mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-12-08 21:45:27 +00:00
Zipformer for Common Voice (#997)
* Add soft links in pruned_transducer_stateless7 for CommonVoice * Add python files * Update prepare.sh * Update normalization * Fix for soft links * Add some docs * Add export * Update egs/commonvoice/ASR/RESULTS.md Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com> * Add export for onnx --------- Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>
This commit is contained in:
Yifan Yang
GitHub
parent
34d1b07c3d
commit
8838fe0bd2
18
egs/commonvoice/ASR/README.md
Normal file
18
egs/commonvoice/ASR/README.md
Normal file
@ -0,0 +1,18 @@
|
||||
# Introduction
|
||||
|
||||
This recipe includes some different ASR models trained with Common Voice
|
||||
|
||||
[./RESULTS.md](./RESULTS.md) contains the latest results.
|
||||
|
||||
# Transducers
|
||||
|
||||
There are various folders containing the name `transducer` in this folder.
|
||||
The following table lists the differences among them.
|
||||
|
||||
| | Encoder | Decoder | Comment |
|
||||
|---------------------------------------|---------------------|--------------------|---------------------------------------------------|
|
||||
| `pruned_transducer_stateless7` | Zipformer | Embedding + Conv1d | First experiment with Zipformer from Dan |
|
||||
|
||||
The decoder in `transducer_stateless` is modified from the paper
|
||||
[RNN-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
|
||||
We place an additional Conv1d layer right after the input embedding layer.
|
||||
59
egs/commonvoice/ASR/RESULTS.md
Normal file
59
egs/commonvoice/ASR/RESULTS.md
Normal file
@ -0,0 +1,59 @@
|
||||
## Results
|
||||
### GigaSpeech BPE training results (Pruned Stateless Transducer 7)
|
||||
|
||||
#### [pruned_transducer_stateless7](./pruned_transducer_stateless7)
|
||||
|
||||
See #997 for more details.
|
||||
|
||||
Number of model parameters: 70369391, i.e., 70.37 M
|
||||
|
||||
The best WER, as of 2023-04-17, for Common Voice English 13.0 (cv-corpus-13.0-2023-03-09/en) is below:
|
||||
|
||||
Results are:
|
||||
|
||||
| | Dev | Test |
|
||||
|----------------------|-------|-------|
|
||||
| greedy search | 9.96 | 12.54 |
|
||||
| modified beam search | 9.86 | 12.48 |
|
||||
|
||||
To reproduce the above result, use the following commands for training:
|
||||
|
||||
```bash
|
||||
export CUDA_VISIBLE_DEVICES="0,1,2,3"
|
||||
./pruned_transducer_stateless7/train.py \
|
||||
--world-size 4 \
|
||||
--num-epochs 30 \
|
||||
--start-epoch 1 \
|
||||
--use-fp16 1 \
|
||||
--exp-dir pruned_transducer_stateless7/exp \
|
||||
--max-duration 550
|
||||
```
|
||||
|
||||
and the following commands for decoding:
|
||||
|
||||
```bash
|
||||
# greedy search
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 30 \
|
||||
--avg 5 \
|
||||
--decoding-method greedy_search \
|
||||
--exp-dir pruned_transducer_stateless7/exp \
|
||||
--bpe-model data/en/lang_bpe_500/bpe.model \
|
||||
--max-duration 600
|
||||
|
||||
# modified beam search
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 30 \
|
||||
--avg 5 \
|
||||
--decoding-method modified_beam_search \
|
||||
--beam-size 4 \
|
||||
--exp-dir pruned_transducer_stateless7/exp \
|
||||
--bpe-model data/en/lang_bpe_500/bpe.model \
|
||||
--max-duration 600
|
||||
```
|
||||
|
||||
Pretrained model is available at
|
||||
<https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17>
|
||||
|
||||
The tensorboard log for training is available at
|
||||
<https://tensorboard.dev/experiment/j4pJQty6RMOkMJtRySREKw/>
|
||||
@ -90,7 +90,7 @@ def compute_fbank_commonvoice_splits(args):
|
||||
subset = "train"
|
||||
num_splits = args.num_splits
|
||||
language = args.language
|
||||
output_dir = f"data/{language}/fbank/{subset}_split_{num_splits}"
|
||||
output_dir = f"data/{language}/fbank/cv-{language}_{subset}_split_{num_splits}"
|
||||
output_dir = Path(output_dir)
|
||||
assert output_dir.exists(), f"{output_dir} does not exist!"
|
||||
|
||||
|
||||
1
egs/commonvoice/ASR/local/prepare_lang_bpe.py
Symbolic link
1
egs/commonvoice/ASR/local/prepare_lang_bpe.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/local/prepare_lang_bpe.py
|
||||
@ -17,6 +17,7 @@
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import re
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
@ -42,6 +43,11 @@ def get_args():
|
||||
return parser.parse_args()
|
||||
|
||||
|
||||
def normalize_text(utt: str) -> str:
|
||||
utt = re.sub(r"[{0}]+".format("-"), " ", utt)
|
||||
return re.sub(r"[^a-zA-Z\s]", "", utt).upper()
|
||||
|
||||
|
||||
def preprocess_commonvoice(
|
||||
language: str,
|
||||
dataset: Optional[str] = None,
|
||||
@ -84,6 +90,17 @@ def preprocess_commonvoice(
|
||||
logging.info(f"{partition} already exists - skipping")
|
||||
continue
|
||||
|
||||
logging.info(f"Normalizing text in {partition}")
|
||||
for sup in m["supervisions"]:
|
||||
text = str(sup.text)
|
||||
orig_text = text
|
||||
sup.text = normalize_text(sup.text)
|
||||
text = str(sup.text)
|
||||
if len(orig_text) != len(text):
|
||||
logging.info(
|
||||
f"\nOriginal text vs normalized text:\n{orig_text}\n{text}"
|
||||
)
|
||||
|
||||
# Create long-recording cut manifests.
|
||||
cut_set = CutSet.from_manifests(
|
||||
recordings=m["recordings"],
|
||||
@ -92,12 +109,6 @@ def preprocess_commonvoice(
|
||||
|
||||
# Run data augmentation that needs to be done in the
|
||||
# time domain.
|
||||
if "train" in partition:
|
||||
logging.info(
|
||||
f"Speed perturb for {partition} with factors 0.9 and 1.1 "
|
||||
"(Perturbing may take 2 minutes and saving may take 7 minutes)"
|
||||
)
|
||||
cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
|
||||
logging.info(f"Saving to {raw_cuts_path}")
|
||||
cut_set.to_file(raw_cuts_path)
|
||||
|
||||
|
||||
1
egs/commonvoice/ASR/local/train_bpe_model.py
Symbolic link
1
egs/commonvoice/ASR/local/train_bpe_model.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/local/train_bpe_model.py
|
||||
1
egs/commonvoice/ASR/local/validate_bpe_lexicon.py
Symbolic link
1
egs/commonvoice/ASR/local/validate_bpe_lexicon.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/local/validate_bpe_lexicon.py
|
||||
@ -126,7 +126,7 @@ fi
|
||||
|
||||
if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
|
||||
log "Stage 5: Split train subset into ${num_splits} pieces"
|
||||
split_dir=data/${lang}/fbank/train_split_${num_splits}
|
||||
split_dir=data/${lang}/fbank/cv-${lang}_train_split_${num_splits}
|
||||
if [ ! -e $split_dir/.cv-${lang}_train_split.done ]; then
|
||||
lhotse split $num_splits ./data/${lang}/fbank/cv-${lang}_cuts_train_raw.jsonl.gz $split_dir
|
||||
touch $split_dir/.cv-${lang}_train_split.done
|
||||
@ -147,10 +147,98 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
|
||||
fi
|
||||
|
||||
if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
|
||||
log "Stage 7: Compute fbank for musan"
|
||||
log "Stage 7: Combine features for train"
|
||||
if [ ! -f data/${lang}/fbank/cv-${lang}_cuts_train.jsonl.gz ]; then
|
||||
pieces=$(find data/${lang}/fbank/cv-${lang}_train_split_${num_splits} -name "cv-${lang}_cuts_train.*.jsonl.gz")
|
||||
lhotse combine $pieces data/${lang}/fbank/cv-${lang}_cuts_train.jsonl.gz
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
|
||||
log "Stage 8: Compute fbank for musan"
|
||||
mkdir -p data/fbank
|
||||
if [ ! -e data/fbank/.musan.done ]; then
|
||||
./local/compute_fbank_musan.py
|
||||
touch data/fbank/.musan.done
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
|
||||
log "Stage 9: Prepare BPE based lang"
|
||||
|
||||
for vocab_size in ${vocab_sizes[@]}; do
|
||||
lang_dir=data/${lang}/lang_bpe_${vocab_size}
|
||||
mkdir -p $lang_dir
|
||||
|
||||
if [ ! -f $lang_dir/transcript_words.txt ]; then
|
||||
log "Generate data for BPE training"
|
||||
file=$(
|
||||
find "data/${lang}/fbank/cv-${lang}_cuts_train.jsonl.gz"
|
||||
)
|
||||
gunzip -c ${file} | awk -F '"' '{print $30}' > $lang_dir/transcript_words.txt
|
||||
|
||||
# Ensure space only appears once
|
||||
sed -i 's/\t/ /g' $lang_dir/transcript_words.txt
|
||||
sed -i 's/[ ][ ]*/ /g' $lang_dir/transcript_words.txt
|
||||
fi
|
||||
|
||||
if [ ! -f $lang_dir/words.txt ]; then
|
||||
cat $lang_dir/transcript_words.txt | sed 's/ /\n/g' \
|
||||
| sort -u | sed '/^$/d' > $lang_dir/words.txt
|
||||
(echo '!SIL'; echo '<SPOKEN_NOISE>'; echo '<UNK>'; ) |
|
||||
cat - $lang_dir/words.txt | sort | uniq | awk '
|
||||
BEGIN {
|
||||
print "<eps> 0";
|
||||
}
|
||||
{
|
||||
if ($1 == "<s>") {
|
||||
print "<s> is in the vocabulary!" | "cat 1>&2"
|
||||
exit 1;
|
||||
}
|
||||
if ($1 == "</s>") {
|
||||
print "</s> is in the vocabulary!" | "cat 1>&2"
|
||||
exit 1;
|
||||
}
|
||||
printf("%s %d\n", $1, NR);
|
||||
}
|
||||
END {
|
||||
printf("#0 %d\n", NR+1);
|
||||
printf("<s> %d\n", NR+2);
|
||||
printf("</s> %d\n", NR+3);
|
||||
}' > $lang_dir/words || exit 1;
|
||||
mv $lang_dir/words $lang_dir/words.txt
|
||||
fi
|
||||
|
||||
if [ ! -f $lang_dir/bpe.model ]; then
|
||||
./local/train_bpe_model.py \
|
||||
--lang-dir $lang_dir \
|
||||
--vocab-size $vocab_size \
|
||||
--transcript $lang_dir/transcript_words.txt
|
||||
fi
|
||||
|
||||
if [ ! -f $lang_dir/L_disambig.pt ]; then
|
||||
./local/prepare_lang_bpe.py --lang-dir $lang_dir
|
||||
|
||||
log "Validating $lang_dir/lexicon.txt"
|
||||
./local/validate_bpe_lexicon.py \
|
||||
--lexicon $lang_dir/lexicon.txt \
|
||||
--bpe-model $lang_dir/bpe.model
|
||||
fi
|
||||
|
||||
if [ ! -f $lang_dir/L.fst ]; then
|
||||
log "Converting L.pt to L.fst"
|
||||
./shared/convert-k2-to-openfst.py \
|
||||
--olabels aux_labels \
|
||||
$lang_dir/L.pt \
|
||||
$lang_dir/L.fst
|
||||
fi
|
||||
|
||||
if [ ! -f $lang_dir/L_disambig.fst ]; then
|
||||
log "Converting L_disambig.pt to L_disambig.fst"
