mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-04 14:44:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add results.

Browse Source
This commit is contained in:
Fangjun Kuang 2022-04-29 12:03:22 +08:00
parent 187534df2e
commit fc7574f6d2
6 changed files with 578 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
egs/librispeech/ASR/README.md
View File

@ -1,9 +1,4 @@
# Introduction
Please refer to <https://icefall.readthedocs.io/en/latest/recipes/librispeech/index.html>
for how to run models in this recipe.
# Introduction Please refer to <https://icefall.readthedocs.io/en/latest/recipes/librispeech/index.html> for how to run models in this recipe.
# Transducers
There are various folders containing the name `transducer` in this folder.
@ -17,6 +12,7 @@ The following table lists the differences among them.
| `transducer_stateless_multi_datasets` | Conformer | Embedding + Conv1d | Using data from GigaSpeech as extra training data |
| `pruned_transducer_stateless` | Conformer | Embedding + Conv1d | Using k2 pruned RNN-T loss |
| `pruned_transducer_stateless2` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss |
| `pruned_transducer_stateless3` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss + using GigaSpeech as extra training data |
The decoder in `transducer_stateless` is modified from the paper

100
egs/librispeech/ASR/RESULTS.md
View File

@ -1,5 +1,105 @@
## Results
### LibriSpeech BPE training results (Pruned Transducer 3)
[pruned_transducer_stateless3](./pruned_transducer_stateless3)
Same as `Pruned Transducer 2` but using the XL subset from
[GigaSpeech](https://github.com/SpeechColab/GigaSpeech) as extra training data.
During training, it selects either a batch from GigaSpeech with prob `giga_prob`
or a batch from LibriSpeech with prob `1 - giga_prob`. All utterances within
a batch comes from the same dataset.
See <https://github.com/k2-fsa/icefall/pull/312>
The WERs are:
| | test-clean | test-other | comment |
|-------------------------------------|------------|------------|----------------------------------------|
| greedy search (max sym per frame 1) | 2.21 | 5.09 | --epoch 27 --avg 2 --max-duration 600 |
| greedy search (max sym per frame 1) | 2.25 | 5.02 | --epoch 27 --avg 12 --max-duration 600 |
| modified beam search | 2.19 | 5.03 | --epoch 25 --avg 6 --max-duration 600 |
| modified beam search | 2.23 | 4.94 | --epoch 27 --avg 10 --max-duration 600 |
| beam search | 2.16 | 4.95 | --epoch 25 --avg 7 --max-duration 600 |
| fast beam search | 2.21 | 4.96 | --epoch 27 --avg 10 --max-duration 600 |
| fast beam search | 2.19 | 4.97 | --epoch 27 --avg 12 --max-duration 600 |
The training commands are:
```bash
./prepare.sh
./prepare_giga_speech.sh
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
./pruned_transducer_stateless3/train.py \
--world-size 8 \
--num-epochs 30 \
--start-epoch 0 \
--full-libri 1 \
--exp-dir pruned_transducer_stateless3/exp \
--max-duration 300 \
--use-fp16 1 \
--lr-epochs 4 \
--num-workers 2 \
--giga-prob 0.8
```
The tensorboard log can be found at
<https://tensorboard.dev/experiment/gaD34WeYSMCOkzoo3dZXGg/>
(Note: The training process is killed manually at `epoch-28.pt`.)
Pretrained models, training logs, decoding logs, and decoding results
are available at
<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-04-29>
Decoding commands are:
```bash
# greedy search
./pruned_transducer_stateless3/decode.py \
--epoch 27 \
--avg 2 \
--exp-dir ./pruned_transducer_stateless3/exp \
--max-duration 600 \
--decoding-method greedy_search \
--max-sym-per-frame 1
# modified beam search
./pruned_transducer_stateless3/decode.py \
--epoch 25 \
--avg 6 \
--exp-dir ./pruned_transducer_stateless3/exp \
--max-duration 600 \
--decoding-method modified_beam_search \
--max-sym-per-frame 1
# beam search
./pruned_transducer_stateless3/decode.py \
--epoch 25 \
--avg 7 \
--exp-dir ./pruned_transducer_stateless3/exp \
--max-duration 600 \
--decoding-method beam_search \
--max-sym-per-frame 1
# fast beam search
for epoch in 27; do
for avg in 10 12; do
./pruned_transducer_stateless3/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir ./pruned_transducer_stateless3/exp \
--max-duration 600 \
--decoding-method fast_beam_search \
--max-states 32 \
--beam 8
done
done
```
### LibriSpeech BPE training results (Pruned Transducer 2)
[pruned_transducer_stateless2](./pruned_transducer_stateless2)

