mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-26 18:24:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

add decode.py

Browse Source
This commit is contained in:
yaozengwei 2023-01-16 10:21:32 +08:00
parent 2fc4ce2751
commit 8a7c43f3f3
3 changed files with 980 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
egs/librispeech/ASR/zipformer_ctc_attn/attention_decoder.py
View File

@ -152,7 +152,7 @@ class AttentionDecoderModel(nn.Module):
encoder_out_lens: (batch,)
token_ids: A list of token id list.
Return: A tensor of shape (batch,).
Return: A tensor of shape (batch, num_tokens).
"""
ys_in_pad, ys_in_lens, ys_out_pad = self._pre_ys_in_out(
token_ids, encoder_out.device
@ -168,7 +168,7 @@ class AttentionDecoderModel(nn.Module):
ignore_index=self.ignore_id,
reduction="none",
)
nll = nll.view(batch_size, -1).sum(1)
nll = nll.view(batch_size, -1)
return nll

841
egs/librispeech/ASR/zipformer_ctc_attn/decode.py Executable file
View File

@ -0,0 +1,841 @@
#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# Liyong Guo,
# Quandong Wang,
# Zengwei Yao)
#
# 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) ctc-decoding
./zipformer_ctc_attn/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./zipformer_ctc_attn/exp \
--max-duration 600 \
--decoding-method ctc-decoding
(2) 1best
./zipformer_ctc_attn/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./zipformer_ctc_attn/exp \
--max-duration 600 \
--hlg-scale 0.8 \
--decoding-method 1best
(3) nbest
./zipformer_ctc_attn/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./zipformer_ctc_attn/exp \
--max-duration 600 \
--hlg-scale 0.8 \
--decoding-method nbest
(4) nbest-rescoring
./zipformer_ctc_attn/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./zipformer_ctc_attn/exp \
--max-duration 600 \
--hlg-scale 0.8 \
--lm-dir data/lm \
--decoding-method nbest-rescoring
(5) whole-lattice-rescoring
./zipformer_ctc_attn/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./zipformer_ctc_attn/exp \
--max-duration 600 \
--hlg-scale 0.8 \
--lm-dir data/lm \
--decoding-method whole-lattice-rescoring
"""
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 train import add_model_arguments, get_ctc_attention_model, get_params
from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
from icefall.checkpoint import (
average_checkpoints,
average_checkpoints_with_averaged_model,
find_checkpoints,
load_checkpoint,
)
from icefall.decode import (
get_lattice,
nbest_decoding,
nbest_oracle,
one_best_decoding,
rescore_with_attention_decoder2,
rescore_with_n_best_list,
rescore_with_whole_lattice,
)
from icefall.lexicon import Lexicon
from icefall.utils import (
AttributeDict,
get_texts,
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=15,
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="zipformer_ctc_attn/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="ctc-decoding",
help="""Decoding method.
Supported values are:
- (1) 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.
- (2) 1best. Extract the best path from the decoding lattice as the
decoding result.
- (3) nbest. Extract n paths from the decoding lattice; the path
with the highest score is the decoding result.
- (4) 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
the highest score is the decoding result.
- (5) whole-lattice-rescoring. Rescore the decoding lattice with an
n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
is the decoding result.
you have trained an RNN LM using ./rnn_lm/train.py
- (6) nbest-oracle. Its WER is the lower bound of any n-best
rescoring method can achieve. Useful for debugging n-best
rescoring method.
""",
)
parser.add_argument(
"--num-paths",
type=int,
default=100,
help="""Number of paths for n-best based decoding method.
Used only when "method" is one of the following values:
nbest, nbest-rescoring, and nbest-oracle
""",
)
parser.add_argument(
"--nbest-scale",
type=float,
default=0.5,
help="""The scale to be applied to `lattice.scores`.
It's needed if you use any kinds of n-best based rescoring.
