mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-26 02:06:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

remove outdated recipes

Browse Source
This commit is contained in:
root 2024-05-02 09:40:40 +09:00
parent 10504555c2
commit 45a122555c
49 changed files with 0 additions and 17601 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

0
egs/reazonspeech/ASR/pruned_transducer_stateless2/__init__.py
View File

377
egs/reazonspeech/ASR/pruned_transducer_stateless2/aishell.py
View File

@ -1,377 +0,0 @@
# Copyright 2021 Piotr Żelasko
# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import inspect
import logging
from functools import lru_cache
from pathlib import Path
from typing import Any, Dict, List, Optional
from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
from lhotse.dataset import (
CutConcatenate,
CutMix,
DynamicBucketingSampler,
K2SpeechRecognitionDataset,
PrecomputedFeatures,
SimpleCutSampler,
SpecAugment,
)
from lhotse.dataset.input_strategies import OnTheFlyFeatures
from torch.utils.data import DataLoader
from icefall.utils import str2bool
class AishellAsrDataModule:
"""
DataModule for k2 ASR experiments.
It assumes there is always one train and valid dataloader,
but there can be multiple test dataloaders (e.g. LibriSpeech 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(
"--manifest-dir",
type=Path,
default=Path("data/fbank"),
help="Path to directory with train/valid/test 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. ",
)
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.
"""
logging.info("About to get Musan cuts")
cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
transforms = []
if self.args.enable_musan:
logging.info("Enable MUSAN")
transforms.append(
CutMix(cuts=cuts_musan, p=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(
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 SimpleCutSampler.")
train_sampler = SimpleCutSampler(
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)
train_dl = DataLoader(
train,
sampler=train_sampler,
batch_size=None,
num_workers=self.args.num_workers,
persistent_workers=False,
)
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.info("About to create test dataset")
test = K2SpeechRecognitionDataset(
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
if self.args.on_the_fly_feats
else PrecomputedFeatures(),
return_cuts=self.args.return_cuts,
)
sampler = DynamicBucketingSampler(
cuts,
max_duration=self.args.max_duration,
shuffle=False,
)
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")
cuts_train = load_manifest_lazy(
self.args.manifest_dir / "aishell_cuts_train.jsonl.gz"
)
return cuts_train
@lru_cache()
def valid_cuts(self) -> CutSet:
logging.info("About to get dev cuts")
return load_manifest_lazy(self.args.manifest_dir / "aishell_cuts_dev.jsonl.gz")
@lru_cache()
def test_cuts(self) -> List[CutSet]:
logging.info("About to get test cuts")
return load_manifest_lazy(self.args.manifest_dir / "aishell_cuts_test.jsonl.gz")

394
egs/reazonspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
View File

@ -1,394 +0,0 @@
# Copyright 2021 Piotr Żelasko
#
# 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, List, Optional
import torch
from lhotse import (
CutSet,
Fbank,
FbankConfig,
load_manifest,
load_manifest_lazy,
set_caching_enabled,
)
from lhotse.dataset import (
CutConcatenate,
CutMix,
DynamicBucketingSampler,
K2SpeechRecognitionDataset,
PrecomputedFeatures,
SimpleCutSampler,
SpecAugment,
)
from lhotse.dataset.input_strategies import 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 ReazonSpeechAsrDataModule:
"""
DataModule for k2 ASR experiments.
It assumes there is always one train and valid dataloader,
but there can be multiple test dataloaders (e.g. LibriSpeech 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(
"--manifest-dir",
type=Path,
default=Path("data/fbank"),
help="Path to directory with train/valid/test 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(
"--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. ",
# )
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.
"""
# logging.info("About to get Musan cuts")
# cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
transforms = []
# if self.args.enable_musan:
# logging.info("Enable MUSAN")
# transforms.append(
# CutMix(cuts=cuts_musan, p=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(
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,
buffer_size=300000,
drop_last=True,
)
else:
logging.info("Using SimpleCutSampler.")
train_sampler = SimpleCutSampler(
cuts_train,
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
)
logging.info("About to create train dataloader")
# '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,
)
if sampler_state_dict is not None:
logging.info("Loading sampler state dict")
train_dl.sampler.load_state_dict(sampler_state_dict)
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,
batch_size=None,
sampler=valid_sampler,
num_workers=self.args.num_workers,
persistent_workers=False,
)
return valid_dl
def test_dataloaders(self, cuts: CutSet) -> DataLoader:
logging.info("About to create test dataset")
test = K2SpeechRecognitionDataset(
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
if self.args.on_the_fly_feats
else PrecomputedFeatures(),
return_cuts=self.args.return_cuts,
)
sampler = DynamicBucketingSampler(
cuts,
max_duration=self.args.max_duration,
shuffle=False,
)
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.manifest_dir / "reazonspeech_cuts_train.jsonl.gz")
@lru_cache()
def valid_cuts(self) -> CutSet:
logging.info("About to get dev cuts")
return load_manifest_lazy(self.args.manifest_dir / "reazonspeech_cuts_dev.jsonl.gz")
@lru_cache()
def test_cuts(self) -> List[CutSet]:
logging.info("About to get test cuts")
return load_manifest_lazy(self.args.manifest_dir / "reazonspeech_cuts_test.jsonl.gz")

2942
egs/reazonspeech/ASR/pruned_transducer_stateless2/beam_search.py
View File

File diff suppressed because it is too large Load Diff

1600
egs/reazonspeech/ASR/pruned_transducer_stateless2/conformer.py
View File

File diff suppressed because it is too large Load Diff

690
egs/reazonspeech/ASR/pruned_transducer_stateless2/decode.py
View File

@ -1,690 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
When training with the L subset, usage:
(1) greedy search
./pruned_transducer_stateless2/decode.py \
--epoch 10 \
--avg 2 \
--exp-dir ./pruned_transducer_stateless2/exp \
--lang-dir data/lang_char \
--max-duration 100 \
--decoding-method greedy_search
(2) modified beam search
./pruned_transducer_stateless2/decode.py \
--epoch 10 \
--avg 2 \
--exp-dir ./pruned_transducer_stateless2/exp \
--lang-dir data/lang_char \
--max-duration 100 \
--decoding-method modified_beam_search \
--beam-size 4
(3) fast beam search (1best)
./pruned_transducer_stateless2/decode.py \
--epoch 10 \
--avg 2 \
--exp-dir ./pruned_transducer_stateless2/exp \
--lang-dir data/lang_char \
--max-duration 1500 \
--decoding-method fast_beam_search \
--beam 4 \
--max-contexts 4 \
--max-states 8
(4) fast beam search (nbest)
./pruned_transducer_stateless2/decode.py \
--epoch 10 \
--avg 2 \
--exp-dir ./pruned_transducer_stateless2/exp \
--lang-dir data/lang_char \
--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
(5) fast beam search (nbest oracle WER)
./pruned_transducer_stateless2/decode.py \
--epoch 10 \
--avg 2 \
--exp-dir ./pruned_transducer_stateless2/exp \
--lang-dir data/lang_char \
--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
(6) fast beam search (with LG)
./pruned_transducer_stateless2/decode.py \
--epoch 10 \
--avg 2 \
--exp-dir ./pruned_transducer_stateless2/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method fast_beam_search_nbest_LG \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
"""
import argparse
import logging
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import k2
import torch
import torch.nn as nn
from asr_datamodule import ReazonSpeechAsrDataModule
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,
)
from train import get_params, get_transducer_model
from tokenizer import Tokenizer
from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
from icefall.utils import (
AttributeDict,
setup_logger,
store_transcripts,
write_error_stats,
)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=28,
help="It specifies the checkpoint to use for decoding."
"Note: Epoch counts from 0.",
)
parser.add_argument(
"--batch",
type=int,
default=None,
help="It specifies the batch checkpoint to use for decoding."
"Note: Epoch counts from 0.",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch'. ",
)
parser.add_argument(
"--avg-last-n",
type=int,
default=0,
help="""If positive, --epoch and --avg are ignored and it
will use the last n checkpoints exp_dir/checkpoint-xxx.pt
where xxx is the number of processed batches while
saving that checkpoint.
""",
)
parser.add_argument(
"--exp-dir",
type=str,
default="pruned_transducer_stateless2/exp",
help="The experiment dir",
)
parser.add_argument(
"--lang-dir",
type=str,
default="data/lang_char",
help="""The lang dir
It contains language related input files such as
"lexicon.txt"
""",
)
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
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=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 --decoding-method is fast_beam_search""",
)
parser.add_argument(
"--ngram-lm-scale",
type=float,
default=0.35,
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=4,
help="""Used only when --decoding-method is
fast_beam_search""",
)
parser.add_argument(
"--max-states",
type=int,
default=8,
help="""Used only when --decoding-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 --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""",
)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
sp: Tokenizer,
batch: dict,
word_talbe: Optional[k2.SymbolTable] = None,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[List[str]]]:
"""Decode one batch and return the result in a dict. The dict has the
following format:
- key: It indicates the setting used for decoding. For example,
if 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.
Returns:
Return the decoding result. See above description for the format of
the returned dict.
"""
device = model.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(sp.text2word(hyp))
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_talbe[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(sp.text2word(hyp))
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(sp.text2word(hyp))
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(sp.text2word(hyp))
elif params.decoding_method == "modified_beam_search":
hyp_tokens = modified_beam_search(
model=model,
encoder_out=encoder_out,
beam=params.beam_size,
encoder_out_lens=encoder_out_lens,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(sp.text2word(hyp))
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.text2word(sp.decode(hyp)))
if params.decoding_method == "greedy_search":
return {"greedy_search": hyps}
elif params.decoding_method == "fast_beam_search":
return {
(
f"beam_{params.beam}_"
f"max_contexts_{params.max_contexts}_"
f"max_states_{params.max_states}"
): hyps
}
else:
return {f"beam_size_{params.beam_size}": hyps}
def decode_dataset(
dl: torch.utils.data.DataLoader,
params: AttributeDict,
model: nn.Module,
sp: Tokenizer,
word_table: Optional[k2.SymbolTable] = None,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
"""Decode dataset.
Args:
dl:
PyTorch's dataloader containing the dataset to decode.
params:
It is returned by :func:`get_params`.
model:
The neural model.
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.
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 = 100
else:
log_interval = 2
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_talbe=word_table,
batch=batch,
)
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 = sp.text2word(ref_text)
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()
ReazonSpeechAsrDataModule.add_arguments(parser)
Tokenizer.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",
)
params.res_dir = params.exp_dir / params.decoding_method
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 params.decoding_method == "fast_beam_search_nbest_LG":
params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
if (
params.decoding_method == "fast_beam_search_nbest"
or params.decoding_method == "fast_beam_search_nbest_oracle"
):
params.suffix += f"-nbest-scale-{params.nbest_scale}"
elif "beam_search" in params.decoding_method:
params.suffix += f"-beam-{params.beam_size}"
else:
params.suffix += f"-context-{params.context_size}"
params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
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 = Tokenizer.load(params.lang_dir, params.lang_type)
# <blk> and <unk> are defined in local/prepare_lang_char.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 params.avg_last_n > 0:
filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
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)
elif params.batch is not None:
filenames = f"{params.exp_dir}/checkpoint-{params.batch}.pt"
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints([filenames], device=device))
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if start >= 0:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
model.to(device)
model.eval()
model.device = device
if "fast_beam_search" in params.decoding_method:
if params.decoding_method == "fast_beam_search_nbest_LG":
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:
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
else:
decoding_graph = 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
reazonspeech = ReazonSpeechAsrDataModule(args)
dev_cuts = reazonspeech.valid_cuts()
dev_dl = reazonspeech.valid_dataloaders(dev_cuts)
test_cuts = reazonspeech.test_cuts()
test_dl = reazonspeech.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,
decoding_graph=decoding_graph,
)
save_results(
params=params,
test_set_name=test_set,
results_dict=results_dict,
)
logging.info("Done!")
if __name__ == "__main__":
main()

