|
|
|
return {key: hyps}
|
|
else:
|
|
return {f"beam_size_{params.beam_size}": hyps}
|
|
|
|
|
|
def decode_dataset(
|
|
dl: torch.utils.data.DataLoader,
|
|
params: AttributeDict,
|
|
model: nn.Module,
|
|
sp: spm.SentencePieceProcessor,
|
|
word_table: Optional[k2.SymbolTable] = None,
|
|
decoding_graph: Optional[k2.Fsa] = None,
|
|
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
|
|
"""Decode dataset.
|
|
|
|
Args:
|
|
dl:
|
|
PyTorch's dataloader containing the dataset to decode.
|
|
params:
|
|
It is returned by :func:`get_params`.
|
|
model:
|
|
The neural model.
|
|
sp:
|
|
The BPE model.
|
|
word_table:
|
|
The word symbol table.
|
|
decoding_graph:
|
|
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
|
|
only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
|
|
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
|
|
Returns:
|
|
Return a dict, whose key may be "greedy_search" if greedy search
|
|
is used, or it may be "beam_7" if beam size of 7 is used.
|
|
Its value is a list of tuples. Each tuple contains two elements:
|
|
The first is the reference transcript, and the second is the
|
|
predicted result.
|
|
"""
|
|
num_cuts = 0
|
|
|
|
try:
|
|
num_batches = len(dl)
|
|
except TypeError:
|
|
num_batches = "?"
|
|
|
|
if params.decoding_method == "greedy_search":
|
|
log_interval = 50
|
|
else:
|
|
log_interval = 20
|
|
|
|
results = defaultdict(list)
|
|
for batch_idx, batch in enumerate(dl):
|
|
texts = batch["supervisions"]["text"]
|
|
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
|
|
|
|
hyps_dict = decode_one_batch(
|
|
params=params,
|
|
model=model,
|
|
sp=sp,
|
|
decoding_graph=decoding_graph,
|
|
word_table=word_table,
|
|
batch=batch,
|
|
)
|
|
|
|
for name, hyps in hyps_dict.items():
|
|
this_batch = []
|
|
assert len(hyps) == len(texts)
|
|
for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
|
|
ref_words = ref_text.split()
|
|
this_batch.append((cut_id, ref_words, hyp_words))
|
|
|
|
results[name].extend(this_batch)
|
|
|
|
num_cuts += len(texts)
|
|
|
|
if batch_idx % log_interval == 0:
|
|
batch_str = f"{batch_idx}/{num_batches}"
|
|
|
|
logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
|
|
return results
|
|
|
|
|
|
def save_results(
|
|
params: AttributeDict,
|
|
test_set_name: str,
|
|
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
|
|
):
|
|
test_set_wers = dict()
|
|
for key, results in results_dict.items():
|
|
recog_path = (
|
|
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
|
|
)
|
|
results = sorted(results)
|
|
store_transcripts(filename=recog_path, texts=results)
|
|
logging.info(f"The transcripts are stored in {recog_path}")
|
|
|
|
# The following prints out WERs, per-word error statistics and aligned
|
|
# ref/hyp pairs.
|
|
errs_filename = (
|
|
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
|
|
)
|
|
with open(errs_filename, "w") as f:
|
|
wer = write_error_stats(
|
|
f, f"{test_set_name}-{key}", results, enable_log=True
|
|
)
|
|
test_set_wers[key] = wer
|
|
|
|
logging.info("Wrote detailed error stats to {}".format(errs_filename))
|
|
|
|
test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
|
|
errs_info = params.res_dir / f"{wer}-{test_set_name}-{key}-{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)
|
|
note = "\tbest for {}".format(test_set_name)
|
|
for key, val in test_set_wers:
|
|
s += "{}\t{}{}\n".format(key, val, note)
|
|
note = ""
|
|
logging.info(s)
|
|
|
|
|
|
@torch.no_grad()
|
|
def main():
|
|
parser = get_parser()
|
|
LibriSpeechAsrDataModule.add_arguments(parser)
|
|
args = parser.parse_args()
|
|
args.exp_dir = Path(args.exp_dir)
|
|
|
|
params = get_params()
|
|
params.update(vars(args))
|
|
|
|
assert params.decoding_method in (
|
|
"greedy_search",
|
|
"beam_search",
|
|
"fast_beam_search",
|
|
"fast_beam_search_LG",
|
|
"fast_beam_search_nbest",
|
|
"fast_beam_search_nbest_LG",
|
|
"fast_beam_search_nbest_oracle",
|
|
"modified_beam_search",
|
|
)
|
|
if "lang_phone" in str(params.lang_dir):
|
|
assert params.decoding_method in (
|
|
"fast_beam_search_LG",
|
|
"fast_beam_search_nbest_LG",
|
|
), "For phone lexicon, use a decoding method with LG."
