mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-10 10:32:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

remove unneeded and unsupported files

Browse Source
This commit is contained in:
Piotr Żelasko 2022-01-21 19:23:49 +00:00
parent cb329d1342
commit ec8fa55bcf
4 changed files with 0 additions and 1009 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

75
egs/fisher_swbd/ASR/conformer_ctc/README.md
View File

@ -1,75 +0,0 @@
## Introduction
Please visit
<https://icefall.readthedocs.io/en/latest/recipes/librispeech/conformer_ctc.html>
for how to run this recipe.
## How to compute framewise alignment information
### Step 1: Train a model
Please use `conformer_ctc/train.py` to train a model.
See <https://icefall.readthedocs.io/en/latest/recipes/librispeech/conformer_ctc.html>
for how to do it.
### Step 2: Compute framewise alignment
Run
```
# Choose a checkpoint and determine the number of checkpoints to average
epoch=30
avg=15
./conformer_ctc/ali.py \
--epoch $epoch \
--avg $avg \
--max-duration 500 \
--bucketing-sampler 0 \
--full-libri 1 \
--exp-dir conformer_ctc/exp \
--lang-dir data/lang_bpe_500 \
--ali-dir data/ali_500
```
and you will get four files inside the folder `data/ali_500`:
```
$ ls -lh data/ali_500
total 546M
-rw-r--r-- 1 kuangfangjun root 1.1M Sep 28 08:06 test_clean.pt
-rw-r--r-- 1 kuangfangjun root 1.1M Sep 28 08:07 test_other.pt
-rw-r--r-- 1 kuangfangjun root 542M Sep 28 11:36 train-960.pt
-rw-r--r-- 1 kuangfangjun root 2.1M Sep 28 11:38 valid.pt
```
**Note**: It can take more than 3 hours to compute the alignment
for the training dataset, which contains 960 * 3 = 2880 hours of data.
**Caution**: The model parameters in `conformer_ctc/ali.py` have to match those
in `conformer_ctc/train.py`.
**Caution**: You have to set the parameter `preserve_id` to `True` for `CutMix`.
Search `./conformer_ctc/asr_datamodule.py` for `preserve_id`.
### Step 3: Check your extracted alignments
There is a file `test_ali.py` in `icefall/test` that can be used to test your
alignments. It uses pre-computed alignments to modify a randomly generated
`nnet_output` and it checks that we can decode the correct transcripts
from the resulting `nnet_output`.
You should get something like the following if you run that script:
```
$ ./test/test_ali.py
['THE GOOD NATURED AUDIENCE IN PITY TO FALLEN MAJESTY SHOWED FOR ONCE GREATER DEFERENCE TO THE KING THAN TO THE MINISTER AND SUNG THE PSALM WHICH THE FORMER HAD CALLED FOR', 'THE OLD SERVANT TOLD HIM QUIETLY AS THEY CREPT BACK TO DWELL THAT THIS PASSAGE THAT LED FROM THE HUT IN THE PLEASANCE TO SHERWOOD AND THAT GEOFFREY FOR THE TIME WAS HIDING WITH THE OUTLAWS IN THE FOREST', 'FOR A WHILE SHE LAY IN HER CHAIR IN HAPPY DREAMY PLEASURE AT SUN AND BIRD AND TREE', "BUT THE ESSENCE OF LUTHER'S LECTURES IS THERE"]
['THE GOOD NATURED AUDIENCE IN PITY TO FALLEN MAJESTY SHOWED FOR ONCE GREATER DEFERENCE TO THE KING THAN TO THE MINISTER AND SUNG THE PSALM WHICH THE FORMER HAD CALLED FOR', 'THE OLD SERVANT TOLD HIM QUIETLY AS THEY CREPT BACK TO GAMEWELL THAT THIS PASSAGE WAY LED FROM THE HUT IN THE PLEASANCE TO SHERWOOD AND THAT GEOFFREY FOR THE TIME WAS HIDING WITH THE OUTLAWS IN THE FOREST', 'FOR A WHILE SHE LAY IN HER CHAIR IN HAPPY DREAMY PLEASURE AT SUN AND BIRD AND TREE', "BUT THE ESSENCE OF LUTHER'S LECTURES IS THERE"]
```
### Step 4: Use your alignments in training
Please refer to `conformer_mmi/train.py` for usage. Some useful
functions are:
- `load_alignments()`, it loads alignment saved by `conformer_ctc/ali.py`
- `convert_alignments_to_tensor()`, it converts alignments to PyTorch tensors
- `lookup_alignments()`, it returns the alignments of utterances by giving the cut ID of the utterances.