Used only when "method" is one of the following values:
nbest, nbest-rescoring, and nbest-oracle
A smaller value results in more unique paths.
""",
)
parser.add_argument(
"--hlg-scale",
type=float,
default=0.8,
help="""The scale to be applied to `hlg.scores`.
""",
)
parser.add_argument(
"--lm-dir",
type=str,
default="data/lm",
help="""The n-gram LM dir.
It should contain either G_4_gram.pt or G_4_gram.fst.txt
""",
)
add_model_arguments(parser)
return parser
def get_decoding_params() -> AttributeDict:
"""Parameters for decoding."""
params = AttributeDict(
{
"frame_shift_ms": 10,
"search_beam": 20,
"output_beam": 8,
"min_active_states": 30,
"max_active_states": 10000,
"use_double_scores": True,
}
)
return params
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
HLG: Optional[k2.Fsa],
H: Optional[k2.Fsa],
bpe_model: Optional[spm.SentencePieceProcessor],
batch: dict,
word_table: k2.SymbolTable,
G: 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 no rescoring is used, the key is the string `no_rescore`.
If LM rescoring is used, the key is the string `lm_scale_xxx`,
where `xxx` is the value of `lm_scale`. An example key is
`lm_scale_0.7`
- 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`.
- params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
- params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
- params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
- params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
rescoring.
model:
The neural model.
HLG:
The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
H:
The ctc topo. Used only when params.decoding_method is ctc-decoding.
bpe_model:
The BPE model. Used only when params.decoding_method is ctc-decoding.
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.
G:
An LM. It is not None when params.decoding_method is "nbest-rescoring"
or "whole-lattice-rescoring". In general, the G in HLG
is a 3-gram LM, while this G is a 4-gram LM.
Returns:
Return the decoding result. See above description for the format of
the returned dict. Note: If it decodes to nothing, then return None.
"""
if HLG is not None:
device = HLG.device
else:
device = H.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(feature, feature_lens)
nnet_output = model.ctc_output(encoder_out)
# nnet_output is (N, T, C)
supervision_segments = torch.stack(
(
supervisions["sequence_idx"],
supervisions["start_frame"] // params.subsampling_factor,
supervisions["num_frames"] // params.subsampling_factor,
),
1,
).to(torch.int32)
if H is None:
assert HLG is not None
decoding_graph = HLG
else:
assert HLG is None
assert bpe_model is not None
decoding_graph = H
lattice = get_lattice(
nnet_output=nnet_output,
decoding_graph=decoding_graph,
supervision_segments=supervision_segments,
search_beam=params.search_beam,
output_beam=params.output_beam,
min_active_states=params.min_active_states,
max_active_states=params.max_active_states,
subsampling_factor=params.subsampling_factor,
)
if params.decoding_method == "ctc-decoding":
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
# Note: `best_path.aux_labels` contains token IDs, not word IDs
# since we are using H, not HLG here.
#
# 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]
key = "ctc-decoding"
return {key: hyps}
if params.decoding_method == "nbest-oracle":