547
egs/reazonspeech/ASR/pruned_transducer_stateless2/decode_aishell.py
View File

@ -1,547 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Usage:
(1) greedy search
./pruned_transducer_stateless2/decode.py \
--epoch 84 \
--avg 25 \
--exp-dir ./pruned_transducer_stateless2/exp \
--max-duration 600 \
--decoding-method greedy_search
(2) beam search (not recommended)
./pruned_transducer_stateless2/decode.py \
--epoch 84 \
--avg 25 \
--exp-dir ./pruned_transducer_stateless2/exp \
--max-duration 600 \
--decoding-method beam_search \
--beam-size 4
(3) modified beam search
./pruned_transducer_stateless2/decode.py \
--epoch 84 \
--avg 25 \
--exp-dir ./pruned_transducer_stateless2/exp \
--max-duration 600 \
--decoding-method modified_beam_search \
--beam-size 4
(4) fast beam search
./pruned_transducer_stateless2/decode.py \
--epoch 84 \
--avg 25 \
--exp-dir ./pruned_transducer_stateless2/exp \
--max-duration 600 \
--decoding-method fast_beam_search \
--beam 4 \
--max-contexts 4 \
--max-states 8
"""
import argparse
import logging
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import k2
import torch
import torch.nn as nn
from aishell import AishellAsrDataModule
from beam_search import (
beam_search,
fast_beam_search_one_best,
greedy_search,
greedy_search_batch,
modified_beam_search,
)
from finetune import get_params, get_transducer_model
from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
from icefall.lexicon import Lexicon
from icefall.utils import (
AttributeDict,
setup_logger,
store_transcripts,
write_error_stats,
)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=30,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--exp-dir",
type=str,
default="pruned_transducer_stateless2/exp",
help="The experiment dir",
)
parser.add_argument(
"--lang-dir",
type=str,
default="data/lang_char",
help="The lang dir",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_search
- beam_search
- modified_beam_search
- fast_beam_search
""",
)
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=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 --decoding-method is fast_beam_search""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=4,
help="""Used only when --decoding-method is
fast_beam_search""",
)
parser.add_argument(
"--max-states",
type=int,
default=8,
help="""Used only when --decoding-method is
fast_beam_search""",
)
parser.add_argument(
"--context-size",
type=int,
default=1,
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""",
)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
token_table: k2.SymbolTable,
batch: dict,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[List[str]]]:
"""Decode one batch and return the result in a dict. The dict has the
following format:
- key: It indicates the setting used for decoding. For example,
if 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.
token_table:
It maps token ID to a string.
batch:
It is the return value from iterating
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
for the format of the `batch`.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
only when --decoding_method is fast_beam_search.
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)
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,
)
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,
)
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,
)
else:
hyp_tokens = []
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}"
)
hyp_tokens.append(hyp)
hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
if params.decoding_method == "greedy_search":
return {"greedy_search": hyps}
elif params.decoding_method == "fast_beam_search":
return {
(
f"beam_{params.beam}_"
f"max_contexts_{params.max_contexts}_"
f"max_states_{params.max_states}"
): hyps
}
else:
return {f"beam_size_{params.beam_size}": hyps}
def decode_dataset(
dl: torch.utils.data.DataLoader,
params: AttributeDict,
model: nn.Module,
token_table: k2.SymbolTable,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
"""Decode dataset.
Args:
dl:
PyTorch's dataloader containing the dataset to decode.
params:
It is returned by :func:`get_params`.
model:
The neural model.
token_table:
It maps a token ID to a string.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
only when --decoding_method is fast_beam_search.
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,
token_table=token_table,
decoding_graph=decoding_graph,
batch=batch,
)
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"
# we compute CER for aishell dataset.
results_char = []
for res in results:
results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
with open(errs_filename, "w") as f:
wer = write_error_stats(
f, f"{test_set_name}-{key}", results_char, 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()
AishellAsrDataModule.add_arguments(parser)
args = parser.parse_args()
args.exp_dir = Path(args.exp_dir)
args.lang_dir = Path(args.lang_dir)
params = get_params()
params.update(vars(args))
assert params.decoding_method in (
"greedy_search",
"beam_search",
"fast_beam_search",
"modified_beam_search",
)
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}"
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}"
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}")
lexicon = Lexicon(params.lang_dir)
params.blank_id = 0
params.vocab_size = max(lexicon.tokens) + 1
logging.info(params)
logging.info("About to create model")
model = get_transducer_model(params)
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 --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), strict=False
)
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), strict=False
)
model.to(device)
model.eval()
if params.decoding_method == "fast_beam_search":
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
else:
decoding_graph = None
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
aishell = AishellAsrDataModule(args)
test_cuts = aishell.test_cuts()
dev_cuts = aishell.valid_cuts()
test_dl = aishell.test_dataloaders(test_cuts)
dev_dl = aishell.test_dataloaders(dev_cuts)
test_sets = ["test", "dev"]
test_dls = [test_dl, dev_dl]
for test_set, test_dl in zip(test_sets, test_dls):
results_dict = decode_dataset(
dl=test_dl,
params=params,
model=model,
token_table=lexicon.token_table,
decoding_graph=decoding_graph,
)
save_results(
params=params,
test_set_name=test_set,
results_dict=results_dict,
)
logging.info("Done!")
if __name__ == "__main__":
main()

1
egs/reazonspeech/ASR/pruned_transducer_stateless2/decode_stream.py
View File

@ -1 +0,0 @@
/var/data/share20/qc/k2/Github/icefall/egs/librispeech/ASR/pruned_transducer_stateless2/decode_stream.py

122
egs/reazonspeech/ASR/pruned_transducer_stateless2/decoder.py
View File

@ -1,122 +0,0 @@
# 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.
import torch
import torch.nn as nn
import torch.nn.functional as F
from scaling import ScaledConv1d, ScaledEmbedding
from icefall.utils import is_jit_tracing
class Decoder(nn.Module):
"""This class modifies the stateless decoder from the following paper:
RNN-transducer with stateless prediction network
https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
It removes the recurrent connection from the decoder, i.e., the prediction
network. Different from the above paper, it adds an extra Conv1d
right after the embedding layer.
TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
"""
def __init__(
self,
vocab_size: int,
decoder_dim: int,
blank_id: int,
context_size: int,
):
"""
Args:
vocab_size:
Number of tokens of the modeling unit including blank.
decoder_dim:
Dimension of the input embedding, and of the decoder output.
blank_id:
The ID of the blank symbol.
context_size:
Number of previous words to use to predict the next word.
1 means bigram; 2 means trigram. n means (n+1)-gram.
"""
super().__init__()
self.embedding = ScaledEmbedding(
num_embeddings=vocab_size,
embedding_dim=decoder_dim,
)
self.blank_id = blank_id
assert context_size >= 1, context_size
self.context_size = context_size
self.vocab_size = vocab_size
if context_size > 1:
self.conv = ScaledConv1d(
in_channels=decoder_dim,
out_channels=decoder_dim,
kernel_size=context_size,
padding=0,
groups=decoder_dim,
bias=False,
)
else:
# It is to support torch script
self.conv = nn.Identity()
def forward(
self,
y: torch.Tensor,
need_pad: bool = True # Annotation should be Union[bool, torch.Tensor]
# but, torch.jit.script does not support Union.
) -> torch.Tensor:
"""
Args:
y:
A 2-D tensor of shape (N, U).
need_pad:
True to left pad the input. Should be True during training.
False to not pad the input. Should be False during inference.
Returns:
Return a tensor of shape (N, U, decoder_dim).
"""
if isinstance(need_pad, torch.Tensor):
# This is for torch.jit.trace(), which cannot handle the case
# when the input argument is not a tensor.
need_pad = bool(need_pad)
y = y.to(torch.int64)
# this stuff about clamp() is a temporary fix for a mismatch
# at utterance start, we use negative ids in beam_search.py
if torch.jit.is_tracing():
# This is for exporting to PNNX via ONNX
embedding_out = self.embedding(y)
else:
embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
if self.context_size > 1:
embedding_out = embedding_out.permute(0, 2, 1)
if need_pad:
embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
else:
# During inference time, there is no need to do extra padding
# as we only need one output
if not is_jit_tracing():
assert embedding_out.size(-1) == self.context_size
embedding_out = self.conv(embedding_out)
embedding_out = embedding_out.permute(0, 2, 1)
embedding_out = F.relu(embedding_out)
return embedding_out

43
egs/reazonspeech/ASR/pruned_transducer_stateless2/encoder_interface.py
View File

@ -1,43 +0,0 @@
# 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.
from typing import Tuple
import torch
import torch.nn as nn
class EncoderInterface(nn.Module):
def forward(
self, x: torch.Tensor, x_lens: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Args:
x:
A tensor of shape (batch_size, input_seq_len, num_features)
containing the input features.
x_lens:
A tensor of shape (batch_size,) containing the number of frames
in `x` before padding.
Returns:
Return a tuple containing two tensors:
- encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
containing unnormalized probabilities, i.e., the output of a
linear layer.
- encoder_out_lens, a tensor of shape (batch_size,) containing
the number of frames in `encoder_out` before padding.
"""
raise NotImplementedError("Please implement it in a subclass")

518
egs/reazonspeech/ASR/pruned_transducer_stateless2/export-onnx.py
View File

@ -1,518 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
"""
This script exports a transducer model from PyTorch to ONNX.
We use the pre-trained model from
https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2
as an example to show how to use this file.
1. Download the pre-trained model
cd egs/wenetspeech/ASR
repo_url=icefall_asr_wenetspeech_pruned_transducer_stateless2
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
repo=$(basename $repo_url)
pushd $repo
git lfs pull --include "data/lang_char/Linv.pt"
git lfs pull --include "exp/pretrained_epoch_10_avg_2.pt"
cd exp
ln -s pretrained_epoch_10_avg_2.pt epoch-99.pt
popd
2. Export the model to ONNX
./pruned_transducer_stateless2/export-onnx.py \
--lang-dir $repo/data/lang_char \
--epoch 99 \
--avg 1 \
--exp-dir $repo/exp
It will generate the following 3 files inside $repo/exp:
- encoder-epoch-99-avg-1.onnx
- decoder-epoch-99-avg-1.onnx
- joiner-epoch-99-avg-1.onnx
See ./onnx_pretrained.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 torch
import torch.nn as nn
from conformer import Conformer
from decoder import Decoder
from onnxruntime.quantization import QuantType, quantize_dynamic
from scaling_converter import convert_scaled_to_non_scaled
from train import get_params, get_transducer_model
from icefall.checkpoint import (
average_checkpoints,
average_checkpoints_with_averaged_model,
find_checkpoints,
load_checkpoint,
)
from icefall.lexicon import Lexicon
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(
"--exp-dir",
type=str,
default="pruned_transducer_stateless5/exp",
help="""It specifies the directory where all training related
files, e.g., checkpoints, log, etc, are saved
""",
)
parser.add_argument(
"--lang-dir",
type=str,
default="data/lang_char",
help="The lang dir",
)
parser.add_argument(
"--context-size",
type=int,
default=2,
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
)
return parser
def 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 Conformer and the encoder_proj from the joiner"""
def __init__(self, encoder: Conformer, encoder_proj: nn.Linear):
"""
Args:
encoder:
A Conformer 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 Conformer.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": "conformer",
"version": "1",
"model_author": "k2-fsa",
"comment": "stateless5",
}
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)
decoder_model = torch.jit.script(decoder_model)
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}")
lexicon = Lexicon(params.lang_dir)
params.blank_id = 0
params.vocab_size = max(lexicon.tokens) + 1
logging.info(params)
logging.info("About to create model")
model = get_transducer_model(params)
model.to(device)
if params.avg == 1:
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if start >= 0:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
model.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", "Gather"],
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()