399
egs/fisher_swbd/ASR/conformer_ctc/ali.py
View File

@ -1,399 +0,0 @@
#!/usr/bin/env python3
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang)
#
# 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:
./conformer_ctc/ali.py \
--exp-dir ./conformer_ctc/exp \
--lang-dir ./data/lang_bpe_500 \
--epoch 20 \
--avg 10 \
--max-duration 300 \
--dataset train-clean-100 \
--out-dir data/ali
"""
import argparse
import logging
from pathlib import Path
import k2
import numpy as np
import torch
from asr_datamodule import LibriSpeechAsrDataModule
from conformer import Conformer
from lhotse import CutSet
from lhotse.features.io import FeaturesWriter, NumpyHdf5Writer
from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
from icefall.checkpoint import average_checkpoints, load_checkpoint
from icefall.decode import one_best_decoding
from icefall.env import get_env_info
from icefall.lexicon import Lexicon
from icefall.utils import (
AttributeDict,
encode_supervisions,
get_alignments,
setup_logger,
)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=34,
help="It specifies the checkpoint to use for decoding."
"Note: Epoch counts from 0.",
)
parser.add_argument(
"--avg",
type=int,
default=20,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch'. ",
)
parser.add_argument(
"--lang-dir",
type=str,
default="data/lang_bpe_500",
help="The lang dir",
)
parser.add_argument(
"--exp-dir",
type=str,
default="conformer_ctc/exp",
help="The experiment dir",
)
parser.add_argument(
"--out-dir",
type=str,
required=True,
help="""Output directory.
It contains 3 generated files:
- labels_xxx.h5
- aux_labels_xxx.h5
- cuts_xxx.json.gz
where xxx is the value of `--dataset`. For instance, if
`--dataset` is `train-clean-100`, it will contain 3 files:
- `labels_train-clean-100.h5`
- `aux_labels_train-clean-100.h5`
- `cuts_train-clean-100.json.gz`
Note: Both labels_xxx.h5 and aux_labels_xxx.h5 contain framewise
alignment. The difference is that labels_xxx.h5 contains repeats.
""",
)
parser.add_argument(
"--dataset",
type=str,
required=True,
help="""The name of the dataset to compute alignments for.
Possible values are:
- test-clean.
- test-other
- train-clean-100
- train-clean-360
- train-other-500
- dev-clean
- dev-other
""",
)
return parser
def get_params() -> AttributeDict:
params = AttributeDict(
{
"lm_dir": Path("data/lm"),
"feature_dim": 80,
"nhead": 8,
"attention_dim": 512,
"subsampling_factor": 4,
# Set it to 0 since attention decoder
# is not used for computing alignments
"num_decoder_layers": 0,
"vgg_frontend": False,
"use_feat_batchnorm": True,
"output_beam": 10,
"use_double_scores": True,
"env_info": get_env_info(),
}
)
return params
def compute_alignments(
model: torch.nn.Module,
dl: torch.utils.data.DataLoader,
labels_writer: FeaturesWriter,
aux_labels_writer: FeaturesWriter,
params: AttributeDict,
graph_compiler: BpeCtcTrainingGraphCompiler,
) -> CutSet:
"""Compute the framewise alignments of a dataset.
Args:
model:
The neural network model.
dl:
Dataloader containing the dataset.
params:
Parameters for computing alignments.
graph_compiler:
It converts token IDs to decoding graphs.
Returns:
Return a CutSet. Each cut has two custom fields: labels_alignment
and aux_labels_alignment, containing framewise alignments information.
Both are of type `lhotse.array.TemporalArray`. The difference between
the two alignments is that `labels_alignment` contain repeats.