4
egs/librispeech/ASR/pruned_transducer_stateless3/decode.py
View File

@ -511,6 +511,10 @@ def main():
params.suffix += f"-beam-{params.beam}"
params.suffix += f"-max-contexts-{params.max_contexts}"
params.suffix += f"-max-states-{params.max_states}"
elif params.decoding_method == "fast_beam_search_nbest_oracle":
params.suffix += f"-beam-{params.beam}"
params.suffix += f"-max-contexts-{params.max_contexts}"
params.suffix += f"-max-states-{params.max_states}"
params.suffix += f"-num-paths-{params.num_paths}"
params.suffix += f"-nbest-scale-{params.nbest_scale}"
elif "beam_search" in params.decoding_method:

183
egs/librispeech/ASR/pruned_transducer_stateless3/export.py Executable file
View File

@ -0,0 +1,183 @@
#!/usr/bin/env python3
#
# Copyright 2021 Xiaomi Corporation (Author: 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 script converts several saved checkpoints
# to a single one using model averaging.
"""
Usage:
./pruned_transducer_stateless3/export.py \
--exp-dir ./pruned_transducer_stateless3/exp \
--bpe-model data/lang_bpe_500/bpe.model \
--epoch 20 \
--avg 10
It will generate a file exp_dir/pretrained.pt
To use the generated file with `pruned_transducer_stateless3/decode.py`,
you can do:
cd /path/to/exp_dir
ln -s pretrained.pt epoch-9999.pt
cd /path/to/egs/librispeech/ASR
./pruned_transducer_stateless3/decode.py \
--exp-dir ./pruned_transducer_stateless3/exp \
--epoch 9999 \
--avg 1 \
--max-duration 600 \
--decoding-method greedy_search \
--bpe-model data/lang_bpe_500/bpe.model
"""
import argparse
import logging
from pathlib import Path
import sentencepiece as spm
import torch
from train import get_params, get_transducer_model
from icefall.checkpoint import average_checkpoints, load_checkpoint
from icefall.utils import str2bool
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=28,
help="It specifies the checkpoint to use for decoding."
"Note: Epoch counts from 0.",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch'. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="pruned_transducer_stateless3/exp",
help="""It specifies the directory where all training related
files, e.g., checkpoints, log, etc, are saved
""",
)
parser.add_argument(
"--bpe-model",
type=str,
default="data/lang_bpe_500/bpe.model",
help="Path to the BPE model",
)
parser.add_argument(
"--jit",
type=str2bool,
default=False,
help="""True to save a model after applying torch.jit.script.
""",
)
parser.add_argument(
"--context-size",
type=int,
default=2,
help="The context size in the decoder. 1 means bigram; "
"2 means tri-gram",
)
return parser
def main():
args = get_parser().parse_args()
args.exp_dir = Path(args.exp_dir)
assert args.jit is False, "Support torchscript will be added later"
params = get_params()
params.update(vars(args))
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> is defined in local/train_bpe_model.py
params.blank_id = sp.piece_to_id("<blk>")
params.vocab_size = sp.get_piece_size()
logging.info(params)
logging.info("About to create model")
model = get_transducer_model(params)
model.to(device)
if 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 start >= 0:
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))
model.eval()
model.to("cpu")
model.eval()
if params.jit:
logging.info("Using torch.jit.script")
model = torch.jit.script(model)
filename = params.exp_dir / "cpu_jit.pt"
model.save(str(filename))
logging.info(f"Saved to {filename}")
else:
logging.info("Not using torch.jit.script")
# Save it using a format so that it can be loaded
# by :func:`load_checkpoint`
filename = params.exp_dir / "pretrained.pt"
torch.save({"model": model.state_dict()}, str(filename))
logging.info(f"Saved to {filename}")
if __name__ == "__main__":
formatter = (
"%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
)
logging.basicConfig(format=formatter, level=logging.INFO)
main()