238
egs/reazonspeech/ASR/pruned_transducer_stateless2/export.py
View File

@ -1,238 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This script converts several saved checkpoints
# to a single one using model averaging.
"""
Usage:
(1) Export to torchscript model using torch.jit.script()
./pruned_transducer_stateless2/export.py \
--exp-dir ./pruned_transducer_stateless2/exp \
--lang data/lang_char \
--epoch 26 \
--avg 5 \
--jit true
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.
Please refer to
https://k2-fsa.github.io/sherpa/python/offline_asr/conformer/index.html
for how to use `cpu_jit.pt` for speech recognition.
(2) Export `model.state_dict()`
./pruned_transducer_stateless2/export.py \
--exp-dir ./pruned_transducer_stateless2/exp \
--lang data/lang_char \
--epoch 26 \
--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_stateless2/decode.py`,
you can do:
cd /path/to/exp_dir
ln -s pretrained.pt epoch-9999.pt
cd /path/to/egs/reazonspeech/ASR
./pruned_transducer_stateless2/decode.py \
--exp-dir ./pruned_transducer_stateless2/exp \
--epoch 9999 \
--avg 1 \
--max-duration 180 \
--decoding-method greedy_search \
--lang data/lang_char
You can find pretrained models at
https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2/tree/main/exp
"""
import argparse
import logging
from pathlib import Path
import torch
from train import get_params, get_transducer_model
from tokenizer import Tokenizer
from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
from icefall.lexicon import Lexicon
from icefall.utils import str2bool
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=29,
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(
"--exp-dir",
type=str,
default="pruned_transducer_stateless2/exp",
help="""It specifies the directory where all training related
files, e.g., checkpoints, log, etc, are saved
""",
)
parser.add_argument(
"--lang-dir",
type=str,
default="data/lang_char",
help="The lang dir",
)
parser.add_argument(
"--jit",
type=str2bool,
default=False,
help="""True to save a model after applying torch.jit.script.
""",
)
parser.add_argument(
"--context-size",
type=int,
default=1,
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
)
# add_model_arguments(parser)
return parser
def main():
parser = get_parser()
Tokenizer.add_arguments(parser)
args = 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 = Tokenizer.load(params.lang_dir, params.lang_type)
# <blk> is defined in local/prepare_lang_char.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)
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 --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))
model.to("cpu")
model.eval()
if params.jit:
# 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 / f"cpu_jit-epoch-{params.epoch}-avg-{params.avg}.pt"
model.save(str(filename))
logging.info(f"Saved to {filename}")
else:
logging.info("Not using torch.jit.script")
# Save it using a format so that it can be loaded
# by :func:`load_checkpoint`
filename = (
params.exp_dir / f"pretrained-epoch-{params.epoch}-avg-{params.avg}.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()

1054
egs/reazonspeech/ASR/pruned_transducer_stateless2/finetune.py
View File

File diff suppressed because it is too large Load Diff

336
egs/reazonspeech/ASR/pruned_transducer_stateless2/jit_pretrained.py
View File

@ -1,336 +0,0 @@
#!/usr/bin/env python3
# Copyright 2022 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 torchscript models, either exported by `torch.jit.trace()`
or by `torch.jit.script()`, and uses them to decode waves.
You can use the following command to get the exported models:
./pruned_transducer_stateless2/export.py \
--exp-dir ./pruned_transducer_stateless2/exp \
--tokens data/lang_char/tokens.txt \
--epoch 10 \
--avg 2 \
--jit-trace 1
or
./pruned_transducer_stateless2/export.py \
--exp-dir ./pruned_transducer_stateless2/exp \
--tokens data/lang_char/tokens.txt \
--epoch 10 \
--avg 2 \
--jit 1
Usage of this script:
./pruned_transducer_stateless2/jit_pretrained.py \
--encoder-model-filename ./pruned_transducer_stateless2/exp/encoder_jit_trace.pt \
--decoder-model-filename ./pruned_transducer_stateless2/exp/decoder_jit_trace.pt \
--joiner-model-filename ./pruned_transducer_stateless2/exp/joiner_jit_trace.pt \
--tokens data/lang_char/tokens.txt \
/path/to/foo.wav \
/path/to/bar.wav
or
./pruned_transducer_stateless2/jit_pretrained.py \
--encoder-model-filename ./pruned_transducer_stateless2/exp/encoder_jit_script.pt \
--decoder-model-filename ./pruned_transducer_stateless2/exp/decoder_jit_script.pt \
--joiner-model-filename ./pruned_transducer_stateless2/exp/joiner_jit_script.pt \
--tokens data/lang_char/tokens.txt \
/path/to/foo.wav \
/path/to/bar.wav
You can find pretrained models at
https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2/tree/main/exp
"""
import argparse
import logging
import math
from typing import List
import k2
import kaldifeat
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 torchscript model. ",
)
parser.add_argument(
"--decoder-model-filename",
type=str,
required=True,
help="Path to the decoder torchscript model. ",
)
parser.add_argument(
"--joiner-model-filename",
type=str,
required=True,
help="Path to the joiner torchscript 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",
)
parser.add_argument(
"--context-size",
type=int,
default=2,
help="Context size of the decoder model",
)
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
def greedy_search(
decoder: torch.jit.ScriptModule,
joiner: torch.jit.ScriptModule,
encoder_out: torch.Tensor,
encoder_out_lens: torch.Tensor,
context_size: int,
) -> List[List[int]]:
"""Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
Args:
decoder:
The decoder model.
joiner:
The joiner model.
encoder_out:
A 3-D tensor of shape (N, T, C)
encoder_out_lens:
A 1-D tensor of shape (N,).
context_size:
The context size of the decoder model.
Returns:
Return the decoded results for each utterance.
"""
assert encoder_out.ndim == 3
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,
)
device = encoder_out.device
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)
hyps = [[blank_id] * context_size for _ in range(N)]
decoder_input = torch.tensor(
hyps,
device=device,
dtype=torch.int64,
) # (N, context_size)
decoder_out = decoder(
decoder_input,
need_pad=torch.tensor([False]),
).squeeze(1)
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 = current_encoder_out
# current_encoder_out's shape: (batch_size, encoder_out_dim)
offset = end
decoder_out = decoder_out[:batch_size]
logits = 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,
device=device,
dtype=torch.int64,
)
decoder_out = decoder(
decoder_input,
need_pad=torch.tensor([False]),
)
decoder_out = decoder_out.squeeze(1)
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))
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"device: {device}")
encoder = torch.jit.load(args.encoder_model_filename)
decoder = torch.jit.load(args.decoder_model_filename)
joiner = torch.jit.load(args.joiner_model_filename)
encoder.eval()
decoder.eval()
joiner.eval()
encoder.to(device)
decoder.to(device)
joiner.to(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 = 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,
)
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 = encoder(
x=features,
x_lens=feature_lengths,
)
hyps = greedy_search(
decoder=decoder,
joiner=joiner,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
context_size=args.context_size,
)
symbol_table = k2.SymbolTable.from_file(args.tokens)
s = "\n"
for filename, hyp in zip(args.sound_files, hyps):
words = "".join([symbol_table[i] for i in 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()

67
egs/reazonspeech/ASR/pruned_transducer_stateless2/joiner.py
View File

@ -1,67 +0,0 @@
# 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.
import torch
import torch.nn as nn
from scaling import ScaledLinear
from icefall.utils import is_jit_tracing
class Joiner(nn.Module):
def __init__(
self,
encoder_dim: int,
decoder_dim: int,
joiner_dim: int,
vocab_size: int,
):
super().__init__()
self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim)
self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim)
self.output_linear = ScaledLinear(joiner_dim, vocab_size)
def forward(
self,
encoder_out: torch.Tensor,
decoder_out: torch.Tensor,
project_input: bool = True,
) -> torch.Tensor:
"""
Args:
encoder_out:
Output from the encoder. Its shape is (N, T, s_range, C).
decoder_out:
Output from the decoder. Its shape is (N, T, s_range, C).
project_input:
If true, apply input projections encoder_proj and decoder_proj.
If this is false, it is the user's responsibility to do this
manually.
Returns:
Return a tensor of shape (N, T, s_range, C).
"""
if not is_jit_tracing():
assert encoder_out.ndim == decoder_out.ndim
if project_input:
logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
else:
logit = encoder_out + decoder_out
logit = self.output_linear(torch.tanh(logit))
return logit

102
egs/reazonspeech/ASR/pruned_transducer_stateless2/lstmp.py
View File

@ -1,102 +0,0 @@
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
class LSTMP(nn.Module):
"""LSTM with projection.
PyTorch does not support exporting LSTM with projection to ONNX.
This class reimplements LSTM with projection using basic matrix-matrix
and matrix-vector operations. It is not intended for training.
"""
def __init__(self, lstm: nn.LSTM):
"""
Args:
lstm:
LSTM with proj_size. We support only uni-directional,
1-layer LSTM with projection at present.
"""
super().__init__()
assert lstm.bidirectional is False, lstm.bidirectional
assert lstm.num_layers == 1, lstm.num_layers
assert 0 < lstm.proj_size < lstm.hidden_size, (
lstm.proj_size,
lstm.hidden_size,
)
assert lstm.batch_first is False, lstm.batch_first
state_dict = lstm.state_dict()
w_ih = state_dict["weight_ih_l0"]
w_hh = state_dict["weight_hh_l0"]
b_ih = state_dict["bias_ih_l0"]
b_hh = state_dict["bias_hh_l0"]
w_hr = state_dict["weight_hr_l0"]
self.input_size = lstm.input_size
self.proj_size = lstm.proj_size
self.hidden_size = lstm.hidden_size
self.w_ih = w_ih
self.w_hh = w_hh
self.b = b_ih + b_hh
self.w_hr = w_hr
def forward(
self,
input: torch.Tensor,
hx: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
"""
Args:
input:
A tensor of shape [T, N, hidden_size]
hx:
A tuple containing:
- h0: a tensor of shape (1, N, proj_size)
- c0: a tensor of shape (1, N, hidden_size)
Returns:
Return a tuple containing:
- output: a tensor of shape (T, N, proj_size).
- A tuple containing:
- h: a tensor of shape (1, N, proj_size)
- c: a tensor of shape (1, N, hidden_size)
"""
x_list = input.unbind(dim=0) # We use batch_first=False
if hx is not None:
h0, c0 = hx
else:
h0 = torch.zeros(1, input.size(1), self.proj_size)
c0 = torch.zeros(1, input.size(1), self.hidden_size)
h0 = h0.squeeze(0)
c0 = c0.squeeze(0)
y_list = []
for x in x_list:
gates = F.linear(x, self.w_ih, self.b) + F.linear(h0, self.w_hh)
i, f, g, o = gates.chunk(4, dim=1)
i = i.sigmoid()
f = f.sigmoid()
g = g.tanh()
o = o.sigmoid()
c = f * c0 + i * g
h = o * c.tanh()
h = F.linear(h, self.w_hr)
y_list.append(h)
c0 = c
h0 = h
y = torch.stack(y_list, dim=0)
return y, (h0.unsqueeze(0), c0.unsqueeze(0))

