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
icefall/egs/librispeech/ASR/pruned_transducer_stateless7_contextual/decode.py

1272 lines
41 KiB
Python
Executable File
Raw Blame History

#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao,
# Xiaoyu Yang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Usage:
(1) greedy search
./pruned_transducer_stateless7/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7/exp \
--max-duration 600 \
--decoding-method greedy_search
(2) beam search (not recommended)
./pruned_transducer_stateless7/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7/exp \
--max-duration 600 \
--decoding-method beam_search \
--beam-size 4
(3) modified beam search
./pruned_transducer_stateless7/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7/exp \
--max-duration 600 \
--decoding-method modified_beam_search \
--beam-size 4
(4) fast beam search (one best)
./pruned_transducer_stateless7/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7/exp \
--max-duration 600 \
--decoding-method fast_beam_search \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(5) fast beam search (nbest)
./pruned_transducer_stateless7/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7/exp \
--max-duration 600 \
--decoding-method fast_beam_search_nbest \
--beam 20.0 \
--max-contexts 8 \
--max-states 64 \
--num-paths 200 \
--nbest-scale 0.5
(6) fast beam search (nbest oracle WER)
./pruned_transducer_stateless7/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7/exp \
--max-duration 600 \
--decoding-method fast_beam_search_nbest_oracle \
--beam 20.0 \
--max-contexts 8 \
--max-states 64 \
--num-paths 200 \
--nbest-scale 0.5
(7) fast beam search (with LG)
./pruned_transducer_stateless7/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7/exp \
--max-duration 600 \
--decoding-method fast_beam_search_nbest_LG \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(8) modified beam search with RNNLM shallow fusion
./pruned_transducer_stateless5/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless5/exp \
--max-duration 600 \
--decoding-method modified_beam_search_lm_shallow_fusion \
--beam-size 4 \
--lm-type rnn \
--lm-scale 0.3 \
--lm-exp-dir /path/to/LM \
--rnn-lm-epoch 99 \
--rnn-lm-avg 1 \
--rnn-lm-num-layers 3 \
--rnn-lm-tie-weights 1
(9) modified beam search with LM shallow fusion + LODR
./pruned_transducer_stateless5/decode.py \
--epoch 28 \
--avg 15 \
--max-duration 600 \
--exp-dir ./pruned_transducer_stateless5/exp \
--decoding-method modified_beam_search_LODR \
--beam-size 4 \
--lm-type rnn \
--lm-scale 0.4 \
--lm-exp-dir /path/to/LM \
--rnn-lm-epoch 99 \
--rnn-lm-avg 1 \
--rnn-lm-num-layers 3 \
--rnn-lm-tie-weights 1
--tokens-ngram 2 \
--ngram-lm-scale -0.16 \
"""
import argparse
import logging
import math
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Optional, Tuple
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,
modified_beam_search_lm_shallow_fusion,
modified_beam_search_LODR,
modified_beam_search_ngram_rescoring,
)
from train import add_model_arguments, get_params, get_transducer_model
from egs.librispeech.ASR.pruned_transducer_stateless7_context.context_collector import ContextCollector
from egs.librispeech.ASR.pruned_transducer_stateless7_context.context_encoder import ContextEncoder
from egs.librispeech.ASR.pruned_transducer_stateless7_context.context_encoder_lstm import ContextEncoderLSTM
from egs.librispeech.ASR.pruned_transducer_stateless7_context.context_encoder_pretrained import ContextEncoderPretrained
from egs.librispeech.ASR.pruned_transducer_stateless7_context.word_encoder_bert import BertEncoder
from icefall import LmScorer, NgramLm, BiasedNgramLm
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(
"--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=9,
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(
"--exp-dir",
type=str,
default="pruned_transducer_stateless7/exp",
help="The experiment dir",
)
parser.add_argument(
"--bpe-model",
type=str,
default="data/lang_bpe_500/bpe.model",
help="Path to the BPE model",
)
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="greedy_search",
help="""Possible values are:
- greedy_search
- beam_search
- modified_beam_search
- fast_beam_search
- fast_beam_search_nbest
- fast_beam_search_nbest_oracle
- fast_beam_search_nbest_LG
- modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
- modified_beam_search_LODR
If you use fast_beam_search_nbest_LG, you have to specify
`--lang-dir`, which should contain `LG.pt`.
""",
)
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_nbest, fast_beam_search_nbest_LG,
and fast_beam_search_nbest_oracle
""",
)
parser.add_argument(
"--ngram-lm-scale",
type=float,
default=0.01,
help="""
Used only when --decoding_method is 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_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_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=200,
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""",
)
parser.add_argument(
"--simulate-streaming",
type=str2bool,
default=False,
help="""Whether to simulate streaming in decoding, this is a good way to
test a streaming model.