|
||||
./shared/convert-k2-to-openfst.py \
|
||||
--olabels aux_labels \
|
||||
$lang_dir/L_disambig.pt \
|
||||
$lang_dir/L_disambig.fst
|
||||
fi
|
||||
done
|
||||
fi
|
||||
|
||||
@ -0,0 +1,420 @@
|
||||
# Copyright 2023 Xiaomi Corp. (authors: Yifan 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.
|
||||
|
||||
|
||||
import argparse
|
||||
import inspect
|
||||
import logging
|
||||
from functools import lru_cache
|
||||
from pathlib import Path
|
||||
from typing import Any, Dict, Optional
|
||||
|
||||
import torch
|
||||
from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
|
||||
from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures
|
||||
CutConcatenate,
|
||||
CutMix,
|
||||
DynamicBucketingSampler,
|
||||
K2SpeechRecognitionDataset,
|
||||
PrecomputedFeatures,
|
||||
SingleCutSampler,
|
||||
SpecAugment,
|
||||
)
|
||||
from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples
|
||||
AudioSamples,
|
||||
OnTheFlyFeatures,
|
||||
)
|
||||
from lhotse.utils import fix_random_seed
|
||||
from torch.utils.data import DataLoader
|
||||
|
||||
from icefall.utils import str2bool
|
||||
|
||||
|
||||
class _SeedWorkers:
|
||||
def __init__(self, seed: int):
|
||||
self.seed = seed
|
||||
|
||||
def __call__(self, worker_id: int):
|
||||
fix_random_seed(self.seed + worker_id)
|
||||
|
||||
|
||||
class CommonVoiceAsrDataModule:
|
||||
"""
|
||||
DataModule for k2 ASR experiments.
|
||||
It assumes there is always one train and valid dataloader,
|
||||
but there can be multiple test dataloaders (e.g. CommonVoice test-clean
|
||||
and test-other).
|
||||
|
||||
It contains all the common data pipeline modules used in ASR
|
||||
experiments, e.g.:
|
||||
- dynamic batch size,
|
||||
- bucketing samplers,
|
||||
- cut concatenation,
|
||||
- augmentation,
|
||||
- on-the-fly feature extraction
|
||||
|
||||
This class should be derived for specific corpora used in ASR tasks.
|
||||
"""
|
||||
|
||||
def __init__(self, args: argparse.Namespace):
|
||||
self.args = args
|
||||
|
||||
@classmethod
|
||||
def add_arguments(cls, parser: argparse.ArgumentParser):
|
||||
group = parser.add_argument_group(
|
||||
title="ASR data related options",
|
||||
description="These options are used for the preparation of "
|
||||
"PyTorch DataLoaders from Lhotse CutSet's -- they control the "
|
||||
"effective batch sizes, sampling strategies, applied data "
|
||||
"augmentations, etc.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--language",
|
||||
type=str,
|
||||
default="en",
|
||||
help="""Language of Common Voice""",
|
||||
)
|
||||
group.add_argument(
|
||||
"--cv-manifest-dir",
|
||||
type=Path,
|
||||
default=Path("data/en/fbank"),
|
||||
help="Path to directory with CommonVoice train/dev/test cuts.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--manifest-dir",
|
||||
type=Path,
|
||||
default=Path("data/fbank"),
|
||||
help="Path to directory with the other cuts.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--max-duration",
|
||||
type=int,
|
||||
default=200.0,
|
||||
help="Maximum pooled recordings duration (seconds) in a "
|
||||
"single batch. You can reduce it if it causes CUDA OOM.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--bucketing-sampler",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="When enabled, the batches will come from buckets of "
|
||||
"similar duration (saves padding frames).",
|
||||
)
|
||||
group.add_argument(
|
||||
"--num-buckets",
|
||||
type=int,
|
||||
default=30,
|
||||
help="The number of buckets for the DynamicBucketingSampler"
|
||||
"(you might want to increase it for larger datasets).",
|
||||
)
|
||||
group.add_argument(
|
||||
"--concatenate-cuts",
|
||||
type=str2bool,
|
||||
default=False,
|
||||
help="When enabled, utterances (cuts) will be concatenated "
|
||||
"to minimize the amount of padding.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--duration-factor",
|
||||
type=float,
|
||||
default=1.0,
|
||||
help="Determines the maximum duration of a concatenated cut "
|
||||
"relative to the duration of the longest cut in a batch.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--gap",
|
||||
type=float,
|
||||
default=1.0,
|
||||
help="The amount of padding (in seconds) inserted between "
|
||||
"concatenated cuts. This padding is filled with noise when "
|
||||
"noise augmentation is used.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--on-the-fly-feats",
|
||||
type=str2bool,
|
||||
default=False,
|
||||
help="When enabled, use on-the-fly cut mixing and feature "
|
||||
"extraction. Will drop existing precomputed feature manifests "
|
||||
"if available.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--shuffle",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="When enabled (=default), the examples will be "
|
||||
"shuffled for each epoch.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--drop-last",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="Whether to drop last batch. Used by sampler.",
|
||||
)
|
||||
group.add_argument(
|
||||
"--return-cuts",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="When enabled, each batch will have the "
|
||||
"field: batch['supervisions']['cut'] with the cuts that "
|
||||
"were used to construct it.",
|
||||
)
|
||||
|
||||
group.add_argument(
|
||||
"--num-workers",
|
||||
type=int,
|
||||
default=2,
|
||||
help="The number of training dataloader workers that "
|
||||
"collect the batches.",
|
||||
)
|
||||
|
||||
group.add_argument(
|
||||
"--enable-spec-aug",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="When enabled, use SpecAugment for training dataset.",
|
||||
)
|
||||
|
||||
group.add_argument(
|
||||
"--spec-aug-time-warp-factor",
|
||||
type=int,
|
||||
default=80,
|
||||
help="Used only when --enable-spec-aug is True. "
|
||||
"It specifies the factor for time warping in SpecAugment. "
|
||||
"Larger values mean more warping. "
|
||||
"A value less than 1 means to disable time warp.",
|
||||
)
|
||||
|
||||
group.add_argument(
|
||||
"--enable-musan",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="When enabled, select noise from MUSAN and mix it"
|
||||
"with training dataset. ",
|
||||
)
|
||||
|
||||
group.add_argument(
|
||||
"--input-strategy",
|
||||
type=str,
|
||||
default="PrecomputedFeatures",
|
||||
help="AudioSamples or PrecomputedFeatures",
|
||||
)
|
||||
|
||||
def train_dataloaders(
|
||||
self,
|
||||
cuts_train: CutSet,
|
||||
sampler_state_dict: Optional[Dict[str, Any]] = None,
|
||||
) -> DataLoader:
|
||||
"""
|
||||
Args:
|
||||
cuts_train:
|
||||
CutSet for training.
|
||||
sampler_state_dict:
|
||||
The state dict for the training sampler.
|
||||
"""
|
||||
transforms = []
|
||||
if self.args.enable_musan:
|
||||
logging.info("Enable MUSAN")
|
||||
logging.info("About to get Musan cuts")
|
||||
cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
|
||||
transforms.append(
|
||||
CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
|
||||
)
|
||||
else:
|
||||
logging.info("Disable MUSAN")
|
||||
|
||||
if self.args.concatenate_cuts:
|
||||
logging.info(
|
||||
f"Using cut concatenation with duration factor "
|
||||
f"{self.args.duration_factor} and gap {self.args.gap}."
