mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-06 15:44:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

removed recipes other than AISHELL-1

Browse Source
This commit is contained in:
jinzr 2023-10-24 09:25:55 +08:00
parent bd7bc092b3
commit 94aa178bb9
82 changed files with 0 additions and 10131 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

0
egs/aidatatang_200zh/ASR/zipformer/__init__.py
View File

1
egs/aidatatang_200zh/ASR/zipformer/asr_datamodule.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless2/asr_datamodule.py

1
egs/aidatatang_200zh/ASR/zipformer/beam_search.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless2/beam_search.py

814
egs/aidatatang_200zh/ASR/zipformer/decode.py
View File

@ -1,814 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao
# 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
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method greedy_search
(2) modified beam search
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method modified_beam_search \
--beam-size 4
(3) fast beam search (trivial_graph)
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method fast_beam_search \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(4) fast beam search (LG)
./zipformer/decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method fast_beam_search_LG \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(5) fast beam search (nbest oracle WER)
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/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
"""
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 Aidatatang_200zhAsrDataModule
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 lhotse.cut import Cut
from train import add_model_arguments, get_model, get_params
from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
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,
make_pad_mask,
setup_logger,
store_transcripts,
str2bool,
write_error_stats,
)
LOG_EPS = math.log(1e-10)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=30,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="zipformer/exp",
help="The experiment dir",
)
parser.add_argument(
"--lang-dir",
type=Path,
default="data/lang_char",
help="The lang dir containing word table and LG graph",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_search
- modified_beam_search
- fast_beam_search
- fast_beam_search_LG
- fast_beam_search_nbest_oracle
If you use fast_beam_search_LG, you have to specify
`--lang-dir`, which should contain `LG.pt`.
""",
)
parser.add_argument(
"--beam-size",
type=int,
default=4,
help="""An integer indicating how many candidates we will keep for each
frame. Used only when --decoding-method is beam_search or
modified_beam_search.""",
)
parser.add_argument(
"--beam",
type=float,
default=20.0,
help="""A floating point value to calculate the cutoff score during beam
search (i.e., `cutoff = max-score - beam`), which is the same as the
`beam` in Kaldi.
Used only when --decoding-method is fast_beam_search,
fast_beam_search, fast_beam_search_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_LG.
It specifies the scale for n-gram LM scores.
""",
)
parser.add_argument(
"--ilme-scale",
type=float,
default=0.2,
help="""
Used only when --decoding_method is fast_beam_search_LG.
It specifies the scale for the internal language model estimation.
""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=8,
help="""Used only when --decoding-method is
fast_beam_search, fast_beam_search, fast_beam_search_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, fast_beam_search_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_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 and fast_beam_search_nbest_oracle""",
)
parser.add_argument(
"--blank-penalty",
type=float,
default=0.0,
help="""
The penalty applied on blank symbol during decoding.
Note: It is a positive value that would be applied to logits like
this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
[batch_size, vocab] and blank id is 0).
""",
)
add_model_arguments(parser)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
lexicon: Lexicon,
graph_compiler: CharCtcTrainingGraphCompiler,
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.
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 LG, 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.causal:
# this seems to cause insertions at the end of the utterance if used with zipformer.
pad_len = 30
feature_lens += pad_len
feature = torch.nn.functional.pad(
feature,
pad=(0, 0, 0, pad_len),
value=LOG_EPS,
)
x, x_lens = model.encoder_embed(feature, feature_lens)
src_key_padding_mask = make_pad_mask(x_lens)
x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C)
encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask)
encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C)
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,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
elif params.decoding_method == "fast_beam_search_LG":
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,
blank_penalty=params.blank_penalty,
ilme_scale=params.ilme_scale,
)
for hyp in hyp_tokens:
sentence = "".join([lexicon.word_table[i] for i in hyp])
hyps.append(list(sentence))
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=graph_compiler.texts_to_ids(supervisions["text"]),
nbest_scale=params.nbest_scale,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
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,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
elif params.decoding_method == "modified_beam_search":
hyp_tokens = modified_beam_search(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
blank_penalty=params.blank_penalty,
beam=params.beam_size,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
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,
blank_penalty=params.blank_penalty,
)
elif params.decoding_method == "beam_search":
hyp = beam_search(
model=model,
encoder_out=encoder_out_i,
beam=params.beam_size,
blank_penalty=params.blank_penalty,
)
else:
raise ValueError(
f"Unsupported decoding method: {params.decoding_method}"
)
hyps.append([lexicon.token_table[idx] for idx in hyp])
key = f"blank_penalty_{params.blank_penalty}"