""",
)
parser.add_argument(
"--decode-chunk-size",
type=int,
default=16,
help="The chunk size for decoding (in frames after subsampling)",
)
parser.add_argument(
"--left-context",
type=int,
default=64,
help="left context can be seen during decoding (in frames after subsampling)",
)
parser.add_argument(
"--use-shallow-fusion",
type=str2bool,
default=False,
help="""Use neural network LM for shallow fusion.
If you want to use LODR, you will also need to set this to true
""",
)
parser.add_argument(
"--lm-type",
type=str,
default="rnn",
help="Type of NN lm",
choices=["rnn", "transformer"],
)
parser.add_argument(
"--lm-scale",
type=float,
default=0.3,
help="""The scale of the neural network LM
Used only when `--use-shallow-fusion` is set to True.
""",
)
parser.add_argument(
"--tokens-ngram",
type=int,
default=3,
help="""Token Ngram used for rescoring.
Used only when the decoding method is
modified_beam_search_ngram_rescoring, or LODR
""",
)
parser.add_argument(
"--backoff-id",
type=int,
default=500,
help="""ID of the backoff symbol.
Used only when the decoding method is
modified_beam_search_ngram_rescoring""",
)
parser.add_argument(
"--context-dir",
type=str,
default="data/fbai-speech/is21_deep_bias/",
help="",
)
parser.add_argument(
"--n-distractors",
type=int,
default=100,
help="",
)
parser.add_argument(
"--keep-ratio",
type=float,
default=1.0,
help="",
)
parser.add_argument(
"--no-encoder-biasing",
type=str2bool,
default=False,
help=""".
""",
)
parser.add_argument(
"--no-decoder-biasing",
type=str2bool,
default=False,
help=""".
""",
)
parser.add_argument(
"--no-wfst-lm-biasing",
type=str2bool,
default=True,
help=""".
""",
)
parser.add_argument(
"--is-full-context",
type=str2bool,
default=False,
help="",
)
parser.add_argument(
"--is-predefined",
type=str2bool,
default=False,
help="",
)
parser.add_argument(
"--is-pretrained-context-encoder",
type=str2bool,
default=False,
help="",
)
parser.add_argument(
"--biased-lm-scale",
type=float,
default=0.0,
help="",
)
parser.add_argument(
"--is-reused-context-encoder",
type=str2bool,
default=False,
help="",
)
add_model_arguments(parser)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
context_collector: ContextCollector,
sp: spm.SentencePieceProcessor,
batch: dict,
word_table: Optional[k2.SymbolTable] = None,
decoding_graph: Optional[k2.Fsa] = None,
ngram_lm: Optional[NgramLm] = None,
ngram_lm_scale: float = 1.0,
LM: Optional[LmScorer] = 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_nbest,
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
LM:
A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
set to true.
ngram_lm:
A ngram lm. Used in LODR decoding.
ngram_lm_scale:
The scale of the ngram language model.
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)
if params.simulate_streaming:
feature_lens += params.left_context
feature = torch.nn.functional.pad(
feature,
pad=(0, 0, 0, params.left_context),
value=LOG_EPS,
)
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,
)
else:
encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
model.scratch_space = dict()
model.scratch_space["sp"] = sp
model.scratch_space["biased_lm_scale"] = params.biased_lm_scale
if not params.no_wfst_lm_biasing:
fsa_list, fsa_sizes, num_words_per_utt2 = \
context_collector.get_context_word_wfst(batch)
biased_lm_list = [
BiasedNgramLm(
fst=fsa,
backoff_id=context_collector.backoff_id
) for fsa in fsa_list
]
model.scratch_space["biased_lm_list"] = biased_lm_list
if not model.no_encoder_biasing:
word_list, word_lengths, num_words_per_utt = \
context_collector.get_context_word_list(batch)
word_list = word_list.to(device)
contexts = {
"mode": "get_context_word_list",
"word_list": word_list,
"word_lengths": word_lengths,
"num_words_per_utt": num_words_per_utt,
}
contexts_h, contexts_mask = model.context_encoder.embed_contexts(
contexts,
)
model.scratch_space["contexts_h"] = contexts_h
model.scratch_space["contexts_mask"] = contexts_mask
encoder_biasing_out, attn = model.encoder_biasing_adapter.forward(encoder_out, contexts_h, contexts_mask)
encoder_out = encoder_out + encoder_biasing_out
hyps = []
if params.decoding_method == "fast_beam_search":
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,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
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())
results = modified_beam_search(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam_size,
return_timestamps=True,
)
hyp_tokens = results.hyps
timestamps = results.timestamps
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
hyp_tokens = modified_beam_search_lm_shallow_fusion(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam_size,
sp=sp,
LM=LM,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.decoding_method == "modified_beam_search_LODR":
hyp_tokens = modified_beam_search_LODR(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam_size,
LODR_lm=ngram_lm,
LODR_lm_scale=ngram_lm_scale,
LM=LM,
)
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,
context_collector: ContextCollector,
sp: spm.SentencePieceProcessor,
word_table: Optional[k2.SymbolTable] = None,
decoding_graph: Optional[k2.Fsa] = None,
ngram_lm: Optional[NgramLm] = None,
ngram_lm_scale: float = 1.0,
LM: Optional[LmScorer] = 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.