|
||||
)
|
||||
# Cut concatenation should be the first transform in the list,
|
||||
# so that if we e.g. mix noise in, it will fill the gaps between
|
||||
# different utterances.
|
||||
transforms = [
|
||||
CutConcatenate(
|
||||
duration_factor=self.args.duration_factor, gap=self.args.gap
|
||||
)
|
||||
] + transforms
|
||||
|
||||
input_transforms = []
|
||||
if self.args.enable_spec_aug:
|
||||
logging.info("Enable SpecAugment")
|
||||
logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
|
||||
# Set the value of num_frame_masks according to Lhotse's version.
|
||||
# In different Lhotse's versions, the default of num_frame_masks is
|
||||
# different.
|
||||
num_frame_masks = 10
|
||||
num_frame_masks_parameter = inspect.signature(
|
||||
SpecAugment.__init__
|
||||
).parameters["num_frame_masks"]
|
||||
if num_frame_masks_parameter.default == 1:
|
||||
num_frame_masks = 2
|
||||
logging.info(f"Num frame mask: {num_frame_masks}")
|
||||
input_transforms.append(
|
||||
SpecAugment(
|
||||
time_warp_factor=self.args.spec_aug_time_warp_factor,
|
||||
num_frame_masks=num_frame_masks,
|
||||
features_mask_size=27,
|
||||
num_feature_masks=2,
|
||||
frames_mask_size=100,
|
||||
)
|
||||
)
|
||||
else:
|
||||
logging.info("Disable SpecAugment")
|
||||
|
||||
logging.info("About to create train dataset")
|
||||
train = K2SpeechRecognitionDataset(
|
||||
input_strategy=eval(self.args.input_strategy)(),
|
||||
cut_transforms=transforms,
|
||||
input_transforms=input_transforms,
|
||||
return_cuts=self.args.return_cuts,
|
||||
)
|
||||
|
||||
if self.args.on_the_fly_feats:
|
||||
# NOTE: the PerturbSpeed transform should be added only if we
|
||||
# remove it from data prep stage.
|
||||
# Add on-the-fly speed perturbation; since originally it would
|
||||
# have increased epoch size by 3, we will apply prob 2/3 and use
|
||||
# 3x more epochs.
|
||||
# Speed perturbation probably should come first before
|
||||
# concatenation, but in principle the transforms order doesn't have
|
||||
# to be strict (e.g. could be randomized)
|
||||
# transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms # noqa
|
||||
# Drop feats to be on the safe side.
|
||||
train = K2SpeechRecognitionDataset(
|
||||
cut_transforms=transforms,
|
||||
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
|
||||
input_transforms=input_transforms,
|
||||
return_cuts=self.args.return_cuts,
|
||||
)
|
||||
|
||||
if self.args.bucketing_sampler:
|
||||
logging.info("Using DynamicBucketingSampler.")
|
||||
train_sampler = DynamicBucketingSampler(
|
||||
cuts_train,
|
||||
max_duration=self.args.max_duration,
|
||||
shuffle=self.args.shuffle,
|
||||
num_buckets=self.args.num_buckets,
|
||||
drop_last=self.args.drop_last,
|
||||
)
|
||||
else:
|
||||
logging.info("Using SingleCutSampler.")
|
||||
train_sampler = SingleCutSampler(
|
||||
cuts_train,
|
||||
max_duration=self.args.max_duration,
|
||||
shuffle=self.args.shuffle,
|
||||
)
|
||||
logging.info("About to create train dataloader")
|
||||
|
||||
if sampler_state_dict is not None:
|
||||
logging.info("Loading sampler state dict")
|
||||
train_sampler.load_state_dict(sampler_state_dict)
|
||||
|
||||
# 'seed' is derived from the current random state, which will have
|
||||
# previously been set in the main process.
|
||||
seed = torch.randint(0, 100000, ()).item()
|
||||
worker_init_fn = _SeedWorkers(seed)
|
||||
|
||||
train_dl = DataLoader(
|
||||
train,
|
||||
sampler=train_sampler,
|
||||
batch_size=None,
|
||||
num_workers=self.args.num_workers,
|
||||
persistent_workers=False,
|
||||
worker_init_fn=worker_init_fn,
|
||||
)
|
||||
|
||||
return train_dl
|
||||
|
||||
def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
|
||||
transforms = []
|
||||
if self.args.concatenate_cuts:
|
||||
transforms = [
|
||||
CutConcatenate(
|
||||
duration_factor=self.args.duration_factor, gap=self.args.gap
|
||||
)
|
||||
] + transforms
|
||||
|
||||
logging.info("About to create dev dataset")
|
||||
if self.args.on_the_fly_feats:
|
||||
validate = K2SpeechRecognitionDataset(
|
||||
cut_transforms=transforms,
|
||||
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
|
||||
return_cuts=self.args.return_cuts,
|
||||
)
|
||||
else:
|
||||
validate = K2SpeechRecognitionDataset(
|
||||
cut_transforms=transforms,
|
||||
return_cuts=self.args.return_cuts,
|
||||
)
|
||||
valid_sampler = DynamicBucketingSampler(
|
||||
cuts_valid,
|
||||
max_duration=self.args.max_duration,
|
||||
shuffle=False,
|
||||
)
|
||||
logging.info("About to create dev dataloader")
|
||||
valid_dl = DataLoader(
|
||||
validate,
|
||||
sampler=valid_sampler,
|
||||
batch_size=None,
|
||||
num_workers=2,
|
||||
persistent_workers=False,
|
||||
)
|
||||
|
||||
return valid_dl
|
||||
|
||||
def test_dataloaders(self, cuts: CutSet) -> DataLoader:
|
||||
logging.debug("About to create test dataset")
|
||||
test = K2SpeechRecognitionDataset(
|
||||
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
|
||||
if self.args.on_the_fly_feats
|
||||
else eval(self.args.input_strategy)(),
|
||||
return_cuts=self.args.return_cuts,
|
||||
)
|
||||
sampler = DynamicBucketingSampler(
|
||||
cuts,
|
||||
max_duration=self.args.max_duration,
|
||||
shuffle=False,
|
||||
)
|
||||
logging.debug("About to create test dataloader")
|
||||
test_dl = DataLoader(
|
||||
test,
|
||||
batch_size=None,
|
||||
sampler=sampler,
|
||||
num_workers=self.args.num_workers,
|
||||
)
|
||||
return test_dl
|
||||
|
||||
@lru_cache()
|
||||
def train_cuts(self) -> CutSet:
|
||||
logging.info("About to get train cuts")
|
||||
return load_manifest_lazy(
|
||||
self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_train.jsonl.gz"
|
||||
)
|
||||
|
||||
@lru_cache()
|
||||
def dev_cuts(self) -> CutSet:
|
||||
logging.info("About to get dev cuts")
|
||||
return load_manifest_lazy(
|
||||
self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_dev.jsonl.gz"
|
||||
)
|
||||
|
||||
@lru_cache()
|
||||
def test_cuts(self) -> CutSet:
|
||||
logging.info("About to get test cuts")
|
||||
return load_manifest_lazy(
|
||||
self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_test.jsonl.gz"
|
||||
)
|
||||
1
egs/commonvoice/ASR/pruned_transducer_stateless7/beam_search.py
Symbolic link
1
egs/commonvoice/ASR/pruned_transducer_stateless7/beam_search.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
|
||||
962
egs/commonvoice/ASR/pruned_transducer_stateless7/decode.py
Executable file
962
egs/commonvoice/ASR/pruned_transducer_stateless7/decode.py
Executable file
@ -0,0 +1,962 @@
|
||||
#!/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 CommonVoiceAsrDataModule
|
||||
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 icefall import LmScorer, NgramLm
|
||||
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/en/lang_bpe_500/bpe.model",
|
||||
help="Path to the BPE model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--lang-dir",
|
||||
type=Path,
|
||||
default="data/en/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(
|
||||
"--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""",
|
||||
)
|
||||
add_model_arguments(parser)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
def decode_one_batch(
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
sp: spm.SentencePieceProcessor,
|
||||
batch: dict,
|
||||
word_table: Optional[k2.SymbolTable] = None,
|
||||
decoding_graph: Optional[k2.Fsa] = None,
|
||||
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)
|
||||
|
||||
encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
|
||||
|
||||
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())
|
||||
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,
|
||||
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,
|
||||
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
|
||||
|
||||
results = defaultdict(list)
|
||||
for batch_idx, batch in enumerate(dl):
|
||||
texts = batch["supervisions"]["text"]
|
||||
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
|
||||
|
||||
hyps_dict = decode_one_batch(
|
||||
params=params,
|
||||
model=model,
|
||||
sp=sp,
|
||||
decoding_graph=decoding_graph,
|
||||
word_table=word_table,
|
||||
batch=batch,
|
||||
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)
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
parser = get_parser()
|
||||
CommonVoiceAsrDataModule.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 "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 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}")
|
||||
|
||||
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()
|
||||
|
||||
logging.info(params)
|
||||
|
||||
logging.info("About to create model")
|
||||
model = get_transducer_model(params)
|
||||
|
||||
if not params.use_averaged_model:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
elif params.avg == 1:
|
||||
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
|
||||
else:
|
||||
start = params.epoch - params.avg + 1
|
||||
filenames = []
|
||||
for i in range(start, params.epoch + 1):
|
||||
if i >= 1:
|
||||
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
else:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg + 1
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg + 1:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
filename_start = filenames[-1]
|
||||
filename_end = filenames[0]
|
||||
logging.info(
|
||||
"Calculating the averaged model over iteration checkpoints"
|
||||
f" from {filename_start} (excluded) to {filename_end}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
else:
|
||||
assert params.avg > 0, params.avg
|
||||
start = params.epoch - params.avg
|
||||
assert start >= 1, start
|
||||
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
|
||||
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
|
||||
logging.info(
|
||||
f"Calculating the averaged model over epoch range from "
|
||||
f"{start} (excluded) to {params.epoch}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
|
||||
model.to(device)
|
||||
model.eval()
|
||||
|
||||
# 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
|
||||
commonvoice = CommonVoiceAsrDataModule(args)
|
||||
|
||||
dev_cuts = commonvoice.dev_cuts()
|
||||
test_cuts = commonvoice.test_cuts()
|
||||
|
||||
dev_dl = commonvoice.valid_dataloaders(dev_cuts)
|
||||
test_dl = commonvoice.test_dataloaders(test_cuts)
|
||||
|
||||
test_sets = ["dev", "test"]
|
||||
test_dl = [dev_dl, test_dl]
|
||||
|
||||
for test_set, test_dl in zip(test_sets, test_dl):
|
||||
results_dict = decode_dataset(
|
||||
dl=test_dl,
|
||||
params=params,
|
||||
model=model,
|
||||
sp=sp,
|
||||
word_table=word_table,
|
||||
decoding_graph=decoding_graph,
|
||||
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,
|
||||
)
|
||||
|
||||
logging.info("Done!")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
1
egs/commonvoice/ASR/pruned_transducer_stateless7/decoder.py
Symbolic link
1
egs/commonvoice/ASR/pruned_transducer_stateless7/decoder.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
|
||||
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/transducer_stateless/encoder_interface.py
|
||||
600
egs/commonvoice/ASR/pruned_transducer_stateless7/export-onnx.py
Executable file
600
egs/commonvoice/ASR/pruned_transducer_stateless7/export-onnx.py
Executable file
@ -0,0 +1,600 @@
|
||||
#!/usr/bin/env python3
|
||||
#
|
||||
# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang,
|
||||
# Yifan Yang)
|
||||
|
||||
"""
|
||||
This script exports a transducer model from PyTorch to ONNX.