if params.decoding_method == "greedy_search":
return {"greedy_search_" + key: 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"_ilme_scale_{params.ilme_scale}"
key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
return {key: hyps}
else:
return {f"beam_size_{params.beam_size}_" + key: hyps}
def decode_dataset(
dl: torch.utils.data.DataLoader,
params: AttributeDict,
model: nn.Module,
lexicon: Lexicon,
graph_compiler: CharCtcTrainingGraphCompiler,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[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.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or LG, 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"]
texts = [list("".join(text.split())) for text in texts]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
hyps_dict = decode_one_batch(
params=params,
model=model,
lexicon=lexicon,
graph_compiler=graph_compiler,
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):
this_batch.append((cut_id, ref_text, 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[List[int], List[int]]]],
):
test_set_wers = dict()
for key, results in results_dict.items():
recog_path = (
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
)
results = sorted(results)
store_transcripts(filename=recog_path, texts=results)
logging.info(f"The transcripts are stored in {recog_path}")
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = (
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_filename, "w") as f:
wer = write_error_stats(
f, f"{test_set_name}-{key}", results, 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}-{key}-{params.suffix}.txt"
)
with open(errs_info, "w") as f:
print("settings\tWER", file=f)
for key, val in test_set_wers:
print("{}\t{}".format(key, val), file=f)
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
note = "\tbest for {}".format(test_set_name)
for key, val in test_set_wers:
s += "{}\t{}{}\n".format(key, val, note)
note = ""
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
Aidatatang_200zhAsrDataModule.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",
"modified_beam_search",
"fast_beam_search",
"fast_beam_search_LG",
"fast_beam_search_nbest_oracle",
)
params.res_dir = params.exp_dir / params.decoding_method
if params.iter > 0:
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
else:
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
if params.causal:
assert (
"," not in params.chunk_size
), "chunk_size should be one value in decoding."
assert (
"," not in params.left_context_frames
), "left_context_frames should be one value in decoding."
params.suffix += f"-chunk-{params.chunk_size}"
params.suffix += f"-left-context-{params.left_context_frames}"
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"_ilme_scale_{params.ilme_scale}"
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}"
params.suffix += f"-blank-penalty-{params.blank_penalty}"
if params.use_averaged_model:
params.suffix += "-use-averaged-model"
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
logging.info("Decoding started")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"Device: {device}")
lexicon = Lexicon(params.lang_dir)
params.blank_id = lexicon.token_table["<blk>"]
params.vocab_size = max(lexicon.tokens) + 1
graph_compiler = CharCtcTrainingGraphCompiler(
lexicon=lexicon,
device=device,
)
logging.info(params)
logging.info("About to create model")
model = get_model(params)
if not params.use_averaged_model:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
elif params.avg == 1:
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if i >= 1:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
else:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg + 1
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg + 1:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
filename_start = filenames[-1]
filename_end = filenames[0]
logging.info(
"Calculating the averaged model over iteration checkpoints"
f" from {filename_start} (excluded) to {filename_end}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
else:
assert params.avg > 0, params.avg
start = params.epoch - params.avg
assert start >= 1, start
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
logging.info(
f"Calculating the averaged model over epoch range from "
f"{start} (excluded) to {params.epoch}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
model.to(device)
model.eval()
if "fast_beam_search" in params.decoding_method:
if "LG" in params.decoding_method:
lexicon = Lexicon(params.lang_dir)
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
aidatatang_200zh = Aidatatang_200zhAsrDataModule(args)
def remove_short_utt(c: Cut):
T = ((c.num_frames - 7) // 2 + 1) // 2
if T <= 0:
logging.warning(
f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
)
return T > 0
dev_cuts = aidatatang_200zh.valid_cuts()
dev_cuts = dev_cuts.filter(remove_short_utt)
dev_dl = aidatatang_200zh.valid_dataloaders(dev_cuts)
test_cuts = aidatatang_200zh.test_cuts()
test_cuts = test_cuts.filter(remove_short_utt)
test_dl = aidatatang_200zh.test_dataloaders(test_cuts)
test_sets = ["valid_cuts", "test_cuts"]
test_dls = [dev_dl, test_dl]
for test_set, test_dl in zip(test_sets, test_dls):
results_dict = decode_dataset(
dl=test_dl,
params=params,
model=model,
lexicon=lexicon,
graph_compiler=graph_compiler,
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/aidatatang_200zh/ASR/zipformer/decode_stream.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/decoder.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/encoder_interface.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless2/encoder_interface.py

1
egs/aidatatang_200zh/ASR/zipformer/export-onnx-streaming.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/zipformer/export-onnx-streaming.py

1
egs/aidatatang_200zh/ASR/zipformer/export-onnx.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/export.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/jit_pretrained.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/jit_pretrained_streaming.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/joiner.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/model.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/onnx_check.py
View File