LM:
A neural network LM, used during shallow fusion
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
device = next(model.parameters()).device
results = defaultdict(list)
for batch_idx, batch in enumerate(dl):
texts = batch["supervisions"]["text"]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
# if "1998-29455-0019-602" in cut_ids:
# logging.info(cut_ids)
# logging.info(cut_ids.index("1998-29455-0019-602"))
# # import pdb; pdb.set_trace()
# else:
# continue
hyps_dict = decode_one_batch(
params=params,
model=model,
context_collector=context_collector,
sp=sp,
decoding_graph=decoding_graph,
word_table=word_table,
batch=batch,
ngram_lm=ngram_lm,
ngram_lm_scale=ngram_lm_scale,
LM=LM,
)
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}-{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}-{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-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)
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)
def rare_word_score(
params: AttributeDict,
test_set_name: str,
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
cuts,
):
from collections import namedtuple
from score import main as score_main
from lhotse import CutSet
logging.info(f"test_set_name: {test_set_name}")
cuts = cuts[0]
cuts = [c for c in cuts]
cuts = CutSet.from_cuts(cuts)
args = namedtuple('A', ['refs', 'hyps', 'lenient'])
if params.n_distractors > 0:
args.refs = params.context_dir / f"ref/{test_set_name}.biasing_{params.n_distractors}.tsv"
else:
args.refs = params.context_dir / f"ref/{test_set_name}.biasing_100.tsv"
args.lenient = True
for key, results in results_dict.items():
print()
logging.info(f"{key}")
args.hyps = dict()
for cut_id, ref, hyp in results:
u_id = cuts[cut_id].supervisions[0].id
hyp = " ".join(hyp)
hyp = hyp.lower()
args.hyps[u_id] = hyp
score_main(args)
print()
@torch.no_grad()
def main():
parser = get_parser()
LibriSpeechAsrDataModule.add_arguments(parser)
LmScorer.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_nbest",
"fast_beam_search_nbest_LG",
"fast_beam_search_nbest_oracle",
"modified_beam_search",
"modified_beam_search_lm_shallow_fusion",
"modified_beam_search_LODR",
)
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 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}"
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 "ngram" in params.decoding_method:
params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
if params.use_shallow_fusion:
if params.lm_type == "rnn":
params.suffix += f"-rnnlm-lm-scale-{params.lm_scale}"
elif params.lm_type == "transformer":
params.suffix += f"-transformer-lm-scale-{params.lm_scale}"
if "LODR" in params.decoding_method:
params.suffix += (
f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
)
if not params.no_wfst_lm_biasing:
params.suffix += f"-wfst-biasing-{params.biased_lm_scale}"
if not params.no_encoder_biasing:
params.suffix += f"-encoder-biasing"
if not params.no_decoder_biasing:
params.suffix += f"-decoder-biasing"
if params.use_averaged_model:
params.suffix += "-use-averaged-model"
# import time
# timestr = time.strftime("%Y%m%d-%H%M%S")
from datetime import datetime
timestr = datetime.utcnow().strftime('%Y%m%d-%H%M%S-%f')[:-3]
params.suffix += f"-{timestr}"
setup_logger(f"{params.res_dir}/log-decode-{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}")
sp = spm.SentencePieceProcessor()
sp.load(params.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()
if params.simulate_streaming:
assert (
params.causal_convolution
), "Decoding in streaming requires causal convolution"
logging.info(params)
logging.info("About to load context collector")
params.context_dir = Path(params.context_dir)
if params.is_pretrained_context_encoder:
# Use pretrained encoder, e.g., BERT
bert_encoder = BertEncoder(device=device)
context_collector = ContextCollector(
path_is21_deep_bias=params.context_dir,
sp=None,
bert_encoder=bert_encoder,
is_predefined=params.is_predefined,