|
||||
|
||||
We use the pre-trained model from
|
||||
https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
|
||||
as an example to show how to use this file.
|
||||
|
||||
1. Download the pre-trained model
|
||||
|
||||
cd egs/librispeech/ASR
|
||||
|
||||
repo_url=https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
|
||||
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
|
||||
repo=$(basename $repo_url)
|
||||
|
||||
pushd $repo
|
||||
git lfs pull --include "data/lang_bpe_500/bpe.model"
|
||||
git lfs pull --include "exp/pretrained.pt"
|
||||
|
||||
cd exp
|
||||
ln -s pretrained.pt epoch-9999.pt
|
||||
popd
|
||||
|
||||
2. Export the model to ONNX
|
||||
|
||||
./pruned_transducer_stateless7/export-onnx.py \
|
||||
--bpe-model $repo/data/en/lang_bpe_500/bpe.model \
|
||||
--use-averaged-model 0 \
|
||||
--epoch 9999 \
|
||||
--avg 1 \
|
||||
--exp-dir $repo/exp
|
||||
|
||||
It will generate the following 3 files inside $repo/exp:
|
||||
|
||||
- encoder-epoch-9999-avg-1.onnx
|
||||
- decoder-epoch-9999-avg-1.onnx
|
||||
- joiner-epoch-9999-avg-1.onnx
|
||||
|
||||
See ./onnx_pretrained.py and ./onnx_check.py for how to
|
||||
use the exported ONNX models.
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
from pathlib import Path
|
||||
from typing import Dict, Tuple
|
||||
|
||||
import onnx
|
||||
import sentencepiece as spm
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from decoder import Decoder
|
||||
from onnxruntime.quantization import QuantType, quantize_dynamic
|
||||
from scaling_converter import convert_scaled_to_non_scaled
|
||||
from train import add_model_arguments, get_params, get_transducer_model
|
||||
from zipformer import Zipformer
|
||||
|
||||
from icefall.checkpoint import (
|
||||
average_checkpoints,
|
||||
average_checkpoints_with_averaged_model,
|
||||
find_checkpoints,
|
||||
load_checkpoint,
|
||||
)
|
||||
from icefall.utils import setup_logger, 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 averaging.
|
||||
Note: Epoch counts from 0.
|
||||
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="pruned_transducer_stateless7/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/en/lang_bpe_500/bpe.model",
|
||||
help="Path to the BPE model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--context-size",
|
||||
type=int,
|
||||
default=2,
|
||||
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
|
||||
)
|
||||
|
||||
add_model_arguments(parser)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
def add_meta_data(filename: str, meta_data: Dict[str, str]):
|
||||
"""Add meta data to an ONNX model. It is changed in-place.
|
||||
|
||||
Args:
|
||||
filename:
|
||||
Filename of the ONNX model to be changed.
|
||||
meta_data:
|
||||
Key-value pairs.
|
||||
"""
|
||||
model = onnx.load(filename)
|
||||
for key, value in meta_data.items():
|
||||
meta = model.metadata_props.add()
|
||||
meta.key = key
|
||||
meta.value = value
|
||||
|
||||
onnx.save(model, filename)
|
||||
|
||||
|
||||
class OnnxEncoder(nn.Module):
|
||||
"""A wrapper for Zipformer and the encoder_proj from the joiner"""
|
||||
|
||||
def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
|
||||
"""
|
||||
Args:
|
||||
encoder:
|
||||
A Zipformer encoder.
|
||||
encoder_proj:
|
||||
The projection layer for encoder from the joiner.
|
||||
"""
|
||||
super().__init__()
|
||||
self.encoder = encoder
|
||||
self.encoder_proj = encoder_proj
|
||||
|
||||
def forward(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
x_lens: torch.Tensor,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
"""Please see the help information of Zipformer.forward
|
||||
|
||||
Args:
|
||||
x:
|
||||
A 3-D tensor of shape (N, T, C)
|
||||
x_lens:
|
||||
A 1-D tensor of shape (N,). Its dtype is torch.int64
|
||||
Returns:
|
||||
Return a tuple containing:
|
||||
- encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
|
||||
- encoder_out_lens, A 1-D tensor of shape (N,)
|
||||
"""
|
||||
encoder_out, encoder_out_lens = self.encoder(x, x_lens)
|
||||
|
||||
encoder_out = self.encoder_proj(encoder_out)
|
||||
# Now encoder_out is of shape (N, T, joiner_dim)
|
||||
|
||||
return encoder_out, encoder_out_lens
|
||||
|
||||
|
||||
class OnnxDecoder(nn.Module):
|
||||
"""A wrapper for Decoder and the decoder_proj from the joiner"""
|
||||
|
||||
def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
|
||||
super().__init__()
|
||||
self.decoder = decoder
|
||||
self.decoder_proj = decoder_proj
|
||||
|
||||
def forward(self, y: torch.Tensor) -> torch.Tensor:
|
||||
"""
|
||||
Args:
|
||||
y:
|
||||
A 2-D tensor of shape (N, context_size).
|
||||
Returns
|
||||
Return a 2-D tensor of shape (N, joiner_dim)
|
||||
"""
|
||||
need_pad = False
|
||||
decoder_output = self.decoder(y, need_pad=need_pad)
|
||||
decoder_output = decoder_output.squeeze(1)
|
||||
output = self.decoder_proj(decoder_output)
|
||||
|
||||
return output
|
||||
|
||||
|
||||
class OnnxJoiner(nn.Module):
|
||||
"""A wrapper for the joiner"""
|
||||
|
||||
def __init__(self, output_linear: nn.Linear):
|
||||
super().__init__()
|
||||
self.output_linear = output_linear
|
||||
|
||||
def forward(
|
||||
self,
|
||||
encoder_out: torch.Tensor,
|
||||
decoder_out: torch.Tensor,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Args:
|
||||
encoder_out:
|
||||
A 2-D tensor of shape (N, joiner_dim)
|
||||
decoder_out:
|
||||
A 2-D tensor of shape (N, joiner_dim)
|
||||
Returns:
|
||||
Return a 2-D tensor of shape (N, vocab_size)
|
||||
"""
|
||||
logit = encoder_out + decoder_out
|
||||
logit = self.output_linear(torch.tanh(logit))
|
||||
return logit
|
||||
|
||||
|
||||
def export_encoder_model_onnx(
|
||||
encoder_model: OnnxEncoder,
|
||||
encoder_filename: str,
|
||||
opset_version: int = 11,
|
||||
) -> None:
|
||||
"""Export the given encoder model to ONNX format.
|
||||
The exported model has two inputs:
|
||||
|
||||
- x, a tensor of shape (N, T, C); dtype is torch.float32
|
||||
- x_lens, a tensor of shape (N,); dtype is torch.int64
|
||||
|
||||
and it has two outputs:
|
||||
|
||||
- encoder_out, a tensor of shape (N, T', joiner_dim)
|
||||
- encoder_out_lens, a tensor of shape (N,)
|
||||
|
||||
Args:
|
||||
encoder_model:
|
||||
The input encoder model
|
||||
encoder_filename:
|
||||
The filename to save the exported ONNX model.
|
||||
opset_version:
|
||||
The opset version to use.