# Note: You can also pass rescored lattices to it.
# We choose the HLG decoded lattice for speed reasons
# as HLG decoding is faster and the oracle WER
# is only slightly worse than that of rescored lattices.
best_path = nbest_oracle(
lattice=lattice,
num_paths=params.num_paths,
ref_texts=supervisions["text"],
word_table=word_table,
nbest_scale=params.nbest_scale,
oov="<UNK>",
)
hyps = get_texts(best_path)
hyps = [[word_table[i] for i in ids] for ids in hyps]
key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa
return {key: hyps}
if params.decoding_method in ["1best", "nbest"]:
if params.decoding_method == "1best":
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
key = "no_rescore"
else:
best_path = nbest_decoding(
lattice=lattice,
num_paths=params.num_paths,
use_double_scores=params.use_double_scores,
nbest_scale=params.nbest_scale,
)
key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa
hyps = get_texts(best_path)
hyps = [[word_table[i] for i in ids] for ids in hyps]
return {key: hyps}
assert params.decoding_method in [
"nbest-rescoring",
"whole-lattice-rescoring",
"attention-decoder",
]
lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
if params.decoding_method == "nbest-rescoring":
best_path_dict = rescore_with_n_best_list(
lattice=lattice,
G=G,
num_paths=params.num_paths,
lm_scale_list=lm_scale_list,
nbest_scale=params.nbest_scale,
)
elif params.decoding_method == "whole-lattice-rescoring":
best_path_dict = rescore_with_whole_lattice(
lattice=lattice,
G_with_epsilon_loops=G,
lm_scale_list=lm_scale_list,
)
elif params.decoding_method == "attention-decoder":
# 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_decoder2(
lattice=rescored_lattice,
num_paths=params.num_paths,
attention_decoder=model.decoder,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
nbest_scale=params.nbest_scale,
)
else:
assert False, f"Unsupported decoding method: {params.decoding_method}"
ans = dict()
if best_path_dict is not None:
for lm_scale_str, best_path in best_path_dict.items():
hyps = get_texts(best_path)
hyps = [[word_table[i] for i in ids] for ids in hyps]
ans[lm_scale_str] = hyps
else:
ans = None
return ans
def decode_dataset(
dl: torch.utils.data.DataLoader,
params: AttributeDict,
model: nn.Module,
HLG: Optional[k2.Fsa],
H: Optional[k2.Fsa],
bpe_model: Optional[spm.SentencePieceProcessor],
word_table: k2.SymbolTable,
G: 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.
HLG:
The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
H:
The ctc topo. Used only when params.decoding_method is ctc-decoding.
bpe_model:
The BPE model. Used only when params.decoding_method is ctc-decoding.
word_table:
It is the word symbol table.
G:
An LM. It is not None when params.decoding_method is "nbest-rescoring"
or "whole-lattice-rescoring". In general, the G in HLG
is a 3-gram LM, while this G is a 4-gram LM.
Returns:
Return a dict, whose key may be "no-rescore" if no LM rescoring
is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?"
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,
HLG=HLG,
H=H,
bpe_model=bpe_model,
batch=batch,
word_table=word_table,
G=G,
)
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 % 100 == 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)
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}-{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)
args.lang_dir = Path(args.lang_dir)
args.lm_dir = Path(args.lm_dir)
params = get_params()
# add decoding params
params.update(get_decoding_params())
params.update(vars(args))
assert params.decoding_method in (
"ctc-decoding",
"1best",
"nbest",
"nbest-rescoring",
"whole-lattice-rescoring",
"nbest-oracle",
"attention-decoder",
)
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.use_averaged_model:
params.suffix += "-use-averaged-model"
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}")
assert "lang_bpe" in str(params.lang_dir), "Currently only supports bpe model."
graph_compiler = BpeCtcTrainingGraphCompiler(
params.lang_dir,
device=device,
sos_token="<sos/eos>",
eos_token="<sos/eos>",
)
# sos_id, eos_id, ignore_id will be used in AttentionDecoderModel
params.sos_id = graph_compiler.sos_id
params.eos_id = graph_compiler.eos_id
# <blk> is defined in local/train_bpe_model.py
params.vocab_size = graph_compiler.sp.get_piece_size()
logging.info(params)
lexicon = Lexicon(params.lang_dir)
max_token_id = max(lexicon.tokens)
if params.decoding_method == "ctc-decoding":
HLG = None
H = k2.ctc_topo(
max_token=max_token_id,
modified=False,
device=device,
)
bpe_model = spm.SentencePieceProcessor()
bpe_model.load(str(params.lang_dir / "bpe.model"))
else:
H = None
bpe_model = None
HLG = k2.Fsa.from_dict(
torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
)
assert HLG.requires_grad is False
HLG.scores *= params.hlg_scale
if not hasattr(HLG, "lm_scores"):
HLG.lm_scores = HLG.scores.clone()
if params.decoding_method in (
"nbest-rescoring",
"whole-lattice-rescoring",
"attention-decoder",
):
if not (params.lm_dir / "G_4_gram.pt").is_file():
logging.info("Loading G_4_gram.fst.txt")
logging.warning("It may take 8 minutes.")