"""
try:
num_batches = len(dl)
except TypeError:
num_batches = "?"
num_cuts = 0
device = graph_compiler.device
cuts = []
for batch_idx, batch in enumerate(dl):
feature = batch["inputs"]
# at entry, feature is [N, T, C]
assert feature.ndim == 3
feature = feature.to(device)
supervisions = batch["supervisions"]
cut_list = supervisions["cut"]
for cut in cut_list:
assert len(cut.supervisions) == 1, f"{len(cut.supervisions)}"
nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
# nnet_output is [N, T, C]
supervision_segments, texts = encode_supervisions(
supervisions, subsampling_factor=params.subsampling_factor
)
# we need also to sort cut_ids as encode_supervisions()
# reorders "texts".
# In general, new2old is an identity map since lhotse sorts the returned
# cuts by duration in descending order
new2old = supervision_segments[:, 0].tolist()
cut_list = [cut_list[i] for i in new2old]
token_ids = graph_compiler.texts_to_ids(texts)
decoding_graph = graph_compiler.compile(token_ids)
dense_fsa_vec = k2.DenseFsaVec(
nnet_output,
supervision_segments,
allow_truncate=params.subsampling_factor - 1,
)
lattice = k2.intersect_dense(
decoding_graph,
dense_fsa_vec,
params.output_beam,
)
best_path = one_best_decoding(
lattice=lattice,
use_double_scores=params.use_double_scores,
)
labels_ali = get_alignments(best_path, kind="labels")
aux_labels_ali = get_alignments(best_path, kind="aux_labels")
assert len(labels_ali) == len(aux_labels_ali) == len(cut_list)
for cut, labels, aux_labels in zip(
cut_list, labels_ali, aux_labels_ali
):
cut.labels_alignment = labels_writer.store_array(
key=cut.id,
value=np.asarray(labels, dtype=np.int32),
# frame shift is 0.01s, subsampling_factor is 4
frame_shift=0.04,
temporal_dim=0,
start=0,
)
cut.aux_labels_alignment = aux_labels_writer.store_array(
key=cut.id,
value=np.asarray(aux_labels, dtype=np.int32),
# frame shift is 0.01s, subsampling_factor is 4
frame_shift=0.04,
temporal_dim=0,
start=0,
)
cuts += cut_list
num_cuts += len(cut_list)
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 CutSet.from_cuts(cuts)
@torch.no_grad()
def main():
parser = get_parser()
LibriSpeechAsrDataModule.add_arguments(parser)
args = parser.parse_args()
args.enable_spec_aug = False
args.enable_musan = False
args.return_cuts = True
args.concatenate_cuts = False
params = get_params()
params.update(vars(args))
setup_logger(f"{params.exp_dir}/log-ali")
logging.info(f"Computing alignments for {params.dataset} - started")
logging.info(params)
out_dir = Path(params.out_dir)
out_dir.mkdir(exist_ok=True)
out_labels_ali_filename = out_dir / f"labels_{params.dataset}.h5"
out_aux_labels_ali_filename = out_dir / f"aux_labels_{params.dataset}.h5"
out_manifest_filename = out_dir / f"cuts_{params.dataset}.json.gz"
for f in (
out_labels_ali_filename,
out_aux_labels_ali_filename,
out_manifest_filename,
):
if f.exists():
logging.info(f"{f} exists - skipping")
return
lexicon = Lexicon(params.lang_dir)
max_token_id = max(lexicon.tokens)
num_classes = max_token_id + 1 # +1 for the blank
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"device: {device}")
graph_compiler = BpeCtcTrainingGraphCompiler(
params.lang_dir,
device=device,
sos_token="<sos/eos>",
eos_token="<sos/eos>",
)
logging.info("About to create model")
model = Conformer(
num_features=params.feature_dim,
nhead=params.nhead,
d_model=params.attention_dim,
num_classes=num_classes,
subsampling_factor=params.subsampling_factor,
num_decoder_layers=params.num_decoder_layers,
vgg_frontend=params.vgg_frontend,
use_feat_batchnorm=params.use_feat_batchnorm,
)
model.to(device)
if params.avg == 1:
load_checkpoint(
f"{params.exp_dir}/epoch-{params.epoch}.pt", model, strict=False
)
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if start >= 0:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.load_state_dict(
average_checkpoints(filenames, device=device), strict=False
)
model.eval()
librispeech = LibriSpeechAsrDataModule(args)
if params.dataset == "test-clean":
test_clean_cuts = librispeech.test_clean_cuts()
dl = librispeech.test_dataloaders(test_clean_cuts)
elif params.dataset == "test-other":
test_other_cuts = librispeech.test_other_cuts()
dl = librispeech.test_dataloaders(test_other_cuts)
elif params.dataset == "train-clean-100":