|
||||
"""
|
||||
x = torch.zeros(1, 100, 80, dtype=torch.float32)
|
||||
x_lens = torch.tensor([100], dtype=torch.int64)
|
||||
|
||||
torch.onnx.export(
|
||||
encoder_model,
|
||||
(x, x_lens),
|
||||
encoder_filename,
|
||||
verbose=False,
|
||||
opset_version=opset_version,
|
||||
input_names=["x", "x_lens"],
|
||||
output_names=["encoder_out", "encoder_out_lens"],
|
||||
dynamic_axes={
|
||||
"x": {0: "N", 1: "T"},
|
||||
"x_lens": {0: "N"},
|
||||
"encoder_out": {0: "N", 1: "T"},
|
||||
"encoder_out_lens": {0: "N"},
|
||||
},
|
||||
)
|
||||
|
||||
meta_data = {
|
||||
"model_type": "zipformer",
|
||||
"version": "1",
|
||||
"model_author": "k2-fsa",
|
||||
"comment": "stateless7",
|
||||
}
|
||||
logging.info(f"meta_data: {meta_data}")
|
||||
|
||||
add_meta_data(filename=encoder_filename, meta_data=meta_data)
|
||||
|
||||
|
||||
def export_decoder_model_onnx(
|
||||
decoder_model: OnnxDecoder,
|
||||
decoder_filename: str,
|
||||
opset_version: int = 11,
|
||||
) -> None:
|
||||
"""Export the decoder model to ONNX format.
|
||||
|
||||
The exported model has one input:
|
||||
|
||||
- y: a torch.int64 tensor of shape (N, decoder_model.context_size)
|
||||
|
||||
and has one output:
|
||||
|
||||
- decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
|
||||
|
||||
Args:
|
||||
decoder_model:
|
||||
The decoder model to be exported.
|
||||
decoder_filename:
|
||||
Filename to save the exported ONNX model.
|
||||
opset_version:
|
||||
The opset version to use.
|
||||
"""
|
||||
context_size = decoder_model.decoder.context_size
|
||||
vocab_size = decoder_model.decoder.vocab_size
|
||||
|
||||
y = torch.zeros(10, context_size, dtype=torch.int64)
|
||||
torch.onnx.export(
|
||||
decoder_model,
|
||||
y,
|
||||
decoder_filename,
|
||||
verbose=False,
|
||||
opset_version=opset_version,
|
||||
input_names=["y"],
|
||||
output_names=["decoder_out"],
|
||||
dynamic_axes={
|
||||
"y": {0: "N"},
|
||||
"decoder_out": {0: "N"},
|
||||
},
|
||||
)
|
||||
|
||||
meta_data = {
|
||||
"context_size": str(context_size),
|
||||
"vocab_size": str(vocab_size),
|
||||
}
|
||||
add_meta_data(filename=decoder_filename, meta_data=meta_data)
|
||||
|
||||
|
||||
def export_joiner_model_onnx(
|
||||
joiner_model: nn.Module,
|
||||
joiner_filename: str,
|
||||
opset_version: int = 11,
|
||||
) -> None:
|
||||
"""Export the joiner model to ONNX format.
|
||||
The exported joiner model has two inputs:
|
||||
|
||||
- encoder_out: a tensor of shape (N, joiner_dim)
|
||||
- decoder_out: a tensor of shape (N, joiner_dim)
|
||||
|
||||
and produces one output:
|
||||
|
||||
- logit: a tensor of shape (N, vocab_size)
|
||||
"""
|
||||
joiner_dim = joiner_model.output_linear.weight.shape[1]
|
||||
logging.info(f"joiner dim: {joiner_dim}")
|
||||
|
||||
projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
|
||||
projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
|
||||
|
||||
torch.onnx.export(
|
||||
joiner_model,
|
||||
(projected_encoder_out, projected_decoder_out),
|
||||
joiner_filename,
|
||||
verbose=False,
|
||||
opset_version=opset_version,
|
||||
input_names=[
|
||||
"encoder_out",
|
||||
"decoder_out",
|
||||
],
|
||||
output_names=["logit"],
|
||||
dynamic_axes={
|
||||
"encoder_out": {0: "N"},
|
||||
"decoder_out": {0: "N"},
|
||||
"logit": {0: "N"},
|
||||
},
|
||||
)
|
||||
meta_data = {
|
||||
"joiner_dim": str(joiner_dim),
|
||||
}
|
||||
add_meta_data(filename=joiner_filename, meta_data=meta_data)
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
args = get_parser().parse_args()
|
||||
args.exp_dir = Path(args.exp_dir)
|
||||
|
||||
params = get_params()
|
||||
params.update(vars(args))
|
||||
|
||||
device = torch.device("cpu")
|
||||
if torch.cuda.is_available():
|
||||
device = torch.device("cuda", 0)
|
||||
|
||||
setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
|
||||
|
||||
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 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("cpu")
|
||||
model.eval()
|
||||
|
||||
convert_scaled_to_non_scaled(model, inplace=True)
|
||||
|
||||
encoder = OnnxEncoder(
|
||||
encoder=model.encoder,
|
||||
encoder_proj=model.joiner.encoder_proj,
|
||||
)
|
||||
|
||||
decoder = OnnxDecoder(
|
||||
decoder=model.decoder,
|
||||
decoder_proj=model.joiner.decoder_proj,
|
||||
)
|
||||
|
||||
joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
|
||||
|
||||
encoder_num_param = sum([p.numel() for p in encoder.parameters()])
|
||||
decoder_num_param = sum([p.numel() for p in decoder.parameters()])
|
||||
joiner_num_param = sum([p.numel() for p in joiner.parameters()])
|
||||
total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
|
||||
logging.info(f"encoder parameters: {encoder_num_param}")
|
||||
logging.info(f"decoder parameters: {decoder_num_param}")
|
||||
logging.info(f"joiner parameters: {joiner_num_param}")
|
||||
logging.info(f"total parameters: {total_num_param}")
|
||||
|
||||
if params.iter > 0:
|
||||
suffix = f"iter-{params.iter}"
|
||||
else:
|
||||
suffix = f"epoch-{params.epoch}"
|
||||
|
||||
suffix += f"-avg-{params.avg}"
|
||||
|
||||
opset_version = 13
|
||||
|
||||
logging.info("Exporting encoder")
|
||||
encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
|
||||
export_encoder_model_onnx(
|
||||
encoder,
|
||||
encoder_filename,
|
||||
opset_version=opset_version,
|
||||
)
|
||||
logging.info(f"Exported encoder to {encoder_filename}")
|
||||
|
||||
logging.info("Exporting decoder")
|
||||
decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
|
||||
export_decoder_model_onnx(
|
||||
decoder,
|
||||
decoder_filename,
|
||||
opset_version=opset_version,
|
||||
)
|
||||
logging.info(f"Exported decoder to {decoder_filename}")
|
||||
|
||||
logging.info("Exporting joiner")
|
||||
joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
|
||||
export_joiner_model_onnx(
|
||||
joiner,
|
||||
joiner_filename,
|
||||
opset_version=opset_version,
|
||||
)
|
||||
logging.info(f"Exported joiner to {joiner_filename}")
|
||||
|
||||
# Generate int8 quantization models
|
||||
# See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
|
||||
|
||||
logging.info("Generate int8 quantization models")
|
||||
|
||||
encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
|
||||
quantize_dynamic(
|
||||
model_input=encoder_filename,
|
||||
model_output=encoder_filename_int8,
|
||||
op_types_to_quantize=["MatMul"],
|
||||
weight_type=QuantType.QInt8,
|
||||
)
|
||||
|
||||
decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
|
||||
quantize_dynamic(
|
||||
model_input=decoder_filename,
|
||||
model_output=decoder_filename_int8,
|
||||
op_types_to_quantize=["MatMul"],
|
||||
weight_type=QuantType.QInt8,
|
||||
)
|
||||
|
||||
joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
|
||||
quantize_dynamic(
|
||||
model_input=joiner_filename,
|
||||
model_output=joiner_filename_int8,
|
||||
op_types_to_quantize=["MatMul"],
|
||||
weight_type=QuantType.QInt8,
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
|
||||
|
||||
main()
|
||||
321
egs/commonvoice/ASR/pruned_transducer_stateless7/export.py
Executable file
321
egs/commonvoice/ASR/pruned_transducer_stateless7/export.py
Executable file
@ -0,0 +1,321 @@
|
||||
#!/usr/bin/env python3
|
||||
#
|
||||
# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang,
|
||||
# Yifan 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.
|
||||
|
||||
# This script converts several saved checkpoints
|
||||
# to a single one using model averaging.
|
||||
"""
|
||||
|
||||
Usage:
|
||||
|
||||
(1) Export to torchscript model using torch.jit.script()
|
||||
|
||||
./pruned_transducer_stateless7/export.py \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--bpe-model data/en/lang_bpe_500/bpe.model \
|
||||
--epoch 30 \
|
||||
--avg 5 \
|
||||
--jit 1
|
||||
|
||||
It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
|
||||
load it by `torch.jit.load("cpu_jit.pt")`.
|
||||
|
||||
Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
|
||||
are on CPU. You can use `to("cuda")` to move them to a CUDA device.
|
||||
|
||||
Check
|
||||
https://github.com/k2-fsa/sherpa
|
||||
for how to use the exported models outside of icefall.
|
||||
|
||||
(2) Export `model.state_dict()`
|
||||
|
||||
./pruned_transducer_stateless7/export.py \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--bpe-model data/en/lang_bpe_500/bpe.model \
|
||||
--epoch 30 \
|
||||
--avg 5
|
||||
|
||||
It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
|
||||
load it by `icefall.checkpoint.load_checkpoint()`.