with open(params.lm_dir / "G_4_gram.fst.txt") as f:
first_word_disambig_id = lexicon.word_table["#0"]
G = k2.Fsa.from_openfst(f.read(), acceptor=False)
# G.aux_labels is not needed in later computations, so
# remove it here.
del G.aux_labels
# CAUTION: The following line is crucial.
# Arcs entering the back-off state have label equal to #0.
# We have to change it to 0 here.
G.labels[G.labels >= first_word_disambig_id] = 0
# See https://github.com/k2-fsa/k2/issues/874
# for why we need to set G.properties to None
G.__dict__["_properties"] = None
G = k2.Fsa.from_fsas([G]).to(device)
G = k2.arc_sort(G)
# Save a dummy value so that it can be loaded in C++.
# See https://github.com/pytorch/pytorch/issues/67902
# for why we need to do this.
G.dummy = 1
torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
else:
logging.info("Loading pre-compiled G_4_gram.pt")
d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
G = k2.Fsa.from_dict(d)
if params.decoding_method in ["whole-lattice-rescoring", "attention-decoder"]:
# Add epsilon self-loops to G as we will compose
# it with the whole lattice later
G = k2.add_epsilon_self_loops(G)
G = k2.arc_sort(G)
G = G.to(device)
# G.lm_scores is used to replace HLG.lm_scores during
# LM rescoring.
G.lm_scores = G.scores.clone()
else:
G = None
logging.info("About to create model")
model = get_ctc_attention_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()
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
num_param = sum([p.numel() for p in model.decoder.parameters()])
logging.info(f"Number of parameters in attention decoder: {num_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,
HLG=HLG,
H=H,
bpe_model=bpe_model,
word_table=lexicon.word_table,
G=G,
)
save_results(
params=params,
test_set_name=test_set,
results_dict=results_dict,
)
logging.info("Done!")
if __name__ == "__main__":
main()

137
icefall/decode.py
View File

@ -1083,6 +1083,143 @@ def rescore_with_attention_decoder(
return ans
def rescore_with_attention_decoder2(
lattice: k2.Fsa,
num_paths: int,
attention_decoder: torch.nn.Module,
encoder_out: torch.Tensor,
encoder_out_lens: torch.Tensor,
nbest_scale: float = 1.0,
ngram_lm_scale: Optional[float] = None,
attention_scale: Optional[float] = None,
use_double_scores: bool = True,
) -> Dict[str, k2.Fsa]:
"""This function extracts `num_paths` paths from the given lattice and uses
an attention decoder to rescore them. The path with the highest score is
the decoding output.
Args:
lattice:
An FsaVec with axes [utt][state][arc].
num_paths:
Number of paths to extract from the given lattice for rescoring.
attention_decoder:
A transformer model. See the class "Transformer" in
conformer_ctc/transformer.py for its interface.
memory:
The encoder memory of the given model. It is the output of
the last torch.nn.TransformerEncoder layer in the given model.
Its shape is `(T, N, C)`.
nbest_scale:
It's the scale applied to `lattice.scores`. A smaller value
leads to more unique paths at the risk of missing the correct path.
ngram_lm_scale:
Optional. It specifies the scale for n-gram LM scores.
attention_scale:
Optional. It specifies the scale for attention decoder scores.
Returns:
A dict of FsaVec, whose key contains a string
ngram_lm_scale_attention_scale and the value is the
best decoding path for each utterance in the lattice.