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

330
egs/aidatatang_200zh/ASR/zipformer/onnx_decode.py
View File

@ -1,330 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao,
# Xiaoyu Yang,
# 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.
"""
This script loads ONNX exported models and uses them to decode the test sets.
We use the pre-trained model from
https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615
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/pkufool/icefall-asr-zipformer-wenetspeech-20230615
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
repo=$(basename $repo_url)
pushd $repo
git lfs pull --include "data/lang_char/tokens.txt"
git lfs pull --include "exp/pretrained.pt"
cd exp
ln -s pretrained.pt epoch-9999.pt
popd
2. Export the model to ONNX
./zipformer/export-onnx.py \
--tokens $repo/data/lang_char/tokens.txt \
--epoch 9999 \
--avg 1 \
--exp-dir $repo/exp/
It will generate the following 3 files inside $repo/exp:
- encoder-epoch-9999-avg-1.onnx
- decoder-epoch-9999-avg-1.onnx
- joiner-epoch-9999-avg-1.onnx
2. Run this file
./zipformer/onnx_decode.py \
--exp-dir ./zipformer/exp \
--max-duration 600 \
--encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
--decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
--joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
"""
import argparse
import logging
import time
from pathlib import Path
from typing import List, Tuple
import k2
import torch
import torch.nn as nn
from asr_datamodule import Aidatatang_200zhAsrDataModule
from lhotse.cut import Cut
from onnx_pretrained import OnnxModel, greedy_search
from icefall.utils import setup_logger, store_transcripts, write_error_stats
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(
"--exp-dir",
type=str,
default="zipformer/exp",
help="The experiment dir",
)
parser.add_argument(
"--tokens",
type=str,
default="data/lang_char/tokens.txt",
help="Path to the tokens.txt",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="Valid values are greedy_search and modified_beam_search",
)
return parser
def decode_one_batch(
model: OnnxModel, token_table: k2.SymbolTable, batch: dict
) -> List[List[str]]:
"""Decode one batch and return the result.
Currently it only greedy_search is supported.
Args:
model:
The neural model.
token_table:
Mapping ids to tokens.
batch:
It is the return value from iterating
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
for the format of the `batch`.
Returns:
Return the decoded results for each utterance.
"""
feature = batch["inputs"]
assert feature.ndim == 3
# at entry, feature is (N, T, C)
supervisions = batch["supervisions"]
feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
hyps = greedy_search(
model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
)
hyps = [[token_table[h] for h in hyp] for hyp in hyps]
return hyps
def decode_dataset(
dl: torch.utils.data.DataLoader,
model: nn.Module,
token_table: k2.SymbolTable,
) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
"""Decode dataset.
Args:
dl:
PyTorch's dataloader containing the dataset to decode.
model:
The neural model.
token_table:
Mapping ids to tokens.
Returns:
- A list of tuples. Each tuple contains three elements:
- cut_id,
- reference transcript,
- predicted result.
- The total duration (in seconds) of the dataset.
"""
num_cuts = 0
try:
num_batches = len(dl)
except TypeError:
num_batches = "?"
log_interval = 10
total_duration = 0
results = []
for batch_idx, batch in enumerate(dl):
texts = batch["supervisions"]["text"]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
hyps = decode_one_batch(model=model, token_table=token_table, batch=batch)
this_batch = []
assert len(hyps) == len(texts)
for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
ref_words = list(ref_text)
this_batch.append((cut_id, ref_words, hyp_words))
results.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, total_duration
def save_results(
res_dir: Path,
test_set_name: str,
results: List[Tuple[str, List[str], List[str]]],
):
recog_path = res_dir / f"recogs-{test_set_name}.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 = res_dir / f"errs-{test_set_name}.txt"
with open(errs_filename, "w") as f:
wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
logging.info("Wrote detailed error stats to {}".format(errs_filename))
errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
with open(errs_info, "w") as f:
print("WER", file=f)
print(wer, file=f)
s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
Aidatatang_200zhAsrDataModule.add_arguments(parser)
args = parser.parse_args()
assert (
args.decoding_method == "greedy_search"
), "Only supports greedy_search currently."
res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
setup_logger(f"{res_dir}/log-decode")
logging.info("Decoding started")
device = torch.device("cpu")
logging.info(f"Device: {device}")
token_table = k2.SymbolTable.from_file(args.tokens)
assert token_table[0] == "<blk>"
logging.info(vars(args))
logging.info("About to create model")
model = OnnxModel(
encoder_model_filename=args.encoder_model_filename,
decoder_model_filename=args.decoder_model_filename,
joiner_model_filename=args.joiner_model_filename,
)
# we need cut ids to display recognition results.
args.return_cuts = True
aidatatang_200zh = Aidatatang_200zhAsrDataModule(args)
def remove_short_utt(c: Cut):
T = ((c.num_frames - 7) // 2 + 1) // 2
if T <= 0:
logging.warning(
f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
)
return T > 0
dev_cuts = aidatatang_200zh.valid_cuts()
dev_cuts = dev_cuts.filter(remove_short_utt)
dev_dl = aidatatang_200zh.valid_dataloaders(dev_cuts)
test_cuts = aidatatang_200zh.test_cuts()
test_cuts = test_cuts.filter(remove_short_utt)
test_dl = aidatatang_200zh.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):
start_time = time.time()
results, total_duration = decode_dataset(
dl=test_dl, model=model, token_table=token_table
)
end_time = time.time()
elapsed_seconds = end_time - start_time
rtf = elapsed_seconds / total_duration
logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
logging.info(f"Wave duration: {total_duration:.3f} s")
logging.info(
f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
)
save_results(res_dir=res_dir, test_set_name=test_set, results=results)
logging.info("Done!")
if __name__ == "__main__":
main()