|
||||
|
||||
To use the generated file with `pruned_transducer_stateless7/decode.py`,
|
||||
you can do:
|
||||
|
||||
cd /path/to/exp_dir
|
||||
ln -s pretrained.pt epoch-9999.pt
|
||||
|
||||
cd /path/to/egs/commonvoice/ASR
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--epoch 9999 \
|
||||
--avg 1 \
|
||||
--max-duration 600 \
|
||||
--decoding-method greedy_search \
|
||||
--bpe-model data/en/lang_bpe_500/bpe.model
|
||||
|
||||
Check ./pretrained.py for its usage.
|
||||
|
||||
Note: If you don't want to train a model from scratch, we have
|
||||
provided one for you. You can get it at
|
||||
|
||||
https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
|
||||
|
||||
with the following commands:
|
||||
|
||||
sudo apt-get install git-lfs
|
||||
git lfs install
|
||||
git clone https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
|
||||
# You will find the pre-trained model in icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17/exp
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
from pathlib import Path
|
||||
|
||||
import sentencepiece as spm
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from scaling_converter import convert_scaled_to_non_scaled
|
||||
from train import add_model_arguments, get_params, get_transducer_model
|
||||
|
||||
from icefall.checkpoint import (
|
||||
average_checkpoints,
|
||||
average_checkpoints_with_averaged_model,
|
||||
find_checkpoints,
|
||||
load_checkpoint,
|
||||
)
|
||||
from icefall.utils import str2bool
|
||||
|
||||
|
||||
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="""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/en/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.
|
||||
It will generate a file named cpu_jit.pt
|
||||
|
||||
Check ./jit_pretrained.py for how to use it.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--context-size",
|
||||
type=int,
|
||||
default=2,
|
||||
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
|
||||
)
|
||||
|
||||
add_model_arguments(parser)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
args = get_parser().parse_args()
|
||||
args.exp_dir = Path(args.exp_dir)
|
||||
|
||||
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 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("cpu")
|
||||
model.eval()
|
||||
|
||||
if params.jit is True:
|
||||
convert_scaled_to_non_scaled(model, inplace=True)
|
||||
# We won't use the forward() method of the model in C++, so just ignore
|
||||
# it here.
|
||||
# Otherwise, one of its arguments is a ragged tensor and is not
|
||||
# torch scriptabe.
|
||||
model.__class__.forward = torch.jit.ignore(model.__class__.forward)
|
||||
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 torchscript. Export model.state_dict()")
|
||||
# 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()
|
||||
1
egs/commonvoice/ASR/pruned_transducer_stateless7/joiner.py
Symbolic link
1
egs/commonvoice/ASR/pruned_transducer_stateless7/joiner.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
|
||||
1
egs/commonvoice/ASR/pruned_transducer_stateless7/model.py
Symbolic link
1
egs/commonvoice/ASR/pruned_transducer_stateless7/model.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/pruned_transducer_stateless7/model.py
|
||||
240
egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py
Executable file
240
egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py
Executable file
@ -0,0 +1,240 @@
|
||||
#!/usr/bin/env python3
|
||||
#
|
||||
# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang,
|
||||
# Yifan 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.
|
||||
|
||||
"""
|
||||
This script checks that exported onnx models produce the same output
|
||||
with the given torchscript model for the same input.
|
||||
|
||||
We use the pre-trained model from
|
||||
https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
|
||||
as an example to show how to use this file.
|
||||
|
||||
1. Download the pre-trained model
|
||||
|
||||
cd egs/librispeech/ASR
|
||||
|
||||
repo_url=https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
|
||||
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
|
||||
repo=$(basename $repo_url)
|
||||
|
||||
pushd $repo
|
||||
git lfs pull --include "data/lang_bpe_500/bpe.model"
|
||||
git lfs pull --include "exp/pretrained.pt"
|
||||
|
||||
cd exp
|
||||
ln -s pretrained.pt epoch-9999.pt
|
||||
popd
|
||||
|
||||
2. Export the model via torchscript (torch.jit.script())
|
||||
|
||||
./pruned_transducer_stateless7/export.py \
|
||||
--bpe-model $repo/data/en/lang_bpe_500/bpe.model \
|
||||
--epoch 9999 \
|
||||
--avg 1 \
|
||||
--exp-dir $repo/exp/ \
|
||||
--jit 1
|
||||
|
||||
It will generate the following file in $repo/exp:
|
||||
- cpu_jit.pt
|
||||
|
||||
3. Export the model to ONNX
|
||||
|
||||
./pruned_transducer_stateless7/export-onnx.py \
|
||||
--bpe-model $repo/data/en/lang_bpe_500/bpe.model \
|
||||
--epoch 9999 \
|
||||
--avg 1 \
|
||||
--exp-dir $repo/exp/
|
||||
|
||||
It will generate the following 3 files inside $repo/exp:
|
||||
|
||||
- encoder-epoch-9999-avg-1.onnx
|
||||
- decoder-epoch-9999-avg-1.onnx
|
||||
- joiner-epoch-9999-avg-1.onnx
|
||||
|
||||
4. Run this file
|
||||
|
||||
./pruned_transducer_stateless7/onnx_check.py \
|
||||
--jit-filename $repo/exp/cpu_jit.pt \
|
||||
--onnx-encoder-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
|
||||
--onnx-decoder-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
|
||||
--onnx-joiner-filename $repo/exp/joiner-epoch-9999-avg-1.onnx
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
|
||||
from icefall import is_module_available
|
||||
from onnx_pretrained import OnnxModel
|
||||
|
||||
import torch
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--jit-filename",
|
||||
required=True,
|
||||
type=str,
|
||||
help="Path to the torchscript model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--onnx-encoder-filename",
|
||||
required=True,
|
||||
type=str,
|
||||
help="Path to the onnx encoder model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--onnx-decoder-filename",
|
||||
required=True,
|
||||
type=str,
|
||||
help="Path to the onnx decoder model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--onnx-joiner-filename",
|
||||
required=True,
|
||||
type=str,
|
||||
help="Path to the onnx joiner model",
|
||||
)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
def test_encoder(
|
||||
torch_model: torch.jit.ScriptModule,
|
||||
onnx_model: OnnxModel,
|
||||
):
|
||||
C = 80
|
||||
for i in range(3):
|
||||
N = torch.randint(low=1, high=20, size=(1,)).item()
|
||||
T = torch.randint(low=30, high=50, size=(1,)).item()
|
||||
logging.info(f"test_encoder: iter {i}, N={N}, T={T}")
|
||||
|
||||
x = torch.rand(N, T, C)
|
||||
x_lens = torch.randint(low=30, high=T + 1, size=(N,))
|
||||
x_lens[0] = T
|
||||
|
||||
torch_encoder_out, torch_encoder_out_lens = torch_model.encoder(x, x_lens)
|
||||
torch_encoder_out = torch_model.joiner.encoder_proj(torch_encoder_out)
|
||||
|
||||
onnx_encoder_out, onnx_encoder_out_lens = onnx_model.run_encoder(x, x_lens)
|
||||
|
||||
assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-05), (
|
||||
(torch_encoder_out - onnx_encoder_out).abs().max()
|
||||
)
|
||||
|
||||
|
||||
def test_decoder(
|
||||
torch_model: torch.jit.ScriptModule,
|
||||
onnx_model: OnnxModel,
|
||||
):
|
||||
context_size = onnx_model.context_size
|
||||
vocab_size = onnx_model.vocab_size
|
||||
for i in range(10):
|
||||
N = torch.randint(1, 100, size=(1,)).item()
|
||||
logging.info(f"test_decoder: iter {i}, N={N}")
|
||||
x = torch.randint(
|
||||
low=1,
|
||||
high=vocab_size,
|
||||
size=(N, context_size),
|
||||
dtype=torch.int64,
|
||||
)
|
||||
torch_decoder_out = torch_model.decoder(x, need_pad=torch.tensor([False]))
|
||||
torch_decoder_out = torch_model.joiner.decoder_proj(torch_decoder_out)
|
||||
torch_decoder_out = torch_decoder_out.squeeze(1)
|
||||
|
||||
onnx_decoder_out = onnx_model.run_decoder(x)
|
||||
assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), (
|
||||
(torch_decoder_out - onnx_decoder_out).abs().max()
|
||||
)
|
||||
|
||||
|
||||
def test_joiner(
|
||||
torch_model: torch.jit.ScriptModule,
|
||||
onnx_model: OnnxModel,
|
||||
):
|
||||
encoder_dim = torch_model.joiner.encoder_proj.weight.shape[1]
|
||||
decoder_dim = torch_model.joiner.decoder_proj.weight.shape[1]
|
||||
for i in range(10):
|
||||
N = torch.randint(1, 100, size=(1,)).item()
|
||||
logging.info(f"test_joiner: iter {i}, N={N}")
|
||||
encoder_out = torch.rand(N, encoder_dim)
|
||||
decoder_out = torch.rand(N, decoder_dim)
|
||||
|
||||
projected_encoder_out = torch_model.joiner.encoder_proj(encoder_out)
|
||||
projected_decoder_out = torch_model.joiner.decoder_proj(decoder_out)
|
||||
|
||||
torch_joiner_out = torch_model.joiner(encoder_out, decoder_out)
|
||||
onnx_joiner_out = onnx_model.run_joiner(
|
||||
projected_encoder_out, projected_decoder_out
|
||||
)
|
||||
|
||||
assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), (
|
||||
(torch_joiner_out - onnx_joiner_out).abs().max()
|
||||
)
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
args = get_parser().parse_args()
|
||||
logging.info(vars(args))
|
||||
|
||||
torch_model = torch.jit.load(args.jit_filename)
|
||||
|
||||
onnx_model = OnnxModel(
|
||||
encoder_model_filename=args.onnx_encoder_filename,
|
||||
decoder_model_filename=args.onnx_decoder_filename,
|
||||
joiner_model_filename=args.onnx_joiner_filename,
|
||||
)
|
||||
|
||||
logging.info("Test encoder")
|
||||
test_encoder(torch_model, onnx_model)
|
||||
|
||||
logging.info("Test decoder")
|
||||
test_decoder(torch_model, onnx_model)
|
||||
|
||||
logging.info("Test joiner")
|
||||
test_joiner(torch_model, onnx_model)
|
||||
logging.info("Finished checking ONNX models")
|
||||
|
||||
|
||||
torch.set_num_threads(1)
|
||||
torch.set_num_interop_threads(1)