207
egs/reazonspeech/ASR/pruned_transducer_stateless2/model.py
View File

@ -1,207 +0,0 @@
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang)
#
# 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.
from typing import Tuple
import k2
import torch
import torch.nn as nn
from encoder_interface import EncoderInterface
from scaling import ScaledLinear
from icefall.utils import add_sos
class Transducer(nn.Module):
"""It implements https://arxiv.org/pdf/1211.3711.pdf
"Sequence Transduction with Recurrent Neural Networks"
"""
def __init__(
self,
encoder: EncoderInterface,
decoder: nn.Module,
joiner: nn.Module,
encoder_dim: int,
decoder_dim: int,
joiner_dim: int,
vocab_size: int,
):
"""
Args:
encoder:
It is the transcription network in the paper. Its accepts
two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
It returns two tensors: `logits` of shape (N, T, encoder_dm) and
`logit_lens` of shape (N,).
decoder:
It is the prediction network in the paper. Its input shape
is (N, U) and its output shape is (N, U, decoder_dim).
It should contain one attribute: `blank_id`.
joiner:
It has two inputs with shapes: (N, T, encoder_dim) and
(N, U, decoder_dim).
Its output shape is (N, T, U, vocab_size). Note that its output
contains unnormalized probs, i.e., not processed by log-softmax.
"""
super().__init__()
assert isinstance(encoder, EncoderInterface), type(encoder)
assert hasattr(decoder, "blank_id")
self.encoder = encoder
self.decoder = decoder
self.joiner = joiner
self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
def forward(
self,
x: torch.Tensor,
x_lens: torch.Tensor,
y: k2.RaggedTensor,
prune_range: int = 5,
am_scale: float = 0.0,
lm_scale: float = 0.0,
warmup: float = 1.0,
reduction: str = "sum",
delay_penalty: float = 0.0,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Args:
x:
A 3-D tensor of shape (N, T, C).
x_lens:
A 1-D tensor of shape (N,). It contains the number of frames in `x`
before padding.
y:
A ragged tensor with 2 axes [utt][label]. It contains labels of each
utterance.
prune_range:
The prune range for rnnt loss, it means how many symbols(context)
we are considering for each frame to compute the loss.
am_scale:
The scale to smooth the loss with am (output of encoder network)
part
lm_scale:
The scale to smooth the loss with lm (output of predictor network)
part
warmup:
A value warmup >= 0 that determines which modules are active, values
warmup > 1 "are fully warmed up" and all modules will be active.
reduction:
"sum" to sum the losses over all utterances in the batch.
"none" to return the loss in a 1-D tensor for each utterance
in the batch.
delay_penalty:
A constant value used to penalize symbol delay, to encourage
streaming models to emit symbols earlier.
See https://github.com/k2-fsa/k2/issues/955 and
https://arxiv.org/pdf/2211.00490.pdf for more details.
Returns:
Returns:
Return the transducer loss.
Note:
Regarding am_scale & lm_scale, it will make the loss-function one of
the form:
lm_scale * lm_probs + am_scale * am_probs +
(1-lm_scale-am_scale) * combined_probs
"""
assert reduction in ("sum", "none"), reduction
assert x.ndim == 3, x.shape
assert x_lens.ndim == 1, x_lens.shape
assert y.num_axes == 2, y.num_axes
assert x.size(0) == x_lens.size(0) == y.dim0
encoder_out, x_lens = self.encoder(x, x_lens, warmup=warmup)
assert torch.all(x_lens > 0)
# Now for the decoder, i.e., the prediction network
row_splits = y.shape.row_splits(1)
y_lens = row_splits[1:] - row_splits[:-1]
blank_id = self.decoder.blank_id
sos_y = add_sos(y, sos_id=blank_id)
# sos_y_padded: [B, S + 1], start with SOS.
sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
# decoder_out: [B, S + 1, decoder_dim]
decoder_out = self.decoder(sos_y_padded)
# Note: y does not start with SOS
# y_padded : [B, S]
y_padded = y.pad(mode="constant", padding_value=0)
y_padded = y_padded.to(torch.int64)
boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
boundary[:, 2] = y_lens
boundary[:, 3] = x_lens
lm = self.simple_lm_proj(decoder_out)
am = self.simple_am_proj(encoder_out)
with torch.cuda.amp.autocast(enabled=False):
simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
lm=lm.float(),
am=am.float(),
symbols=y_padded,
termination_symbol=blank_id,
lm_only_scale=lm_scale,
am_only_scale=am_scale,
boundary=boundary,
reduction=reduction,
delay_penalty=delay_penalty,
return_grad=True,
)
# ranges : [B, T, prune_range]
ranges = k2.get_rnnt_prune_ranges(
px_grad=px_grad,
py_grad=py_grad,
boundary=boundary,
s_range=prune_range,
)
# am_pruned : [B, T, prune_range, encoder_dim]
# lm_pruned : [B, T, prune_range, decoder_dim]
am_pruned, lm_pruned = k2.do_rnnt_pruning(
am=self.joiner.encoder_proj(encoder_out),
lm=self.joiner.decoder_proj(decoder_out),
ranges=ranges,
)
# logits : [B, T, prune_range, vocab_size]
# project_input=False since we applied the decoder's input projections
# prior to do_rnnt_pruning (this is an optimization for speed).
logits = self.joiner(am_pruned, lm_pruned, project_input=False)
with torch.cuda.amp.autocast(enabled=False):
pruned_loss = k2.rnnt_loss_pruned(
logits=logits.float(),
symbols=y_padded,
ranges=ranges,
termination_symbol=blank_id,
boundary=boundary,
delay_penalty=delay_penalty,
reduction=reduction,
)
return (simple_loss, pruned_loss)

303
egs/reazonspeech/ASR/pruned_transducer_stateless2/onnx_check.py
View File

@ -1,303 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This script checks that exported onnx models produce the same output
with the given torchscript model for the same input.
Usage:
./pruned_transducer_stateless2/onnx_check.py \
--jit-filename ./t/cpu_jit.pt \
--onnx-encoder-filename ./t/encoder.onnx \
--onnx-decoder-filename ./t/decoder.onnx \
--onnx-joiner-filename ./t/joiner.onnx \
--onnx-joiner-encoder-proj-filename ./t/joiner_encoder_proj.onnx \
--onnx-joiner-decoder-proj-filename ./t/joiner_decoder_proj.onnx
You can generate cpu_jit.pt, encoder.onnx, decoder.onnx, and other
xxx.onnx files using ./export.py
We provide pretrained models at:
https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2/tree/main/exp
"""
import argparse
import logging
from icefall import is_module_available
if not is_module_available("onnxruntime"):
raise ValueError("Please 'pip install onnxruntime' first.")
import onnxruntime as ort
import torch
ort.set_default_logger_severity(3)
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 exported by torch.jit.script",
)
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",
)
parser.add_argument(
"--onnx-joiner-encoder-proj-filename",
required=True,
type=str,
help="Path to the onnx joiner encoder projection model",
)
parser.add_argument(
"--onnx-joiner-decoder-proj-filename",
required=True,
type=str,
help="Path to the onnx joiner decoder projection model",
)
return parser
def test_encoder(
model: torch.jit.ScriptModule,
encoder_session: ort.InferenceSession,
):
inputs = encoder_session.get_inputs()
outputs = encoder_session.get_outputs()
input_names = [n.name for n in inputs]
output_names = [n.name for n in outputs]
assert inputs[0].shape == ["N", "T", 80]
assert inputs[1].shape == ["N"]
for N in [1, 5]:
for T in [12, 25]:
print("N, T", N, T)
x = torch.rand(N, T, 80, dtype=torch.float32)
x_lens = torch.randint(low=10, high=T + 1, size=(N,))
x_lens[0] = T
encoder_inputs = {
input_names[0]: x.numpy(),
input_names[1]: x_lens.numpy(),
}
encoder_out, encoder_out_lens = encoder_session.run(
output_names,
encoder_inputs,
)
torch_encoder_out, torch_encoder_out_lens = model.encoder(x, x_lens)
encoder_out = torch.from_numpy(encoder_out)
assert torch.allclose(encoder_out, torch_encoder_out, atol=1e-05), (
(encoder_out - torch_encoder_out).abs().max(),
encoder_out.shape,
torch_encoder_out.shape,
)
def test_decoder(
model: torch.jit.ScriptModule,
decoder_session: ort.InferenceSession,
):
inputs = decoder_session.get_inputs()
outputs = decoder_session.get_outputs()
input_names = [n.name for n in inputs]
output_names = [n.name for n in outputs]
assert inputs[0].shape == ["N", 2]
for N in [1, 5, 10]:
y = torch.randint(low=1, high=500, size=(10, 2))
decoder_inputs = {input_names[0]: y.numpy()}
decoder_out = decoder_session.run(
output_names,
decoder_inputs,
)[0]
decoder_out = torch.from_numpy(decoder_out)
torch_decoder_out = model.decoder(y, need_pad=False)
assert torch.allclose(decoder_out, torch_decoder_out, atol=1e-5), (
(decoder_out - torch_decoder_out).abs().max()
)
def test_joiner(
model: torch.jit.ScriptModule,
joiner_session: ort.InferenceSession,
joiner_encoder_proj_session: ort.InferenceSession,
joiner_decoder_proj_session: ort.InferenceSession,
):
joiner_inputs = joiner_session.get_inputs()
joiner_outputs = joiner_session.get_outputs()
joiner_input_names = [n.name for n in joiner_inputs]
joiner_output_names = [n.name for n in joiner_outputs]
assert joiner_inputs[0].shape == ["N", 512]
assert joiner_inputs[1].shape == ["N", 512]
joiner_encoder_proj_inputs = joiner_encoder_proj_session.get_inputs()
encoder_proj_input_name = joiner_encoder_proj_inputs[0].name
assert joiner_encoder_proj_inputs[0].shape == ["N", 512]
joiner_encoder_proj_outputs = joiner_encoder_proj_session.get_outputs()
encoder_proj_output_name = joiner_encoder_proj_outputs[0].name
joiner_decoder_proj_inputs = joiner_decoder_proj_session.get_inputs()
decoder_proj_input_name = joiner_decoder_proj_inputs[0].name
assert joiner_decoder_proj_inputs[0].shape == ["N", 512]
joiner_decoder_proj_outputs = joiner_decoder_proj_session.get_outputs()
decoder_proj_output_name = joiner_decoder_proj_outputs[0].name
for N in [1, 5, 10]:
encoder_out = torch.rand(N, 512)
decoder_out = torch.rand(N, 512)
projected_encoder_out = torch.rand(N, 512)
projected_decoder_out = torch.rand(N, 512)
joiner_inputs = {
joiner_input_names[0]: projected_encoder_out.numpy(),
joiner_input_names[1]: projected_decoder_out.numpy(),
}
joiner_out = joiner_session.run(joiner_output_names, joiner_inputs)[0]
joiner_out = torch.from_numpy(joiner_out)
torch_joiner_out = model.joiner(
projected_encoder_out,
projected_decoder_out,
project_input=False,
)
assert torch.allclose(joiner_out, torch_joiner_out, atol=1e-5), (
(joiner_out - torch_joiner_out).abs().max()
)
# Now test encoder_proj
joiner_encoder_proj_inputs = {encoder_proj_input_name: encoder_out.numpy()}
joiner_encoder_proj_out = joiner_encoder_proj_session.run(
[encoder_proj_output_name], joiner_encoder_proj_inputs
)[0]
joiner_encoder_proj_out = torch.from_numpy(joiner_encoder_proj_out)
torch_joiner_encoder_proj_out = model.joiner.encoder_proj(encoder_out)
assert torch.allclose(
joiner_encoder_proj_out, torch_joiner_encoder_proj_out, atol=1e-5
), ((joiner_encoder_proj_out - torch_joiner_encoder_proj_out).abs().max())
# Now test decoder_proj
joiner_decoder_proj_inputs = {decoder_proj_input_name: decoder_out.numpy()}
joiner_decoder_proj_out = joiner_decoder_proj_session.run(
[decoder_proj_output_name], joiner_decoder_proj_inputs
)[0]
joiner_decoder_proj_out = torch.from_numpy(joiner_decoder_proj_out)
torch_joiner_decoder_proj_out = model.joiner.decoder_proj(decoder_out)
assert torch.allclose(
joiner_decoder_proj_out, torch_joiner_decoder_proj_out, atol=1e-5
), ((joiner_decoder_proj_out - torch_joiner_decoder_proj_out).abs().max())
@torch.no_grad()
def main():
args = get_parser().parse_args()
logging.info(vars(args))
model = torch.jit.load(args.jit_filename)
options = ort.SessionOptions()
options.inter_op_num_threads = 1
options.intra_op_num_threads = 1
logging.info("Test encoder")
encoder_session = ort.InferenceSession(
args.onnx_encoder_filename,
sess_options=options,
providers=["CPUExecutionProvider"],
)
test_encoder(model, encoder_session)
logging.info("Test decoder")
decoder_session = ort.InferenceSession(
args.onnx_decoder_filename,
sess_options=options,
providers=["CPUExecutionProvider"],
)
test_decoder(model, decoder_session)
logging.info("Test joiner")
joiner_session = ort.InferenceSession(
args.onnx_joiner_filename,
sess_options=options,
providers=["CPUExecutionProvider"],
)
joiner_encoder_proj_session = ort.InferenceSession(
args.onnx_joiner_encoder_proj_filename,
sess_options=options,
providers=["CPUExecutionProvider"],
)
joiner_decoder_proj_session = ort.InferenceSession(
args.onnx_joiner_decoder_proj_filename,
sess_options=options,
providers=["CPUExecutionProvider"],
)
test_joiner(
model,
joiner_session,
joiner_encoder_proj_session,
joiner_decoder_proj_session,
)
logging.info("Finished checking ONNX models")
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()