train_clean_100_cuts = librispeech.train_clean_100_cuts()
dl = librispeech.train_dataloaders(train_clean_100_cuts)
elif params.dataset == "train-clean-360":
train_clean_360_cuts = librispeech.train_clean_360_cuts()
dl = librispeech.train_dataloaders(train_clean_360_cuts)
elif params.dataset == "train-other-500":
train_other_500_cuts = librispeech.train_other_500_cuts()
dl = librispeech.train_dataloaders(train_other_500_cuts)
elif params.dataset == "dev-clean":
dev_clean_cuts = librispeech.dev_clean_cuts()
dl = librispeech.valid_dataloaders(dev_clean_cuts)
else:
assert params.dataset == "dev-other", f"{params.dataset}"
dev_other_cuts = librispeech.dev_other_cuts()
dl = librispeech.valid_dataloaders(dev_other_cuts)
logging.info(f"Processing {params.dataset}")
with NumpyHdf5Writer(out_labels_ali_filename) as labels_writer:
with NumpyHdf5Writer(out_aux_labels_ali_filename) as aux_labels_writer:
cut_set = compute_alignments(
model=model,
dl=dl,
labels_writer=labels_writer,
aux_labels_writer=aux_labels_writer,
params=params,
graph_compiler=graph_compiler,
)
cut_set.to_file(out_manifest_filename)
logging.info(
f"For dataset {params.dataset}, its alignments with repeats are "
f"saved to {out_labels_ali_filename}, the alignments without repeats "
f"are saved to {out_aux_labels_ali_filename}, and the cut manifest "
f"file is {out_manifest_filename}. Number of cuts: {len(cut_set)}"
)
torch.set_num_threads(1)
torch.set_num_interop_threads(1)
if __name__ == "__main__":
main()

435
egs/fisher_swbd/ASR/conformer_ctc/pretrained.py
View File

@ -1,435 +0,0 @@
#!/usr/bin/env python3
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang,
# Mingshuang Luo)
#
# 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.
import argparse
import logging
import math
from typing import List
import k2
import kaldifeat
import sentencepiece as spm
import torch
import torchaudio
from conformer import Conformer
from torch.nn.utils.rnn import pad_sequence
from icefall.decode import (
get_lattice,
one_best_decoding,
rescore_with_attention_decoder,
rescore_with_whole_lattice,
)
from icefall.utils import AttributeDict, get_texts
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--checkpoint",
type=str,
required=True,
help="Path to the checkpoint. "
"The checkpoint is assumed to be saved by "
"icefall.checkpoint.save_checkpoint().",
)
parser.add_argument(
"--words-file",
type=str,
help="""Path to words.txt.
Used only when method is not ctc-decoding.
""",
)
parser.add_argument(
"--HLG",
type=str,
help="""Path to HLG.pt.
Used only when method is not ctc-decoding.
""",
)
parser.add_argument(
"--bpe-model",
type=str,
help="""Path to bpe.model.
Used only when method is ctc-decoding.
""",
)
parser.add_argument(
"--method",
type=str,
default="1best",
help="""Decoding method.
Possible values are:
(0) 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.
(1) 1best - Use the best path as decoding output. Only
the transformer encoder output is used for decoding.
We call it HLG decoding.
(2) whole-lattice-rescoring - Use an LM to rescore the
decoding lattice and then use 1best to decode the
rescored lattice.
We call it HLG decoding + n-gram LM rescoring.
(3) attention-decoder - Extract n paths from the rescored
lattice and use the transformer attention decoder for
rescoring.
We call it HLG decoding + n-gram LM rescoring + attention
decoder rescoring.
""",
)
parser.add_argument(
"--G",
type=str,
help="""An LM for rescoring.
Used only when method is
whole-lattice-rescoring or attention-decoder.
It's usually a 4-gram LM.
""",
)
parser.add_argument(
"--num-paths",
type=int,
default=100,
help="""
Used only when method is attention-decoder.
It specifies the size of n-best list.""",
)
parser.add_argument(
"--ngram-lm-scale",
type=float,
default=1.3,
help="""
Used only when method is whole-lattice-rescoring and attention-decoder.
It specifies the scale for n-gram LM scores.
(Note: You need to tune it on a dataset.)
""",
)
parser.add_argument(
"--attention-decoder-scale",
type=float,
default=1.2,
help="""
Used only when method is attention-decoder.
It specifies the scale for attention decoder scores.
(Note: You need to tune it on a dataset.)
""",
)
parser.add_argument(
"--nbest-scale",
type=float,
default=0.5,
help="""
Used only when method is attention-decoder.