|
||||
|
||||
# See https://github.com/pytorch/pytorch/issues/38342
|
||||
# and https://github.com/pytorch/pytorch/issues/33354
|
||||
#
|
||||
# If we don't do this, the delay increases whenever there is
|
||||
# a new request that changes the actual batch size.
|
||||
# If you use `py-spy dump --pid <server-pid> --native`, you will
|
||||
# see a lot of time is spent in re-compiling the torch script model.
|
||||
torch._C._jit_set_profiling_executor(False)
|
||||
torch._C._jit_set_profiling_mode(False)
|
||||
torch._C._set_graph_executor_optimize(False)
|
||||
if __name__ == "__main__":
|
||||
torch.manual_seed(20220727)
|
||||
formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
|
||||
|
||||
logging.basicConfig(format=formatter, level=logging.INFO)
|
||||
main()
|
||||
419
egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_pretrained.py
Executable file
419
egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_pretrained.py
Executable file
@ -0,0 +1,419 @@
|
||||
#!/usr/bin/env python3
|
||||
# Copyright 2023 Xiaomi Corp. (authors: Fangjun Kuang,
|
||||
# Yifan 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.
|
||||
"""
|
||||
This script loads ONNX models and uses them to decode waves.
|
||||
You can use the following command to get the exported models:
|
||||
|
||||
We use the pre-trained model from
|
||||
https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
|
||||
as an example to show how to use this file.
|
||||
|
||||
1. Download the pre-trained model
|
||||
|
||||
cd egs/librispeech/ASR
|
||||
|
||||
repo_url=https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
|
||||
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
|
||||
repo=$(basename $repo_url)
|
||||
|
||||
pushd $repo
|
||||
git lfs pull --include "data/lang_bpe_500/bpe.model"
|
||||
git lfs pull --include "exp/pretrained.pt"
|
||||
|
||||
cd exp
|
||||
ln -s pretrained.pt epoch-9999.pt
|
||||
popd
|
||||
|
||||
2. Export the model to ONNX
|
||||
|
||||
./pruned_transducer_stateless7/export-onnx.py \
|
||||
--bpe-model $repo/data/en/lang_bpe_500/bpe.model \
|
||||
--epoch 9999 \
|
||||
--avg 1 \
|
||||
--exp-dir $repo/exp/
|
||||
|
||||
It will generate the following 3 files inside $repo/exp:
|
||||
|
||||
- encoder-epoch-9999-avg-1.onnx
|
||||
- decoder-epoch-9999-avg-1.onnx
|
||||
- joiner-epoch-9999-avg-1.onnx
|
||||
|
||||
3. Run this file
|
||||
|
||||
./pruned_transducer_stateless7/onnx_pretrained.py \
|
||||
--encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
|
||||
--decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
|
||||
--joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
|
||||
--tokens $repo/data/en/lang_bpe_500/tokens.txt \
|
||||
$repo/test_wavs/1089-134686-0001.wav \
|
||||
$repo/test_wavs/1221-135766-0001.wav \
|
||||
$repo/test_wavs/1221-135766-0002.wav
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import math
|
||||
from typing import List, Tuple
|
||||
|
||||
import k2
|
||||
import kaldifeat
|
||||
import numpy as np
|
||||
import onnxruntime as ort
|
||||
import torch
|
||||
import torchaudio
|
||||
from torch.nn.utils.rnn import pad_sequence
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--encoder-model-filename",
|
||||
type=str,
|
||||
required=True,
|
||||
help="Path to the encoder onnx model. ",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--decoder-model-filename",
|
||||
type=str,
|
||||
required=True,
|
||||
help="Path to the decoder onnx model. ",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--joiner-model-filename",
|
||||
type=str,
|
||||
required=True,
|
||||
help="Path to the joiner onnx model. ",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--tokens",
|
||||
type=str,
|
||||
help="""Path to tokens.txt.""",
|
||||
)
|
||||
|
||||
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",
|
||||
)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
class OnnxModel:
|
||||
def __init__(
|
||||
self,
|
||||
encoder_model_filename: str,
|
||||
decoder_model_filename: str,
|
||||
joiner_model_filename: str,
|
||||
):
|
||||
session_opts = ort.SessionOptions()
|
||||
session_opts.inter_op_num_threads = 1
|
||||
session_opts.intra_op_num_threads = 1
|
||||
|
||||
self.session_opts = session_opts
|
||||
|
||||
self.init_encoder(encoder_model_filename)
|
||||
self.init_decoder(decoder_model_filename)
|
||||
self.init_joiner(joiner_model_filename)
|
||||
|
||||
def init_encoder(self, encoder_model_filename: str):
|
||||
self.encoder = ort.InferenceSession(
|
||||
encoder_model_filename,
|
||||
sess_options=self.session_opts,
|
||||
)
|
||||
|
||||
def init_decoder(self, decoder_model_filename: str):
|
||||
self.decoder = ort.InferenceSession(
|
||||
decoder_model_filename,
|
||||
sess_options=self.session_opts,
|
||||
)
|
||||
|
||||
decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
|
||||
self.context_size = int(decoder_meta["context_size"])
|
||||
self.vocab_size = int(decoder_meta["vocab_size"])
|
||||
|
||||
logging.info(f"context_size: {self.context_size}")
|
||||
logging.info(f"vocab_size: {self.vocab_size}")
|
||||
|
||||
def init_joiner(self, joiner_model_filename: str):
|
||||
self.joiner = ort.InferenceSession(
|
||||
joiner_model_filename,
|
||||
sess_options=self.session_opts,
|
||||
)
|
||||
|
||||
joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
|
||||
self.joiner_dim = int(joiner_meta["joiner_dim"])
|
||||
|
||||
logging.info(f"joiner_dim: {self.joiner_dim}")
|
||||
|
||||
def run_encoder(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
x_lens: torch.Tensor,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Args:
|
||||
x:
|
||||
A 3-D tensor of shape (N, T, C)
|
||||
x_lens:
|
||||
A 2-D tensor of shape (N,). Its dtype is torch.int64
|
||||
Returns:
|
||||
Return a tuple containing:
|
||||
- encoder_out, its shape is (N, T', joiner_dim)
|
||||
- encoder_out_lens, its shape is (N,)
|
||||
"""
|
||||
out = self.encoder.run(
|
||||
[
|
||||
self.encoder.get_outputs()[0].name,
|
||||
self.encoder.get_outputs()[1].name,
|
||||
],
|
||||
{
|
||||
self.encoder.get_inputs()[0].name: x.numpy(),
|
||||
self.encoder.get_inputs()[1].name: x_lens.numpy(),
|
||||
},
|
||||
)
|
||||
return torch.from_numpy(out[0]), torch.from_numpy(out[1])
|
||||
|
||||
def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
|
||||
"""
|
||||
Args:
|
||||
decoder_input:
|
||||
A 2-D tensor of shape (N, context_size)
|
||||
Returns:
|
||||
Return a 2-D tensor of shape (N, joiner_dim)
|
||||
"""
|
||||
out = self.decoder.run(
|
||||
[self.decoder.get_outputs()[0].name],
|
||||
{self.decoder.get_inputs()[0].name: decoder_input.numpy()},
|
||||
)[0]
|
||||
|
||||
return torch.from_numpy(out)
|
||||
|
||||
def run_joiner(
|
||||
self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Args:
|
||||
encoder_out:
|
||||
A 2-D tensor of shape (N, joiner_dim)
|
||||
decoder_out:
|
||||
A 2-D tensor of shape (N, joiner_dim)
|
||||
Returns:
|
||||
Return a 2-D tensor of shape (N, vocab_size)
|
||||
"""
|
||||
out = self.joiner.run(
|
||||
[self.joiner.get_outputs()[0].name],
|
||||
{
|
||||
self.joiner.get_inputs()[0].name: encoder_out.numpy(),
|
||||
self.joiner.get_inputs()[1].name: decoder_out.numpy(),
|
||||
},
|
||||
)[0]
|
||||
|
||||
return torch.from_numpy(out)
|
||||
|
||||
|
||||
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}. Given: {sample_rate}"
|
||||
# We use only the first channel
|
||||
ans.append(wave[0])
|
||||
return ans
|
||||
|
||||
|
||||
def greedy_search(
|
||||
model: OnnxModel,
|
||||
encoder_out: torch.Tensor,
|
||||
encoder_out_lens: torch.Tensor,
|
||||
) -> List[List[int]]:
|
||||
"""Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
|
||||
Args:
|
||||
model:
|
||||
The transducer model.
|
||||
encoder_out:
|
||||
A 3-D tensor of shape (N, T, joiner_dim)
|
||||
encoder_out_lens:
|
||||
A 1-D tensor of shape (N,).