428
egs/reazonspeech/ASR/pruned_transducer_stateless2/onnx_pretrained.py
View File

@ -1,428 +0,0 @@
#!/usr/bin/env python3
# Copyright 2022 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 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/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
as an example to show how to use this file.
1. Download the pre-trained model
cd egs/wenetspeech/ASR
repo_url=https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
repo=$(basename $repo_url)
pushd $repo
git lfs pull --include "data/lang_char/Linv.pt"
git lfs pull --include "exp/pretrained_epoch_4_avg_1.pt"
git lfs pull --include "exp/cpu_jit_epoch_4_avg_1_torch.1.7.1.pt"
cd exp
ln -s pretrained_epoch_9_avg_1_torch.1.7.1.pt epoch-99.pt
popd
2. Export the model to ONNX
./pruned_transducer_stateless5/export-onnx.py \
--lang-dir $repo/data/lang_char \
--epoch 99 \
--avg 1 \
--use-averaged-model 0 \
--exp-dir $repo/exp \
--num-encoder-layers 24 \
--dim-feedforward 1536 \
--nhead 8 \
--encoder-dim 384 \
--decoder-dim 512 \
--joiner-dim 512
It will generate the following 3 files inside $repo/exp:
- encoder-epoch-99-avg-1.onnx
- decoder-epoch-99-avg-1.onnx
- joiner-epoch-99-avg-1.onnx
3. Run this file
./pruned_transducer_stateless5/onnx_pretrained.py \
--encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
--decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
--joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
--tokens $repo/data/lang_char/tokens.txt \
$repo/test_wavs/DEV_T0000000000.wav \
$repo/test_wavs/DEV_T0000000001.wav \
$repo/test_wavs/DEV_T0000000002.wav
"""
import argparse
import logging
import math
from typing import List, Tuple
import k2
import kaldifeat
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 = 4
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,
providers=["CPUExecutionProvider"],
)
def init_decoder(self, decoder_model_filename: str):
self.decoder = ort.InferenceSession(
decoder_model_filename,
sess_options=self.session_opts,
providers=["CPUExecutionProvider"],
)
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,
providers=["CPUExecutionProvider"],
)
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()

319
egs/reazonspeech/ASR/pruned_transducer_stateless2/optim.py
View File

@ -1,319 +0,0 @@
# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey)
#
# 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.
from typing import List, Optional, Union
import torch
from torch.optim import Optimizer
class Eve(Optimizer):
r"""
Implements Eve algorithm. This is a modified version of AdamW with a special
way of setting the weight-decay / shrinkage-factor, which is designed to make the
rms of the parameters approach a particular target_rms (default: 0.1). This is
for use with networks with 'scaled' versions of modules (see scaling.py), which
will be close to invariant to the absolute scale on the parameter matrix.
The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
Eve is unpublished so far.
Arguments:
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float, optional): learning rate (default: 1e-3)
betas (Tuple[float, float], optional): coefficients used for computing
running averages of gradient and its square (default: (0.9, 0.999))
eps (float, optional): term added to the denominator to improve
numerical stability (default: 1e-8)
weight_decay (float, optional): weight decay coefficient (default: 3e-4;
this value means that the weight would decay significantly after
about 3k minibatches. Is not multiplied by learning rate, but
is conditional on RMS-value of parameter being > target_rms.
target_rms (float, optional): target root-mean-square value of
parameters, if they fall below this we will stop applying weight decay.
.. _Adam\: A Method for Stochastic Optimization:
https://arxiv.org/abs/1412.6980
.. _Decoupled Weight Decay Regularization:
https://arxiv.org/abs/1711.05101
.. _On the Convergence of Adam and Beyond:
https://openreview.net/forum?id=ryQu7f-RZ
"""
def __init__(
self,
params,
lr=1e-3,
betas=(0.9, 0.98),
eps=1e-8,
weight_decay=1e-3,
target_rms=0.1,
):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
if not 0 <= weight_decay <= 0.1:
raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
if not 0 < target_rms <= 10.0:
raise ValueError("Invalid target_rms value: {}".format(target_rms))
defaults = dict(
lr=lr,
betas=betas,
eps=eps,
weight_decay=weight_decay,
target_rms=target_rms,
)
super(Eve, self).__init__(params, defaults)
def __setstate__(self, state):
super(Eve, self).__setstate__(state)
@torch.no_grad()
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
for p in group["params"]:
if p.grad is None:
continue
# Perform optimization step
grad = p.grad
if grad.is_sparse:
raise RuntimeError("AdamW does not support sparse gradients")
state = self.state[p]
# State initialization
if len(state) == 0:
state["step"] = 0
# Exponential moving average of gradient values
state["exp_avg"] = torch.zeros_like(
p, memory_format=torch.preserve_format
)
# Exponential moving average of squared gradient values
state["exp_avg_sq"] = torch.zeros_like(
p, memory_format=torch.preserve_format
)
exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
beta1, beta2 = group["betas"]
state["step"] += 1
bias_correction1 = 1 - beta1 ** state["step"]
bias_correction2 = 1 - beta2 ** state["step"]
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
group["eps"]
)
step_size = group["lr"] / bias_correction1
target_rms = group["target_rms"]
weight_decay = group["weight_decay"]
if p.numel() > 1:
# avoid applying this weight-decay on "scaling factors"
# (which are scalar).
is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
p.mul_(1 - (weight_decay * is_above_target_rms))
p.addcdiv_(exp_avg, denom, value=-step_size)
# Constrain the range of scalar weights
if p.numel() == 1:
p.clamp_(min=-10, max=2)
return loss
class LRScheduler(object):
"""
Base-class for learning rate schedulers where the learning-rate depends on both the
batch and the epoch.
"""
def __init__(self, optimizer: Optimizer, verbose: bool = False):
# Attach optimizer
if not isinstance(optimizer, Optimizer):
raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
self.optimizer = optimizer
self.verbose = verbose
for group in optimizer.param_groups:
group.setdefault("initial_lr", group["lr"])
self.base_lrs = [group["initial_lr"] for group in optimizer.param_groups]
self.epoch = 0
self.batch = 0
def state_dict(self):
"""Returns the state of the scheduler as a :class:`dict`.
It contains an entry for every variable in self.__dict__ which
is not the optimizer.
"""
return {
"base_lrs": self.base_lrs,
"epoch": self.epoch,
"batch": self.batch,
}
def load_state_dict(self, state_dict):
"""Loads the schedulers state.
Args:
state_dict (dict): scheduler state. Should be an object returned
from a call to :meth:`state_dict`.
"""
self.__dict__.update(state_dict)
def get_last_lr(self) -> List[float]:
"""Return last computed learning rate by current scheduler. Will be a list of float."""
return self._last_lr
def get_lr(self):
# Compute list of learning rates from self.epoch and self.batch and
# self.base_lrs; this must be overloaded by the user.
# e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
raise NotImplementedError
def step_batch(self, batch: Optional[int] = None) -> None:
# Step the batch index, or just set it. If `batch` is specified, it
# must be the batch index from the start of training, i.e. summed over
# all epochs.
# You can call this in any order; if you don't provide 'batch', it should
# of course be called once per batch.
if batch is not None:
self.batch = batch
else:
self.batch = self.batch + 1
self._set_lrs()
def step_epoch(self, epoch: Optional[int] = None):
# Step the epoch index, or just set it. If you provide the 'epoch' arg,
# you should call this at the start of the epoch; if you don't provide the 'epoch'
# arg, you should call it at the end of the epoch.
if epoch is not None:
self.epoch = epoch
else:
self.epoch = self.epoch + 1
self._set_lrs()
def _set_lrs(self):
values = self.get_lr()
assert len(values) == len(self.optimizer.param_groups)
for i, data in enumerate(zip(self.optimizer.param_groups, values)):
param_group, lr = data
param_group["lr"] = lr
self.print_lr(self.verbose, i, lr)
self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
def print_lr(self, is_verbose, group, lr):
"""Display the current learning rate."""
if is_verbose:
print(
f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
f" of group {group} to {lr:.4e}."
)
class Eden(LRScheduler):
"""
Eden scheduler.
lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
(((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25))
E.g. suggest initial-lr = 0.003 (passed to optimizer).
Args:
optimizer: the optimizer to change the learning rates on
lr_batches: the number of batches after which we start significantly
decreasing the learning rate, suggest 5000.
lr_epochs: the number of epochs after which we start significantly
decreasing the learning rate, suggest 6 if you plan to do e.g.
20 to 40 epochs, but may need smaller number if dataset is huge
and you will do few epochs.
"""
def __init__(
self,
optimizer: Optimizer,
lr_batches: Union[int, float],
lr_epochs: Union[int, float],
verbose: bool = False,
):
super(Eden, self).__init__(optimizer, verbose)
self.lr_batches = lr_batches
self.lr_epochs = lr_epochs
def get_lr(self):
factor = (
(self.batch**2 + self.lr_batches**2) / self.lr_batches**2
) ** -0.25 * (
((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
)
return [x * factor for x in self.base_lrs]
def _test_eden():
m = torch.nn.Linear(100, 100)
optim = Eve(m.parameters(), lr=0.003)
scheduler = Eden(optim, lr_batches=30, lr_epochs=2, verbose=True)
for epoch in range(10):
scheduler.step_epoch(epoch) # sets epoch to `epoch`
for step in range(20):
x = torch.randn(200, 100).detach()
x.requires_grad = True
y = m(x)
dy = torch.randn(200, 100).detach()
f = (y * dy).sum()
f.backward()
optim.step()
scheduler.step_batch()
optim.zero_grad()
print("last lr = ", scheduler.get_last_lr())
print("state dict = ", scheduler.state_dict())
if __name__ == "__main__":
_test_eden()