It specifies the scale for lattice.scores when
extracting n-best lists. A smaller value results in
more unique number of paths with the risk of missing
the best path.
""",
)
parser.add_argument(
"--sos-id",
type=int,
default=1,
help="""
Used only when method is attention-decoder.
It specifies ID for the SOS token.
""",
)
parser.add_argument(
"--num-classes",
type=int,
default=500,
help="""
Vocab size in the BPE model.
""",
)
parser.add_argument(
"--eos-id",
type=int,
default=1,
help="""
Used only when method is attention-decoder.
It specifies ID for the EOS token.
""",
)
parser.add_argument(
"sound_files",
type=str,
nargs="+",
help="The input sound file(s) to transcribe. "
"Supported formats are those supported by torchaudio.load(). "
"For example, wav and flac are supported. "
"The sample rate has to be 16kHz.",
)
return parser
def get_params() -> AttributeDict:
params = AttributeDict(
{
"sample_rate": 16000,
# parameters for conformer
"subsampling_factor": 4,
"vgg_frontend": False,
"use_feat_batchnorm": True,
"feature_dim": 80,
"nhead": 8,
"attention_dim": 512,
"num_decoder_layers": 6,
# parameters for decoding
"search_beam": 20,
"output_beam": 8,
"min_active_states": 30,
"max_active_states": 10000,
"use_double_scores": True,
}
)
return params
def read_sound_files(
filenames: List[str], expected_sample_rate: float
) -> List[torch.Tensor]:
"""Read a list of sound files into a list 1-D float32 torch tensors.
Args:
filenames:
A list of sound filenames.
expected_sample_rate:
The expected sample rate of the sound files.
Returns:
Return a list of 1-D float32 torch tensors.
"""
ans = []
for f in filenames:
wave, sample_rate = torchaudio.load(f)
assert sample_rate == expected_sample_rate, (
f"expected sample rate: {expected_sample_rate}. "
f"Given: {sample_rate}"
)
# We use only the first channel
ans.append(wave[0])
return ans
def main():
parser = get_parser()
args = parser.parse_args()
params = get_params()
if args.method != "attention-decoder":
# to save memory as the attention decoder
# will not be used
params.num_decoder_layers = 0
params.update(vars(args))
logging.info(f"{params}")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"device: {device}")
logging.info("Creating model")
model = Conformer(
num_features=params.feature_dim,
nhead=params.nhead,
d_model=params.attention_dim,
num_classes=params.num_classes,
subsampling_factor=params.subsampling_factor,
num_decoder_layers=params.num_decoder_layers,
vgg_frontend=params.vgg_frontend,
use_feat_batchnorm=params.use_feat_batchnorm,
)
checkpoint = torch.load(args.checkpoint, map_location="cpu")
model.load_state_dict(checkpoint["model"], strict=False)
model.to(device)
model.eval()
logging.info("Constructing Fbank computer")
opts = kaldifeat.FbankOptions()
opts.device = device
opts.frame_opts.dither = 0
opts.frame_opts.snip_edges = False
opts.frame_opts.samp_freq = params.sample_rate
opts.mel_opts.num_bins = params.feature_dim
fbank = kaldifeat.Fbank(opts)
logging.info(f"Reading sound files: {params.sound_files}")
waves = read_sound_files(
filenames=params.sound_files, expected_sample_rate=params.sample_rate
)
waves = [w.to(device) for w in waves]
logging.info("Decoding started")
features = fbank(waves)
features = pad_sequence(
features, batch_first=True, padding_value=math.log(1e-10)
)
# Note: We don't use key padding mask for attention during decoding
with torch.no_grad():
nnet_output, memory, memory_key_padding_mask = model(features)
batch_size = nnet_output.shape[0]
supervision_segments = torch.tensor(
[[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
dtype=torch.int32,
)
if params.method == "ctc-decoding":
logging.info("Use CTC decoding")
bpe_model = spm.SentencePieceProcessor()
bpe_model.load(params.bpe_model)
max_token_id = params.num_classes - 1
H = k2.ctc_topo(
max_token=max_token_id,
modified=False,
device=device,
)
lattice = get_lattice(
nnet_output=nnet_output,
decoding_graph=H,
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,
)
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
token_ids = get_texts(best_path)
hyps = bpe_model.decode(token_ids)
hyps = [s.split() for s in hyps]
elif params.method in [
"1best",
"whole-lattice-rescoring",
"attention-decoder",
]:
logging.info(f"Loading HLG from {params.HLG}")
HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
HLG = HLG.to(device)
if not hasattr(HLG, "lm_scores"):
# For whole-lattice-rescoring and attention-decoder
HLG.lm_scores = HLG.scores.clone()
if params.method in [
"whole-lattice-rescoring",
"attention-decoder",
]:
logging.info(f"Loading G from {params.G}")
G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
# Add epsilon self-loops to G as we will compose
# it with the whole lattice later
G = G.to(device)
G = k2.add_epsilon_self_loops(G)
G = k2.arc_sort(G)
G.lm_scores = G.scores.clone()
lattice = get_lattice(
nnet_output=nnet_output,
decoding_graph=HLG,
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.method == "1best":
logging.info("Use HLG decoding")
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
elif params.method == "whole-lattice-rescoring":
logging.info("Use HLG decoding + LM rescoring")
best_path_dict = rescore_with_whole_lattice(
lattice=lattice,
G_with_epsilon_loops=G,
lm_scale_list=[params.ngram_lm_scale],
)
best_path = next(iter(best_path_dict.values()))
elif params.method == "attention-decoder":
logging.info("Use HLG + LM rescoring + attention decoder rescoring")
rescored_lattice = rescore_with_whole_lattice(
lattice=lattice, G_with_epsilon_loops=G, lm_scale_list=None
)
best_path_dict = rescore_with_attention_decoder(
lattice=rescored_lattice,
num_paths=params.num_paths,
model=model,
memory=memory,
memory_key_padding_mask=memory_key_padding_mask,
sos_id=params.sos_id,
eos_id=params.eos_id,
nbest_scale=params.nbest_scale,
ngram_lm_scale=params.ngram_lm_scale,
attention_scale=params.attention_decoder_scale,
)
best_path = next(iter(best_path_dict.values()))
hyps = get_texts(best_path)
word_sym_table = k2.SymbolTable.from_file(params.words_file)
hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
else:
raise ValueError(f"Unsupported decoding method: {params.method}")
s = "\n"
for filename, hyp in zip(params.sound_files, hyps):
words = " ".join(hyp)
s += f"{filename}:\n{words}\n\n"
logging.info(s)
logging.info("Decoding Done")
if __name__ == "__main__":
formatter = (
"%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
)
logging.basicConfig(format=formatter, level=logging.INFO)
main()

100
egs/fisher_swbd/ASR/local/generate_unique_lexicon.py
View File

@ -1,100 +0,0 @@
#!/usr/bin/env python3
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang)
#
# 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.
"""
This file takes as input a lexicon.txt and output a new lexicon,
in which each word has a unique pronunciation.
The way to do this is to keep only the first pronunciation of a word
in lexicon.txt.
"""
import argparse
import logging
from pathlib import Path
from typing import List, Tuple
from icefall.lexicon import read_lexicon, write_lexicon
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--lang-dir",
type=str,
help="""Input and output directory.
It should contain a file lexicon.txt.
This file will generate a new file uniq_lexicon.txt
in it.
""",
)
return parser.parse_args()
def filter_multiple_pronunications(
lexicon: List[Tuple[str, List[str]]]
) -> List[Tuple[str, List[str]]]:
"""Remove multiple pronunciations of words from a lexicon.
If a word has more than one pronunciation in the lexicon, only
the first one is kept, while other pronunciations are removed
from the lexicon.
Args:
lexicon:
The input lexicon, containing a list of (word, [p1, p2, ..., pn]),
where "p1, p2, ..., pn" are the pronunciations of the "word".
Returns:
Return a new lexicon where each word has a unique pronunciation.
"""
seen = set()
ans = []
for word, tokens in lexicon:
if word in seen:
continue
seen.add(word)
ans.append((word, tokens))
return ans
def main():
args = get_args()
lang_dir = Path(args.lang_dir)
lexicon_filename = lang_dir / "lexicon.txt"
in_lexicon = read_lexicon(lexicon_filename)
out_lexicon = filter_multiple_pronunications(in_lexicon)
write_lexicon(lang_dir / "uniq_lexicon.txt", out_lexicon)
logging.info(f"Number of entries in lexicon.txt: {len(in_lexicon)}")
logging.info(f"Number of entries in uniq_lexicon.txt: {len(out_lexicon)}")
if __name__ == "__main__":
formatter = (
"%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
)
logging.basicConfig(format=formatter, level=logging.INFO)
main()