"""
max_loop_count = 10
loop_count = 0
while loop_count <= max_loop_count:
try:
nbest = Nbest.from_lattice(
lattice=lattice,
num_paths=num_paths,
use_double_scores=use_double_scores,
nbest_scale=nbest_scale,
)
# nbest.fsa.scores are all 0s at this point
nbest = nbest.intersect(lattice)
break
except RuntimeError as e:
logging.info(f"Caught exception:\n{e}\n")
logging.info(f"num_paths before decreasing: {num_paths}")
num_paths = int(num_paths / 2)
if loop_count >= max_loop_count or num_paths <= 0:
logging.info("Return None as the resulting lattice is too large.")
return None
logging.info(
"This OOM is not an error. You can ignore it. "
"If your model does not converge well, or --max-duration "
"is too large, or the input sound file is difficult to "
"decode, you will meet this exception."
)
logging.info(f"num_paths after decreasing: {num_paths}")
loop_count += 1
# Now nbest.fsa has its scores set.
# Also, nbest.fsa inherits the attributes from `lattice`.
assert hasattr(nbest.fsa, "lm_scores")
am_scores = nbest.compute_am_scores()
ngram_lm_scores = nbest.compute_lm_scores()
# The `tokens` attribute is set inside `compile_hlg.py`
assert hasattr(nbest.fsa, "tokens")
assert isinstance(nbest.fsa.tokens, torch.Tensor)
path_to_utt_map = nbest.shape.row_ids(1).to(torch.long)
# the shape of memory is (T, N, C), so we use axis=1 here
expanded_encoder_out = encoder_out.index_select(0, path_to_utt_map)
expanded_encoder_out_lens = encoder_out_lens.index_select(0, path_to_utt_map)
# remove axis corresponding to states.
tokens_shape = nbest.fsa.arcs.shape().remove_axis(1)
tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.tokens)
tokens = tokens.remove_values_leq(0)
token_ids = tokens.tolist()
if len(token_ids) == 0:
print("Warning: rescore_with_attention_decoder(): empty token-ids")
return None
nll = attention_decoder.nll(
encoder_out=expanded_encoder_out,
encoder_out_lens=expanded_encoder_out_lens,
token_ids=token_ids,
)
assert nll.ndim == 2
assert nll.shape[0] == len(token_ids)
attention_scores = -nll.sum(dim=1)
if ngram_lm_scale is None:
ngram_lm_scale_list = [0.01, 0.05, 0.08]
ngram_lm_scale_list += [0.1, 0.3, 0.5, 0.6, 0.7, 0.9, 1.0]
ngram_lm_scale_list += [1.1, 1.2, 1.3, 1.5, 1.7, 1.9, 2.0]
ngram_lm_scale_list += [2.1, 2.2, 2.3, 2.5, 3.0, 4.0, 5.0]
else:
ngram_lm_scale_list = [ngram_lm_scale]
if attention_scale is None:
attention_scale_list = [0.01, 0.05, 0.08]
attention_scale_list += [0.1, 0.3, 0.5, 0.6, 0.7, 0.9, 1.0]
attention_scale_list += [1.1, 1.2, 1.3, 1.5, 1.7, 1.9, 2.0]
attention_scale_list += [2.1, 2.2, 2.3, 2.5, 3.0, 4.0, 5.0]
else:
attention_scale_list = [attention_scale]
ans = dict()
for n_scale in ngram_lm_scale_list:
for a_scale in attention_scale_list:
tot_scores = (
am_scores.values
+ n_scale * ngram_lm_scores.values
+ a_scale * attention_scores
)
ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
max_indexes = ragged_tot_scores.argmax()
best_path = k2.index_fsa(nbest.fsa, max_indexes)
key = f"ngram_lm_scale_{n_scale}_attention_scale_{a_scale}"
ans[key] = best_path
return ans
def rescore_with_rnn_lm(
lattice: k2.Fsa,
num_paths: int,