339
egs/reazonspeech/ASR/pruned_transducer_stateless2/pretrained.py
View File

@ -1,339 +0,0 @@
#!/usr/bin/env python3
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang)
# 2022 Xiaomi Crop. (authors: Mingshuang Luo)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Usage:
(1) greedy search
./pruned_transducer_stateless2/pretrained.py \
--checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
--lang-dir ./data/lang_char \
--decoding-method greedy_search \
--max-sym-per-frame 1 \
/path/to/foo.wav \
/path/to/bar.wav
(2) modified beam search
./pruned_transducer_stateless2/pretrained.py \
--checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
--lang-dir ./data/lang_char \
--decoding-method modified_beam_search \
--beam-size 4 \
/path/to/foo.wav \
/path/to/bar.wav
(3) fast beam search
./pruned_transducer_stateless2/pretrained.py \
--checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
--lang-dir ./data/lang_char \
--decoding-method fast_beam_search \
--beam 4 \
--max-contexts 4 \
--max-states 8 \
/path/to/foo.wav \
/path/to/bar.wav
You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`.
Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by
./pruned_transducer_stateless2/export.py
"""
import argparse
import logging
import math
from typing import List
import k2
import kaldifeat
import torch
import torchaudio
from beam_search import (
beam_search,
fast_beam_search_one_best,
greedy_search,
greedy_search_batch,
modified_beam_search,
)
from tokenizer import Tokenizer
from torch.nn.utils.rnn import pad_sequence
from train import get_params, get_transducer_model
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--checkpoint",
type=str,
required=True,
help="Path to the checkpoint. "
"The checkpoint is assumed to be saved by "
"icefall.checkpoint.save_checkpoint().",
)
parser.add_argument(
"--lang-dir",
type=str,
help="""Path to lang.
""",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_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="""Used only when --decoding-method is beam_search
and 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 --decoding-method is fast_beam_search""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=4,
help="""Used only when --decoding-method is
fast_beam_search""",
)
parser.add_argument(
"--max-states",
type=int,
default=8,
help="""Used only when --decoding-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
--decoding-method is greedy_search.
""",
)
return parser
def read_sound_files(
filenames: List[str], expected_sample_rate: float
) -> List[torch.Tensor]:
"""Read a list of sound files into a list 1-D float32 torch tensors.
Args:
filenames:
A list of sound filenames.
expected_sample_rate:
The expected sample rate of the sound files.
Returns:
Return a list of 1-D float32 torch tensors.
"""
ans = []
for f in filenames:
wave, sample_rate = torchaudio.load(f)
assert (
sample_rate == expected_sample_rate
), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
# We use only the first channel
ans.append(wave[0])
return ans
@torch.no_grad()
def main():
parser = get_parser()
Tokenizer.add_arguments(parser)
args = parser.parse_args()
params = get_params()
params.update(vars(args))
sp = Tokenizer.load(params.lang_dir, params.lang_type)
# <blk> is defined in local/prepare_lang_char.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
if params.decoding_method == "fast_beam_search":
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
else:
decoding_graph = None
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.decoding_method}"
if params.decoding_method == "beam_search":
msg += f" with beam size {params.beam_size}"
logging.info(msg)
if params.decoding_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.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 == "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.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 method: {params.decoding_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()

1014
egs/reazonspeech/ASR/pruned_transducer_stateless2/scaling.py
View File

File diff suppressed because it is too large Load Diff

320
egs/reazonspeech/ASR/pruned_transducer_stateless2/scaling_converter.py
View File

@ -1,320 +0,0 @@
# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This file provides functions to convert `ScaledLinear`, `ScaledConv1d`,
`ScaledConv2d`, and `ScaledEmbedding` to their non-scaled counterparts:
`nn.Linear`, `nn.Conv1d`, `nn.Conv2d`, and `nn.Embedding`.
The scaled version are required only in the training time. It simplifies our
life by converting them to their non-scaled version during inference.
"""
import copy
import re
from typing import List
import torch
import torch.nn as nn
from lstmp import LSTMP
from scaling import (
ActivationBalancer,
BasicNorm,
ScaledConv1d,
ScaledConv2d,
ScaledEmbedding,
ScaledLinear,
ScaledLSTM,
)
class NonScaledNorm(nn.Module):
"""See BasicNorm for doc"""
def __init__(
self,
num_channels: int,
eps_exp: float,
channel_dim: int = -1, # CAUTION: see documentation.
):
super().__init__()
self.num_channels = num_channels
self.channel_dim = channel_dim
self.eps_exp = eps_exp
def forward(self, x: torch.Tensor) -> torch.Tensor:
if not torch.jit.is_tracing():
assert x.shape[self.channel_dim] == self.num_channels
scales = (
torch.mean(x * x, dim=self.channel_dim, keepdim=True) + self.eps_exp
).pow(-0.5)
return x * scales
def scaled_linear_to_linear(scaled_linear: ScaledLinear) -> nn.Linear:
"""Convert an instance of ScaledLinear to nn.Linear.
Args:
scaled_linear:
The layer to be converted.
Returns:
Return a linear layer. It satisfies:
scaled_linear(x) == linear(x)
for any given input tensor `x`.
"""
assert isinstance(scaled_linear, ScaledLinear), type(scaled_linear)
weight = scaled_linear.get_weight()
bias = scaled_linear.get_bias()
has_bias = bias is not None
linear = torch.nn.Linear(
in_features=scaled_linear.in_features,
out_features=scaled_linear.out_features,
bias=True, # otherwise, it throws errors when converting to PNNX format
# device=weight.device, # Pytorch version before v1.9.0 does not have
# this argument. Comment out for now, we will
# see if it will raise error for versions
# after v1.9.0
)
linear.weight.data.copy_(weight)
if has_bias:
linear.bias.data.copy_(bias)
else:
linear.bias.data.zero_()
return linear
def scaled_conv1d_to_conv1d(scaled_conv1d: ScaledConv1d) -> nn.Conv1d:
"""Convert an instance of ScaledConv1d to nn.Conv1d.
Args:
scaled_conv1d:
The layer to be converted.
Returns:
Return an instance of nn.Conv1d that has the same `forward()` behavior
of the given `scaled_conv1d`.
"""
assert isinstance(scaled_conv1d, ScaledConv1d), type(scaled_conv1d)
weight = scaled_conv1d.get_weight()
bias = scaled_conv1d.get_bias()
has_bias = bias is not None
conv1d = nn.Conv1d(
in_channels=scaled_conv1d.in_channels,
out_channels=scaled_conv1d.out_channels,
kernel_size=scaled_conv1d.kernel_size,
stride=scaled_conv1d.stride,
padding=scaled_conv1d.padding,
dilation=scaled_conv1d.dilation,
groups=scaled_conv1d.groups,
bias=scaled_conv1d.bias is not None,
padding_mode=scaled_conv1d.padding_mode,
)
conv1d.weight.data.copy_(weight)
if has_bias:
conv1d.bias.data.copy_(bias)
return conv1d
def scaled_conv2d_to_conv2d(scaled_conv2d: ScaledConv2d) -> nn.Conv2d:
"""Convert an instance of ScaledConv2d to nn.Conv2d.
Args:
scaled_conv2d:
The layer to be converted.
Returns:
Return an instance of nn.Conv2d that has the same `forward()` behavior
of the given `scaled_conv2d`.
"""
assert isinstance(scaled_conv2d, ScaledConv2d), type(scaled_conv2d)
weight = scaled_conv2d.get_weight()
bias = scaled_conv2d.get_bias()
has_bias = bias is not None
conv2d = nn.Conv2d(
in_channels=scaled_conv2d.in_channels,
out_channels=scaled_conv2d.out_channels,
kernel_size=scaled_conv2d.kernel_size,
stride=scaled_conv2d.stride,
padding=scaled_conv2d.padding,
dilation=scaled_conv2d.dilation,
groups=scaled_conv2d.groups,
bias=scaled_conv2d.bias is not None,
padding_mode=scaled_conv2d.padding_mode,
)
conv2d.weight.data.copy_(weight)
if has_bias:
conv2d.bias.data.copy_(bias)
return conv2d
def scaled_embedding_to_embedding(
scaled_embedding: ScaledEmbedding,
) -> nn.Embedding:
"""Convert an instance of ScaledEmbedding to nn.Embedding.
Args:
scaled_embedding:
The layer to be converted.
Returns:
Return an instance of nn.Embedding that has the same `forward()` behavior
of the given `scaled_embedding`.
"""
assert isinstance(scaled_embedding, ScaledEmbedding), type(scaled_embedding)
embedding = nn.Embedding(
num_embeddings=scaled_embedding.num_embeddings,
embedding_dim=scaled_embedding.embedding_dim,
padding_idx=scaled_embedding.padding_idx,
scale_grad_by_freq=scaled_embedding.scale_grad_by_freq,
sparse=scaled_embedding.sparse,
)
weight = scaled_embedding.weight
scale = scaled_embedding.scale
embedding.weight.data.copy_(weight * scale.exp())
return embedding
def convert_basic_norm(basic_norm: BasicNorm) -> NonScaledNorm:
assert isinstance(basic_norm, BasicNorm), type(BasicNorm)
norm = NonScaledNorm(
num_channels=basic_norm.num_channels,
eps_exp=basic_norm.eps.data.exp().item(),
channel_dim=basic_norm.channel_dim,
)
return norm
def scaled_lstm_to_lstm(scaled_lstm: ScaledLSTM) -> nn.LSTM:
"""Convert an instance of ScaledLSTM to nn.LSTM.
Args:
scaled_lstm:
The layer to be converted.
Returns:
Return an instance of nn.LSTM that has the same `forward()` behavior
of the given `scaled_lstm`.
"""
assert isinstance(scaled_lstm, ScaledLSTM), type(scaled_lstm)
lstm = nn.LSTM(
input_size=scaled_lstm.input_size,
hidden_size=scaled_lstm.hidden_size,
num_layers=scaled_lstm.num_layers,
bias=scaled_lstm.bias,
batch_first=scaled_lstm.batch_first,
dropout=scaled_lstm.dropout,
bidirectional=scaled_lstm.bidirectional,
proj_size=scaled_lstm.proj_size,
)
assert lstm._flat_weights_names == scaled_lstm._flat_weights_names
for idx in range(len(scaled_lstm._flat_weights_names)):
scaled_weight = scaled_lstm._flat_weights[idx] * scaled_lstm._scales[idx].exp()
lstm._flat_weights[idx].data.copy_(scaled_weight)
return lstm
# Copied from https://pytorch.org/docs/1.9.0/_modules/torch/nn/modules/module.html#Module.get_submodule # noqa
# get_submodule was added to nn.Module at v1.9.0
def get_submodule(model, target):
if target == "":
return model
atoms: List[str] = target.split(".")
mod: torch.nn.Module = model
for item in atoms:
if not hasattr(mod, item):
raise AttributeError(
mod._get_name() + " has no " "attribute `" + item + "`"
)
mod = getattr(mod, item)
if not isinstance(mod, torch.nn.Module):
raise AttributeError("`" + item + "` is not " "an nn.Module")
return mod
def convert_scaled_to_non_scaled(
model: nn.Module,
inplace: bool = False,
is_onnx: bool = False,
):
"""Convert `ScaledLinear`, `ScaledConv1d`, and `ScaledConv2d`
in the given modle to their unscaled version `nn.Linear`, `nn.Conv1d`,
and `nn.Conv2d`.
Args:
model:
The model to be converted.
inplace:
If True, the input model is modified inplace.
If False, the input model is copied and we modify the copied version.
is_onnx:
If True, we are going to export the model to ONNX. In this case,
we will convert nn.LSTM with proj_size to LSTMP.
Return:
Return a model without scaled layers.
"""
if not inplace:
model = copy.deepcopy(model)
excluded_patterns = r"(self|src)_attn\.(in|out)_proj"
p = re.compile(excluded_patterns)
d = {}
for name, m in model.named_modules():
if isinstance(m, ScaledLinear):
if p.search(name) is not None:
continue
d[name] = scaled_linear_to_linear(m)
elif isinstance(m, ScaledConv1d):
d[name] = scaled_conv1d_to_conv1d(m)
elif isinstance(m, ScaledConv2d):
d[name] = scaled_conv2d_to_conv2d(m)
elif isinstance(m, ScaledEmbedding):
d[name] = scaled_embedding_to_embedding(m)
elif isinstance(m, BasicNorm):
d[name] = convert_basic_norm(m)
elif isinstance(m, ScaledLSTM):
if is_onnx:
d[name] = LSTMP(scaled_lstm_to_lstm(m))
# See
# https://github.com/pytorch/pytorch/issues/47887
# d[name] = torch.jit.script(LSTMP(scaled_lstm_to_lstm(m)))
else:
d[name] = scaled_lstm_to_lstm(m)
elif isinstance(m, ActivationBalancer):
d[name] = nn.Identity()
for k, v in d.items():
if "." in k:
parent, child = k.rsplit(".", maxsplit=1)
setattr(get_submodule(model, parent), child, v)
else:
setattr(model, k, v)
return model