|
|
|
|
params.res_dir = params.exp_dir / params.decoding_method
|
|
|
|
if params.iter > 0:
|
|
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
|
|
else:
|
|
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
|
|
|
|
if "fast_beam_search" in params.decoding_method:
|
|
params.suffix += f"-beam-{params.beam}"
|
|
params.suffix += f"-max-contexts-{params.max_contexts}"
|
|
params.suffix += f"-max-states-{params.max_states}"
|
|
if "nbest" in params.decoding_method:
|
|
params.suffix += f"-nbest-scale-{params.nbest_scale}"
|
|
params.suffix += f"-num-paths-{params.num_paths}"
|
|
if "LG" in params.decoding_method:
|
|
params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
|
|
elif "beam_search" in params.decoding_method:
|
|
params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
|
|
else:
|
|
params.suffix += f"-context-{params.context_size}"
|
|
params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
|
|
|
|
if params.use_averaged_model:
|
|
params.suffix += "-uam"
|
|
|
|
setup_logger(f"{params.res_dir}/log-{params.suffix}")
|
|
logging.info("Decoding started")
|
|
|
|
device = torch.device("cpu")
|
|
if torch.cuda.is_available():
|
|
device = torch.device("cuda", 0)
|
|
|
|
logging.info(f"Device: {device}")
|
|
|
|
lexicon = Lexicon(params.lang_dir)
|
|
if "lang_bpe" in str(params.lang_dir):
|
|
sp = spm.SentencePieceProcessor()
|
|
sp.load(str(params.lang_dir / "bpe.model"))
|
|
|
|
# <blk> and <unk> are defined in local/train_bpe_model.py
|
|
params.blank_id = sp.piece_to_id("<blk>")
|
|
params.unk_id = sp.piece_to_id("<unk>")
|
|
params.vocab_size = sp.get_piece_size()
|
|
else:
|
|
params.blank_id = lexicon.token_table.get("<eps>")
|
|
params.unk_id = lexicon.token_table.get("SPN")
|
|
params.vocab_size = max(lexicon.tokens) + 1
|
|
sp = None
|
|
|
|
logging.info(pprint.pformat(params, indent=2))
|
|
|
|
logging.info("About to create model")
|
|
model = get_transducer_model(params)
|
|
|
|
if not params.use_averaged_model:
|
|
if params.iter > 0:
|
|
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
|
: params.avg
|
|
]
|
|
if len(filenames) == 0:
|
|
raise ValueError(
|
|
f"No checkpoints found for"
|
|
f" --iter {params.iter}, --avg {params.avg}"
|
|
)
|
|
elif len(filenames) < params.avg:
|
|
raise ValueError(
|
|
f"Not enough checkpoints ({len(filenames)}) found for"
|
|
f" --iter {params.iter}, --avg {params.avg}"
|
|
)
|
|
logging.info(f"averaging {filenames}")
|
|
model.to(device)
|
|
model.load_state_dict(average_checkpoints(filenames, device=device))
|
|
elif params.avg == 1:
|
|
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
|
|
else:
|
|
start = params.epoch - params.avg + 1
|
|
filenames = []
|
|
for i in range(start, params.epoch + 1):
|
|
if i >= 1:
|
|
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
|
|
logging.info(f"averaging {filenames}")
|
|
model.to(device)
|
|
model.load_state_dict(average_checkpoints(filenames, device=device))
|
|
else:
|
|
if params.iter > 0:
|
|
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
|
: params.avg + 1
|
|
]
|
|
if len(filenames) == 0:
|
|