n_distractors=params.n_distractors,
keep_ratio=params.keep_ratio,
is_full_context=params.is_full_context,
backoff_id=params.backoff_id,
)
# bert_encoder.free_up()
else:
context_collector = ContextCollector(
path_is21_deep_bias=params.context_dir,
sp=sp,
bert_encoder=None,
is_predefined=params.is_predefined,
n_distractors=params.n_distractors,
keep_ratio=params.keep_ratio,
is_full_context=params.is_full_context,
backoff_id=params.backoff_id,
)
logging.info("About to create model")
model = get_transducer_model(params)
model.no_encoder_biasing = params.no_encoder_biasing
model.no_decoder_biasing = params.no_decoder_biasing
model.no_wfst_lm_biasing = params.no_wfst_lm_biasing
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()
# only load N-gram LM when needed
if "ngram" in params.decoding_method or "LODR" in params.decoding_method:
lm_filename = f"{params.tokens_ngram}gram.fst.txt"
logging.info(f"lm filename: {lm_filename}")
ngram_lm = NgramLm(
str(params.lang_dir / lm_filename),
backoff_id=params.backoff_id,
is_binary=False,
)
logging.info(f"num states: {ngram_lm.lm.num_states}")
ngram_lm_scale = params.ngram_lm_scale
else:
ngram_lm = None
ngram_lm_scale = None
# only load the neural network LM if doing shallow fusion
if params.use_shallow_fusion:
LM = LmScorer(
lm_type=params.lm_type,
params=params,
device=device,
lm_scale=params.lm_scale,
)
LM.to(device)
LM.eval()
else:
LM = None
if "fast_beam_search" in params.decoding_method:
if params.decoding_method == "fast_beam_search_nbest_LG":
lexicon = Lexicon(params.lang_dir)
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()])
logging.info(f"Number of model parameters: {num_param}")
# we need cut ids to display recognition results.
args.return_cuts = True
librispeech = LibriSpeechAsrDataModule(args)
dev_clean_cuts = librispeech.dev_clean_cuts()
dev_other_cuts = librispeech.dev_other_cuts()
test_clean_cuts = librispeech.test_clean_cuts()
test_other_cuts = librispeech.test_other_cuts()
# from lhotse import CutSet
# test_clean_cuts = [c for c in test_clean_cuts][:500]
# test_other_cuts = [c for c in test_other_cuts][:500]
# # test_other_cuts1 = [c for c in test_other_cuts if c.id == "1998-29455-0019-602"]
# # test_other_cuts = test_other_cuts1 + [c for c in test_other_cuts][1700:1710]
# test_clean_cuts = CutSet.from_cuts(test_clean_cuts)
# test_other_cuts = CutSet.from_cuts(test_other_cuts)
# from lhotse import CutSet
# test_clean_cuts = [c for c in test_clean_cuts if c.id == "1089-134686-0016-2185"] + [c for c in test_clean_cuts][:9]
# # test_clean_cuts = [c for c in test_clean_cuts if c.id == "1089-134686-0016-2185"]
# test_clean_cuts = CutSet.from_cuts(test_clean_cuts)
# test_clean_cuts.describe()
# TODO:
# 7729-102255-0015-210
# 1995-1836-0000-8621995-1836-0000
# import random
# random.seed(10)
dev_clean_dl = librispeech.test_dataloaders(dev_clean_cuts)
dev_other_dl = librispeech.test_dataloaders(dev_other_cuts)
test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
test_other_dl = librispeech.test_dataloaders(test_other_cuts)
# test_sets = ["dev-clean", "dev-other", "test-clean", "test-other"]
# test_dl = [dev_clean_dl, dev_other_dl, test_clean_dl, test_other_dl]
test_sets = ["test-clean", "test-other"]
test_dls = [test_clean_dl, test_other_dl]
# test_sets = ["test-clean"]
# test_dl = [test_clean_dl]
# test_sets = ["test-other"]
# test_dl = [test_other_dl]
for test_set, test_dl in zip(test_sets, test_dls):
results_dict = decode_dataset(
dl=test_dl,
params=params,
model=model,
context_collector=context_collector,
sp=sp,
word_table=word_table,
decoding_graph=decoding_graph,
ngram_lm=ngram_lm,
ngram_lm_scale=ngram_lm_scale,
LM=LM,
)
save_results(
params=params,
test_set_name=test_set,
results_dict=results_dict,
)
rare_word_score(
params=params,
test_set_name=test_set,
results_dict=results_dict,
cuts=test_dl.sampler.cuts,
)
logging.info("Done!")
if __name__ == "__main__":
main()
Reference in New Issue View Git Blame Copy Permalink