1
egs/reazonspeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py
View File

@ -1 +0,0 @@
/var/data/share20/qc/k2/Github/icefall/egs/librispeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless2/streaming_decode.py
View File

@ -1 +0,0 @@
/var/data/share20/qc/k2/Github/icefall/egs/librispeech/ASR/pruned_transducer_stateless2/streaming_decode.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless2/test_model.py
View File

@ -1 +0,0 @@
/var/data/share20/qc/k2/Github/icefall/egs/librispeech/ASR/pruned_transducer_stateless2/test_model.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless2/tokenizer.py
View File

@ -1 +0,0 @@
../local/utils/tokenizer.py

1061
egs/reazonspeech/ASR/pruned_transducer_stateless2/train.py
View File

File diff suppressed because it is too large Load Diff

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
View File

@ -1 +0,0 @@
../local/utils/asr_datamodule.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py

846
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decode.py
View File

@ -1,846 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Usage:
(1) greedy search
./pruned_transducer_stateless7_streaming/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--max-duration 600 \
--decode-chunk-len 32 \
--lang data/lang_char \
--decoding-method greedy_search
(2) beam search (not recommended)
./pruned_transducer_stateless7_streaming/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--max-duration 600 \
--decode-chunk-len 32 \
--decoding-method beam_search \
--lang data/lang_char \
--beam-size 4
(3) modified beam search
./pruned_transducer_stateless7_streaming/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--max-duration 600 \
--decode-chunk-len 32 \
--decoding-method modified_beam_search \
--lang data/lang_char \
--beam-size 4
(4) fast beam search (one best)
./pruned_transducer_stateless7_streaming/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--max-duration 600 \
--decode-chunk-len 32 \
--decoding-method fast_beam_search \
--beam 20.0 \
--max-contexts 8 \
--lang data/lang_char \
--max-states 64
(5) fast beam search (nbest)
./pruned_transducer_stateless7_streaming/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--max-duration 600 \
--decode-chunk-len 32 \
--decoding-method fast_beam_search_nbest \
--beam 20.0 \
--max-contexts 8 \
--max-states 64 \
--num-paths 200 \
--lang data/lang_char \
--nbest-scale 0.5
(6) fast beam search (nbest oracle WER)
./pruned_transducer_stateless7_streaming/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--max-duration 600 \
--decode-chunk-len 32 \
--decoding-method fast_beam_search_nbest_oracle \
--beam 20.0 \
--max-contexts 8 \
--max-states 64 \
--num-paths 200 \
--lang data/lang_char \
--nbest-scale 0.5
(7) fast beam search (with LG)
./pruned_transducer_stateless7_streaming/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--max-duration 600 \
--decode-chunk-len 32 \
--decoding-method fast_beam_search_nbest_LG \
--beam 20.0 \
--max-contexts 8 \
--lang data/lang_char \
--max-states 64
"""
import argparse
import logging
import math
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import k2
import torch
import torch.nn as nn
from asr_datamodule import ReazonSpeechAsrDataModule
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,
)
from tokenizer import Tokenizer
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.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(
"--gpu",
type=int,
default=0,
)
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_streaming/exp",
help="The experiment dir",
)
parser.add_argument(
"--res-dir",
type=Path,
default=None,
help="The path to save results.",
)
parser.add_argument(
"--lang-dir",
type=Path,
default="data/lang_char",
help="The lang dir. It should contain at least a word table.",
)
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
If you use fast_beam_search_nbest_LG, you have to specify
`--lang-dir`, which should contain `LG.pt`.
""",
)
parser.add_argument(
"--decoding-graph",
type=str,
default="",
help="""Used only when --decoding-method is
fast_beam_search""",
)
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(
"--pad-feature",
type=int,
default=30,
help="""
Number of frames to pad at the end.
""",
)
add_model_arguments(parser)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
sp: Tokenizer,
batch: dict,
word_table: Optional[k2.SymbolTable] = None,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[List[str]]]:
"""Decode one batch and return the result in a dict. The dict has the
following format:
- key: It indicates the setting used for decoding. For example,
if 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.
Returns:
Return the decoding result. See above description for the format of
the returned dict.
"""
device = next(model.parameters()).device
feature = batch["inputs"]
assert feature.ndim == 3
feature = feature.to(device)
# at entry, feature is (N, T, C)
supervisions = batch["supervisions"]
feature_lens = supervisions["num_frames"].to(device)
if params.pad_feature:
feature_lens += params.pad_feature
feature = torch.nn.functional.pad(
feature,
pad=(0, 0, 0, params.pad_feature),
value=LOG_EPS,
)
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(sp.text2word(hyp))
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(sp.text2word(hyp))
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(sp.text2word(hyp))
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(sp.text2word(hyp))
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(sp.text2word(hyp))
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.text2word(sp.decode(hyp)))
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: Tokenizer,
word_table: Optional[k2.SymbolTable] = None,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
"""Decode dataset.
Args:
dl:
PyTorch's dataloader containing the dataset to decode.
params:
It is returned by :func:`get_params`.
model:
The neural model.
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.
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,
)
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 = sp.text2word(ref_text)
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)
return test_set_wers
@torch.no_grad()
def main():
parser = get_parser()
ReazonSpeechAsrDataModule.add_arguments(parser)
Tokenizer.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",
)
if not params.res_dir:
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}"
params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
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 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", params.gpu)
logging.info(f"Device: {device}")
sp = Tokenizer.load(params.lang, params.lang_type)
# <blk> and <unk> are defined in local/prepare_lang_char.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)
assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
model.encoder.decode_chunk_size,
params.decode_chunk_len,
)
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()
decoding_graph = None
word_table = None
if params.decoding_graph:
decoding_graph = k2.Fsa.from_dict(
torch.load(params.decoding_graph, map_location=device)
)
elif "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)
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
reazonspeech_corpus = ReazonSpeechAsrDataModule(args)
for subdir in ["valid"]:
results_dict = decode_dataset(
dl=reazonspeech_corpus.test_dataloaders(getattr(reazonspeech_corpus, f"{subdir}_cuts")()),
params=params,
model=model,
sp=sp,
decoding_graph=decoding_graph,
)
tot_err = save_results(
params=params,
test_set_name=subdir,
results_dict=results_dict,
)
with (
params.res_dir
/ (
f"{subdir}-{params.decode_chunk_len}_{params.beam_size}"
f"_{params.avg}_{params.epoch}.cer"
)
).open("w") as fout:
if len(tot_err) == 1:
fout.write(f"{tot_err[0][1]}")
else:
fout.write("\n".join(f"{k}\t{v}") for k, v in tot_err)
logging.info("Done!")
if __name__ == "__main__":
main()

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decoder.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decoder.py

1261
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
View File

File diff suppressed because it is too large Load Diff

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py

250
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/export.py
View File

@ -1,250 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This script converts several saved checkpoints
# to a single one using model averaging.
import argparse
import logging
from pathlib import Path
import torch
from scaling_converter import convert_scaled_to_non_scaled
from tokenizer import Tokenizer
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_streaming/exp",
help="""It specifies the directory where all training related
files, e.g., checkpoints, log, etc, are saved
""",
)
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():
parser = get_parser()
Tokenizer.add_arguments(parser)
args = 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 = Tokenizer.load(params.lang, params.lang_type)
# <blk> is defined in local/prepare_lang_char.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/reazonspeech/ASR/pruned_transducer_stateless7_streaming/joiner.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/joiner.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/model.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/model.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/optim.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/optim.py