raise ValueError(
|
|
f"No checkpoints found for"
|
|
f" --iter {params.iter}, --avg {params.avg}"
|
|
)
|
|
elif len(filenames) < params.avg + 1:
|
|
raise ValueError(
|
|
f"Not enough checkpoints ({len(filenames)}) found for"
|
|
f" --iter {params.iter}, --avg {params.avg}"
|
|
)
|
|
filename_start = filenames[-1]
|
|
filename_end = filenames[0]
|
|
logging.info(
|
|
"Calculating the averaged model over iteration checkpoints"
|
|
f" from {filename_start} (excluded) to {filename_end}"
|
|
)
|
|
model.to(device)
|
|
model.load_state_dict(
|
|
average_checkpoints_with_averaged_model(
|
|
filename_start=filename_start,
|
|
filename_end=filename_end,
|
|
device=device,
|
|
)
|
|
)
|
|
else:
|
|
assert params.avg > 0, params.avg
|
|
start = params.epoch - params.avg
|
|
assert start >= 1, start
|
|
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
|
|
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
|
|
logging.info(
|
|
f"Calculating the averaged model over epoch range from "
|
|
f"{start} (excluded) to {params.epoch}"
|
|
)
|
|
model.to(device)
|
|
model.load_state_dict(
|
|
average_checkpoints_with_averaged_model(
|
|
filename_start=filename_start,
|
|
filename_end=filename_end,
|
|
device=device,
|
|
)
|
|
)
|
|
|
|
model.to(device)
|
|
model.eval()
|
|
|
|
if "fast_beam_search" in params.decoding_method:
|
|
if "LG" in params.decoding_method:
|
|
word_table = lexicon.word_table
|
|
lg_filename = params.lang_dir / "LG.pt"
|
|
logging.info(f"Loading {lg_filename}")
|
|
decoding_graph = k2.Fsa.from_dict(
|
|
torch.load(lg_filename, map_location=device)
|
|
)
|
|
decoding_graph.scores *= params.ngram_lm_scale
|
|
else:
|
|
word_table = None
|
|
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
|
|
else:
|
|
decoding_graph = None
|
|
word_table = None
|
|
|
|
num_param = sum([p.numel() for p in model.parameters()])
|
|
enc_param = sum([p.numel() for p in model.encoder.parameters()])
|
|
dec_param = sum([p.numel() for p in model.decoder.parameters()])
|
|
join_param = sum([p.numel() for p in model.joiner.parameters()])
|
|
logging.info(f"Number of model parameters: {num_param}")
|
|
logging.info(
|
|
f"Parameters for transducer decoding: {enc_param + dec_param + join_param}"
|
|
)
|
|
|
|
# we need cut ids to display recognition results.
|
|
args.return_cuts = True
|
|
librispeech = LibriSpeechAsrDataModule(args)
|
|
|
|
test_clean_cuts = librispeech.test_clean_cuts()
|
|
test_other_cuts = librispeech.test_other_cuts()
|
|
|
|
test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
|
|
test_other_dl = librispeech.test_dataloaders(test_other_cuts)
|
|
|
|
test_sets = ["test-clean", "test-other"]
|
|
test_dl = [test_clean_dl, test_other_dl]
|
|
|
|
for test_set, test_dl in zip(test_sets, test_dl):
|
|
results_dict = decode_dataset(
|
|
dl=test_dl,
|
|
params=params,
|
|
model=model,
|
|
sp=sp,
|
|
word_table=word_table,
|
|
decoding_graph=decoding_graph,
|
|
)
|
|
|
|
save_results(
|
|
params=params,
|
|
test_set_name=test_set,
|
|
results_dict=results_dict,
|
|
)
|
|
|
|
logging.info("Done!")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|