|
||||
Returns:
|
||||
Return the decoded results for each utterance.
|
||||
"""
|
||||
assert encoder_out.ndim == 3, encoder_out.shape
|
||||
assert encoder_out.size(0) >= 1, encoder_out.size(0)
|
||||
|
||||
packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
|
||||
input=encoder_out,
|
||||
lengths=encoder_out_lens.cpu(),
|
||||
batch_first=True,
|
||||
enforce_sorted=False,
|
||||
)
|
||||
|
||||
blank_id = 0 # hard-code to 0
|
||||
|
||||
batch_size_list = packed_encoder_out.batch_sizes.tolist()
|
||||
N = encoder_out.size(0)
|
||||
|
||||
assert torch.all(encoder_out_lens > 0), encoder_out_lens
|
||||
assert N == batch_size_list[0], (N, batch_size_list)
|
||||
|
||||
context_size = model.context_size
|
||||
hyps = [[blank_id] * context_size for _ in range(N)]
|
||||
|
||||
decoder_input = torch.tensor(
|
||||
hyps,
|
||||
dtype=torch.int64,
|
||||
) # (N, context_size)
|
||||
|
||||
decoder_out = model.run_decoder(decoder_input)
|
||||
|
||||
offset = 0
|
||||
for batch_size in batch_size_list:
|
||||
start = offset
|
||||
end = offset + batch_size
|
||||
current_encoder_out = packed_encoder_out.data[start:end]
|
||||
# current_encoder_out's shape: (batch_size, joiner_dim)
|
||||
offset = end
|
||||
|
||||
decoder_out = decoder_out[:batch_size]
|
||||
logits = model.run_joiner(current_encoder_out, decoder_out)
|
||||
|
||||
# logits'shape (batch_size, vocab_size)
|
||||
|
||||
assert logits.ndim == 2, logits.shape
|
||||
y = logits.argmax(dim=1).tolist()
|
||||
emitted = False
|
||||
for i, v in enumerate(y):
|
||||
if v != blank_id:
|
||||
hyps[i].append(v)
|
||||
emitted = True
|
||||
if emitted:
|
||||
# update decoder output
|
||||
decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
|
||||
decoder_input = torch.tensor(
|
||||
decoder_input,
|
||||
dtype=torch.int64,
|
||||
)
|
||||
decoder_out = model.run_decoder(decoder_input)
|
||||
|
||||
sorted_ans = [h[context_size:] for h in hyps]
|
||||
ans = []
|
||||
unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
|
||||
for i in range(N):
|
||||
ans.append(sorted_ans[unsorted_indices[i]])
|
||||
|
||||
return ans
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
parser = get_parser()
|
||||
args = parser.parse_args()
|
||||
logging.info(vars(args))
|
||||
model = OnnxModel(
|
||||
encoder_model_filename=args.encoder_model_filename,
|
||||
decoder_model_filename=args.decoder_model_filename,
|
||||
joiner_model_filename=args.joiner_model_filename,
|
||||
)
|
||||
|
||||
logging.info("Constructing Fbank computer")
|
||||
opts = kaldifeat.FbankOptions()
|
||||
opts.device = "cpu"
|
||||
opts.frame_opts.dither = 0
|
||||
opts.frame_opts.snip_edges = False
|
||||
opts.frame_opts.samp_freq = args.sample_rate
|
||||
opts.mel_opts.num_bins = 80
|
||||
|
||||
fbank = kaldifeat.Fbank(opts)
|
||||
|
||||
logging.info(f"Reading sound files: {args.sound_files}")
|
||||
waves = read_sound_files(
|
||||
filenames=args.sound_files,
|
||||
expected_sample_rate=args.sample_rate,
|
||||
)
|
||||
|
||||
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, dtype=torch.int64)
|
||||
encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths)
|
||||
|
||||
hyps = greedy_search(
|
||||
model=model,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
)
|
||||
s = "\n"
|
||||
|
||||
symbol_table = k2.SymbolTable.from_file(args.tokens)
|
||||
|
||||
def token_ids_to_words(token_ids: List[int]) -> str:
|
||||
text = ""
|
||||
for i in token_ids:
|
||||
text += symbol_table[i]
|
||||
return text.replace("▁", " ").strip()
|
||||
|
||||
for filename, hyp in zip(args.sound_files, hyps):
|
||||
words = token_ids_to_words(hyp)
|
||||
s += f"{filename}:\n{words}\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/commonvoice/ASR/pruned_transducer_stateless7/optim.py
Symbolic link
1
egs/commonvoice/ASR/pruned_transducer_stateless7/optim.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
|
||||
355
egs/commonvoice/ASR/pruned_transducer_stateless7/pretrained.py
Executable file
355
egs/commonvoice/ASR/pruned_transducer_stateless7/pretrained.py
Executable file
@ -0,0 +1,355 @@
|
||||
#!/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 script loads a checkpoint and uses it to decode waves.
|
||||
You can generate the checkpoint with the following command:
|
||||
|
||||
./pruned_transducer_stateless7/export.py \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--bpe-model data/en/lang_bpe_500/bpe.model \
|
||||
--epoch 30 \
|
||||
--avg 5
|
||||
|
||||
Usage of this script:
|
||||
|
||||
(1) greedy search
|
||||
./pruned_transducer_stateless7/pretrained.py \
|
||||
--checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
|
||||
--bpe-model ./data/en/lang_bpe_500/bpe.model \
|
||||
--method greedy_search \
|
||||
/path/to/foo.wav \
|
||||
/path/to/bar.wav
|
||||
|
||||
(2) beam search
|
||||
./pruned_transducer_stateless7/pretrained.py \
|
||||
--checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
|
||||
--bpe-model ./data/en/lang_bpe_500/bpe.model \
|
||||
--method beam_search \
|
||||
--beam-size 4 \
|
||||
/path/to/foo.wav \
|
||||
/path/to/bar.wav
|
||||
|
||||
(3) modified beam search
|
||||
./pruned_transducer_stateless7/pretrained.py \
|
||||
--checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
|
||||
--bpe-model ./data/en/lang_bpe_500/bpe.model \
|
||||
--method modified_beam_search \
|
||||
--beam-size 4 \
|
||||
/path/to/foo.wav \
|
||||
/path/to/bar.wav
|
||||
|
||||
(4) fast beam search
|
||||
./pruned_transducer_stateless7/pretrained.py \
|
||||
--checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
|
||||
--bpe-model ./data/en/lang_bpe_500/bpe.model \
|
||||
--method fast_beam_search \
|
||||
--beam-size 4 \
|
||||
/path/to/foo.wav \
|
||||
/path/to/bar.wav
|
||||
|
||||
You can also use `./pruned_transducer_stateless7/exp/epoch-xx.pt`.
|
||||
|
||||
Note: ./pruned_transducer_stateless7/exp/pretrained.pt is generated by
|
||||
./pruned_transducer_stateless7/export.py
|
||||
"""
|
||||
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import math
|
||||
from typing import List
|
||||
|
||||
import k2
|
||||
import kaldifeat
|
||||
import sentencepiece as spm
|
||||
import torch
|
||||
import torchaudio
|
||||
from beam_search import (
|
||||
beam_search,
|
||||
fast_beam_search_one_best,
|
||||
greedy_search,
|
||||
greedy_search_batch,
|
||||
modified_beam_search,
|
||||
)
|
||||
from torch.nn.utils.rnn import pad_sequence
|
||||
from train import add_model_arguments, get_params, get_transducer_model
|
||||
|
||||
from icefall.utils import str2bool
|
||||
|
||||
|
||||
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.""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--method",
|
||||
type=str,
|
||||
default="greedy_search",
|
||||
help="""Possible values are:
|
||||
- greedy_search
|
||||
- beam_search
|
||||
- modified_beam_search
|
||||
- fast_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="""An integer indicating how many candidates we will keep for each
|
||||
frame. Used only when --method is beam_search or
|
||||
modified_beam_search.""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--beam",
|
||||
type=float,
|
||||
default=4,
|
||||
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 --method is fast_beam_search""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--max-contexts",
|
||||
type=int,
|
||||
default=4,
|
||||
help="""Used only when --method is fast_beam_search""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--max-states",
|
||||
type=int,
|
||||
default=8,
|
||||
help="""Used only when --method is fast_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.
|
||||
""",
|
||||
)
|
||||
|
||||
add_model_arguments(parser)
|
||||
|
||||
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}. 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 == "fast_beam_search":
|
||||
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
|
||||
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.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())
|
||||
elif params.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())
|
||||
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/commonvoice/ASR/pruned_transducer_stateless7/scaling.py
Symbolic link
1
egs/commonvoice/ASR/pruned_transducer_stateless7/scaling.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py
|
||||
1250
egs/commonvoice/ASR/pruned_transducer_stateless7/train.py
Executable file
1250
egs/commonvoice/ASR/pruned_transducer_stateless7/train.py
Executable file
File diff suppressed because it is too large
Load Diff
1
egs/commonvoice/ASR/pruned_transducer_stateless7/zipformer.py
Symbolic link
1
egs/commonvoice/ASR/pruned_transducer_stateless7/zipformer.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../../librispeech/ASR/pruned_transducer_stateless7/zipformer.py
|
||||
1
egs/commonvoice/ASR/scaling_converter.py
Symbolic link
1
egs/commonvoice/ASR/scaling_converter.py
Symbolic link
@ -0,0 +1 @@
|
||||
../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
|
||||
Loading…
x
Reference in New Issue
Block a user