347
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py
View File

@ -1,347 +0,0 @@
#!/usr/bin/env python3
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This script loads a checkpoint and uses it to decode waves.
You can generate the checkpoint with the following command:
./pruned_transducer_stateless7_streaming/export.py \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--lang data/lang_char \
--epoch 20 \
--avg 10
Usage of this script:
(1) greedy search
./pruned_transducer_stateless7_streaming/pretrained.py \
--checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
--lang data/lang_char \
--method greedy_search \
/path/to/foo.wav \
/path/to/bar.wav
(2) beam search
./pruned_transducer_stateless7_streaming/pretrained.py \
--checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
--lang data/lang_char \
--method beam_search \
--beam-size 4 \
/path/to/foo.wav \
/path/to/bar.wav
(3) modified beam search
./pruned_transducer_stateless7_streaming/pretrained.py \
--checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
--lang data/lang_char \
--method modified_beam_search \
--beam-size 4 \
/path/to/foo.wav \
/path/to/bar.wav
(4) fast beam search
./pruned_transducer_stateless7_streaming/pretrained.py \
--checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
--lang data/lang_char \
--method fast_beam_search \
--beam-size 4 \
/path/to/foo.wav \
/path/to/bar.wav
You can also use `./pruned_transducer_stateless7_streaming/exp/epoch-xx.pt`.
Note: ./pruned_transducer_stateless7_streaming/exp/pretrained.pt is generated by
./pruned_transducer_stateless7_streaming/export.py
"""
import argparse
import logging
import math
from typing import List
import k2
import kaldifeat
import torch
import torchaudio
from beam_search import (
beam_search,
fast_beam_search_one_best,
greedy_search,
greedy_search_batch,
modified_beam_search,
)
from tokenizer import Tokenizer
from torch.nn.utils.rnn import pad_sequence
from train import add_model_arguments, get_params, get_transducer_model
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--checkpoint",
type=str,
required=True,
help="Path to the checkpoint. "
"The checkpoint is assumed to be saved by "
"icefall.checkpoint.save_checkpoint().",
)
parser.add_argument(
"--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()
Tokenizer.add_arguments(parser)
args = parser.parse_args()
params = get_params()
params.update(vars(args))
sp = Tokenizer.load(params.lang, params.lang_type)
# <blk> is defined in local/prepare_lang_char.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/reazonspeech/ASR/pruned_transducer_stateless7_streaming/scaling.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/scaling.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py

597
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
View File

@ -1,597 +0,0 @@
#!/usr/bin/env python3
# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Usage:
./pruned_transducer_stateless7_streaming/streaming_decode.py \
--epoch 28 \
--avg 15 \
--decode-chunk-len 32 \
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
--decoding_method greedy_search \
--lang data/lang_char \
--num-decode-streams 2000
"""
import argparse
import logging
import math
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import k2
import numpy as np
import torch
import torch.nn as nn
from asr_datamodule import ReazonSpeechAsrDataModule
from decode import save_results
from decode_stream import DecodeStream
from kaldifeat import Fbank, FbankOptions
from lhotse import CutSet
from streaming_beam_search import (
fast_beam_search_one_best,
greedy_search,
modified_beam_search,
)
from tokenizer import Tokenizer
from torch.nn.utils.rnn import pad_sequence
from train import add_model_arguments, get_params, get_transducer_model
from zipformer import stack_states, unstack_states
from icefall.checkpoint import (
average_checkpoints,
average_checkpoints_with_averaged_model,
find_checkpoints,
load_checkpoint,
)
from icefall.utils import AttributeDict, setup_logger, str2bool
LOG_EPS = math.log(1e-10)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=28,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 0.
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(
"--gpu",
type=int,
default=0,
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' 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_stateless2/exp",
help="The experiment dir",
)
parser.add_argument(
"--bpe-model",
type=str,
default="data/lang_bpe_500/bpe.model",
help="Path to the BPE model",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="""Supported decoding methods are:
greedy_search
modified_beam_search
fast_beam_search
""",
)
parser.add_argument(
"--decoding-graph",
type=str,
default="",
help="""Used only when --decoding-method is
fast_beam_search""",
)
parser.add_argument(
"--num_active_paths",
type=int,
default=4,
help="""An interger indicating how many candidates we will keep for each
frame. Used only when --decoding-method is modified_beam_search.""",
)
parser.add_argument(
"--beam",
type=float,
default=4.0,
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""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=4,
help="""Used only when --decoding-method is
fast_beam_search""",
)
parser.add_argument(
"--max-states",
type=int,
default=32,
help="""Used only when --decoding-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(
"--num-decode-streams",
type=int,
default=2000,
help="The number of streams that can be decoded parallel.",
)
parser.add_argument(
"--res-dir",
type=Path,
default=None,
help="The path to save results.",
)
add_model_arguments(parser)
return parser
def decode_one_chunk(
params: AttributeDict,
model: nn.Module,
decode_streams: List[DecodeStream],
) -> List[int]:
"""Decode one chunk frames of features for each decode_streams and
return the indexes of finished streams in a List.
Args:
params:
It's the return value of :func:`get_params`.
model:
The neural model.
decode_streams:
A List of DecodeStream, each belonging to a utterance.
Returns:
Return a List containing which DecodeStreams are finished.
"""
device = model.device
features = []
feature_lens = []
states = []
processed_lens = []
for stream in decode_streams:
feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
features.append(feat)
feature_lens.append(feat_len)
states.append(stream.states)
processed_lens.append(stream.done_frames)
feature_lens = torch.tensor(feature_lens, device=device)
features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
# We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
# factor in encoders is 8.
# After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
tail_length = 23
if features.size(1) < tail_length:
pad_length = tail_length - features.size(1)
feature_lens += pad_length
features = torch.nn.functional.pad(
features,
(0, 0, 0, pad_length),
mode="constant",
value=LOG_EPS,
)
states = stack_states(states)
processed_lens = torch.tensor(processed_lens, device=device)
encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
x=features,
x_lens=feature_lens,
states=states,
)
encoder_out = model.joiner.encoder_proj(encoder_out)
if params.decoding_method == "greedy_search":
greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
elif params.decoding_method == "fast_beam_search":
processed_lens = processed_lens + encoder_out_lens
fast_beam_search_one_best(
model=model,
encoder_out=encoder_out,
processed_lens=processed_lens,
streams=decode_streams,
beam=params.beam,
max_states=params.max_states,
max_contexts=params.max_contexts,
)
elif params.decoding_method == "modified_beam_search":
modified_beam_search(
model=model,
streams=decode_streams,
encoder_out=encoder_out,
num_active_paths=params.num_active_paths,
)
else:
raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
states = unstack_states(new_states)
finished_streams = []
for i in range(len(decode_streams)):
decode_streams[i].states = states[i]
decode_streams[i].done_frames += encoder_out_lens[i]
if decode_streams[i].done:
finished_streams.append(i)
return finished_streams
def decode_dataset(
cuts: CutSet,
params: AttributeDict,
model: nn.Module,
sp: Tokenizer,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[List[str], List[str]]]]:
"""Decode dataset.
Args:
cuts:
Lhotse Cutset containing the dataset to decode.
params:
It is returned by :func:`get_params`.
model:
The neural model.
sp:
The BPE model.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
only when --decoding_method is fast_beam_search.
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.
"""
device = model.device
opts = FbankOptions()
opts.device = device
opts.frame_opts.dither = 0
opts.frame_opts.snip_edges = False
opts.frame_opts.samp_freq = 16000
opts.mel_opts.num_bins = 80
log_interval = 50
decode_results = []
# Contain decode streams currently running.
decode_streams = []
for num, cut in enumerate(cuts):
# each utterance has a DecodeStream.
initial_states = model.encoder.get_init_state(device=device)
decode_stream = DecodeStream(
params=params,
cut_id=cut.id,
initial_states=initial_states,
decoding_graph=decoding_graph,
device=device,
)
audio: np.ndarray = cut.load_audio()
# audio.shape: (1, num_samples)
assert len(audio.shape) == 2
assert audio.shape[0] == 1, "Should be single channel"
assert audio.dtype == np.float32, audio.dtype
# The trained model is using normalized samples
assert audio.max() <= 1, "Should be normalized to [-1, 1])"
samples = torch.from_numpy(audio).squeeze(0)
fbank = Fbank(opts)
feature = fbank(samples.to(device))
decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
decode_stream.ground_truth = cut.supervisions[0].custom[params.transcript_mode]
decode_streams.append(decode_stream)
while len(decode_streams) >= params.num_decode_streams:
finished_streams = decode_one_chunk(
params=params, model=model, decode_streams=decode_streams
)
for i in sorted(finished_streams, reverse=True):
decode_results.append(
(
decode_streams[i].id,
sp.text2word(decode_streams[i].ground_truth),
sp.text2word(sp.decode(decode_streams[i].decoding_result())),
)
)
del decode_streams[i]
if num % log_interval == 0:
logging.info(f"Cuts processed until now is {num}.")
# decode final chunks of last sequences
while len(decode_streams):
finished_streams = decode_one_chunk(
params=params, model=model, decode_streams=decode_streams
)
for i in sorted(finished_streams, reverse=True):
decode_results.append(
(
decode_streams[i].id,
sp.text2word(decode_streams[i].ground_truth),
sp.text2word(sp.decode(decode_streams[i].decoding_result())),
)
)
del decode_streams[i]
if params.decoding_method == "greedy_search":
key = "greedy_search"
elif params.decoding_method == "fast_beam_search":
key = (
f"beam_{params.beam}_"
f"max_contexts_{params.max_contexts}_"
f"max_states_{params.max_states}"
)
elif params.decoding_method == "modified_beam_search":
key = f"num_active_paths_{params.num_active_paths}"
else:
raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
return {key: decode_results}
@torch.no_grad()
def main():
parser = get_parser()
ReazonSpeechAsrDataModule.add_arguments(parser)
Tokenizer.add_arguments(parser)
args = parser.parse_args()
args.exp_dir = Path(args.exp_dir)
params = get_params()
params.update(vars(args))
if not params.res_dir:
params.res_dir = params.exp_dir / "streaming" / 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}"
# for streaming
params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
# for fast_beam_search
if params.decoding_method == "fast_beam_search":
params.suffix += f"-beam-{params.beam}"
params.suffix += f"-max-contexts-{params.max_contexts}"
params.suffix += f"-max-states-{params.max_states}"
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", params.gpu)
logging.info(f"Device: {device}")
sp = Tokenizer.load(params.lang, params.lang_type)
# <blk> and <unk> is defined in local/prepare_lang_char.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 start >= 0:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
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()
model.device = device
decoding_graph = None
if params.decoding_graph:
decoding_graph = k2.Fsa.from_dict(
torch.load(params.decoding_graph, map_location=device)
)
elif params.decoding_method == "fast_beam_search":
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
args.return_cuts = True
reazonspeech_corpus = ReazonSpeechAsrDataModule(args)
for subdir in ["valid"]:
results_dict = decode_dataset(
cuts=getattr(reazonspeech_corpus, f"{subdir}_cuts")(),
params=params,
model=model,
sp=sp,
decoding_graph=decoding_graph,
)
tot_err = save_results(
params=params, test_set_name=subdir, results_dict=results_dict
)
with (
params.res_dir
/ (
f"{subdir}-{params.decode_chunk_len}"
f"_{params.avg}_{params.epoch}.cer"
)
).open("w") as fout:
if len(tot_err) == 1:
fout.write(f"{tot_err[0][1]}")
else:
fout.write("\n".join(f"{k}\t{v}") for k, v in tot_err)
logging.info("Done!")
if __name__ == "__main__":
main()

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/tokenizer.py
View File

@ -1 +0,0 @@
../local/utils/tokenizer.py

1259
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/train.py
View File

File diff suppressed because it is too large Load Diff

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py

1
egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py