289
egs/librispeech/ASR/pruned_transducer_stateless3/pretrained.py Executable file
View File

@ -0,0 +1,289 @@
#!/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:
(1) greedy search
./pruned_transducer_stateless3/pretrained.py \
--checkpoint ./pruned_transducer_stateless3/exp/pretrained.pt \
--bpe-model ./data/lang_bpe_500/bpe.model \
--method greedy_search \
/path/to/foo.wav \
/path/to/bar.wav \
(1) beam search
./pruned_transducer_stateless3/pretrained.py \
--checkpoint ./pruned_transducer_stateless3/exp/pretrained.pt \
--bpe-model ./data/lang_bpe_500/bpe.model \
--method beam_search \
--beam-size 4 \
/path/to/foo.wav \
/path/to/bar.wav \
You can also use `./pruned_transducer_stateless3/exp/epoch-xx.pt`.
Note: ./pruned_transducer_stateless3/exp/pretrained.pt is generated by
./pruned_transducer_stateless3/export.py
"""
import argparse
import logging
import math
from typing import List
import kaldifeat
import sentencepiece as spm
import torch
import torchaudio
from beam_search import (
beam_search,
greedy_search,
greedy_search_batch,
modified_beam_search,
)
from torch.nn.utils.rnn import pad_sequence
from train import get_params, get_transducer_model
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(
"--bpe-model",
type=str,
help="""Path to bpe.model.
Used only when method is ctc-decoding.
""",
)
parser.add_argument(
"--method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_search
- beam_search
- modified_beam_search
""",
)
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.",
)
parser.add_argument(
"--sample-rate",
type=int,
default=16000,
help="The sample rate of the input sound file",
)
parser.add_argument(
"--beam-size",
type=int,
default=4,
help="Used only when --method is beam_search and modified_beam_search",
)
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
--method is greedy_search.
""",
)
return parser
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
@torch.no_grad()
def main():
parser = get_parser()
args = parser.parse_args()
params = get_params()
params.update(vars(args))
sp = spm.SentencePieceProcessor()
sp.load(params.bpe_model)
# <blk> is 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()
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 = get_transducer_model(params)
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
checkpoint = torch.load(args.checkpoint, map_location="cpu")
model.load_state_dict(checkpoint["model"], strict=False)
model.to(device)
model.eval()
model.device = device
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)
feature_lengths = [f.size(0) for f in features]
features = pad_sequence(
features, batch_first=True, padding_value=math.log(1e-10)
)
feature_lengths = torch.tensor(feature_lengths, device=device)
encoder_out, encoder_out_lens = model.encoder(
x=features, x_lens=feature_lengths
)
num_waves = encoder_out.size(0)
hyps = []
msg = f"Using {params.method}"
if params.method == "beam_search":
msg += f" with beam size {params.beam_size}"
logging.info(msg)
if params.method == "modified_beam_search":
hyp_tokens = modified_beam_search(
model=model,
encoder_out=encoder_out,
beam=params.beam_size,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
hyp_tokens = greedy_search_batch(
model=model,
encoder_out=encoder_out,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
else:
for i in range(num_waves):
# fmt: off
encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
# fmt: on
if params.method == "greedy_search":
hyp = greedy_search(
model=model,
encoder_out=encoder_out_i,
max_sym_per_frame=params.max_sym_per_frame,
)
elif params.method == "beam_search":
hyp = beam_search(
model=model,
encoder_out=encoder_out_i,
beam=params.beam_size,
)
else:
raise ValueError(f"Unsupported method: {params.method}")
hyps.append(sp.decode(hyp).split())
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()

1
egs/librispeech/ASR/pruned_transducer_stateless3/train.py
View File

@ -968,7 +968,6 @@ def run(rank, world_size, args):
train_giga_cuts = gigaspeech.train_S_cuts()
train_giga_cuts = filter_short_and_long_utterances(train_giga_cuts)
train_giga_cuts = train_giga_cuts.repeat(times=None)
if args.enable_musan:
cuts_musan = load_manifest(