1
egs/aidatatang_200zh/ASR/zipformer/onnx_pretrained-streaming.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/onnx_pretrained.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/optim.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/pretrained.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/scaling.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/scaling_converter.py
View File

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

1
egs/aidatatang_200zh/ASR/zipformer/streaming_beam_search.py
View File

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

880
egs/aidatatang_200zh/ASR/zipformer/streaming_decode.py
View File

@ -1,880 +0,0 @@
#!/usr/bin/env python3
# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang,
# 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:
./zipformer/streaming_decode.py \
--epoch 28 \
--avg 15 \
--causal 1 \
--chunk-size 16 \
--left-context-frames 256 \
--exp-dir ./zipformer/exp \
--decoding-method greedy_search \
--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
from asr_datamodule import Aidatatang_200zhAsrDataModule
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 torch import Tensor, nn
from torch.nn.utils.rnn import pad_sequence
from train import add_model_arguments, get_model, get_params
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,
make_pad_mask,
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=28,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="zipformer/exp",
help="The experiment dir",
)
parser.add_argument(
"--lang-dir",
type=str,
default="data/lang_char",
help="Path to the lang dir(containing lexicon, tokens, etc.)",
)
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(
"--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,
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(
"--blank-penalty",
type=float,
default=0.0,
help="""
The penalty applied on blank symbol during decoding.
Note: It is a positive value that would be applied to logits like
this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
[batch_size, vocab] and blank id is 0).
""",
)
parser.add_argument(
"--num-decode-streams",
type=int,
default=2000,
help="The number of streams that can be decoded parallel.",
)
add_model_arguments(parser)
return parser
def get_init_states(
model: nn.Module,
batch_size: int = 1,
device: torch.device = torch.device("cpu"),
) -> List[torch.Tensor]:
"""
Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
states[-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
states[-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
"""
states = model.encoder.get_init_states(batch_size, device)
embed_states = model.encoder_embed.get_init_states(batch_size, device)
states.append(embed_states)
processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
states.append(processed_lens)
return states
def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
"""Stack list of zipformer states that correspond to separate utterances
into a single emformer state, so that it can be used as an input for
zipformer when those utterances are formed into a batch.
Args:
state_list:
Each element in state_list corresponding to the internal state
of the zipformer model for a single utterance. For element-n,
state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
cached_val2, cached_conv1, cached_conv2).
state_list[n][-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
state_list[n][-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
Note:
It is the inverse of :func:`unstack_states`.
"""
batch_size = len(state_list)
assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
tot_num_layers = (len(state_list[0]) - 2) // 6
batch_states = []
for layer in range(tot_num_layers):
layer_offset = layer * 6
# cached_key: (left_context_len, batch_size, key_dim)
cached_key = torch.cat(
[state_list[i][layer_offset] for i in range(batch_size)], dim=1
)
# cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
cached_nonlin_attn = torch.cat(
[state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
)
# cached_val1: (left_context_len, batch_size, value_dim)
cached_val1 = torch.cat(
[state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
)
# cached_val2: (left_context_len, batch_size, value_dim)
cached_val2 = torch.cat(
[state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
)
# cached_conv1: (#batch, channels, left_pad)
cached_conv1 = torch.cat(
[state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
)
# cached_conv2: (#batch, channels, left_pad)
cached_conv2 = torch.cat(
[state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
)
batch_states += [
cached_key,
cached_nonlin_attn,
cached_val1,
cached_val2,
cached_conv1,
cached_conv2,
]
cached_embed_left_pad = torch.cat(
[state_list[i][-2] for i in range(batch_size)], dim=0
)
batch_states.append(cached_embed_left_pad)
processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
batch_states.append(processed_lens)
return batch_states
def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
"""Unstack the zipformer state corresponding to a batch of utterances
into a list of states, where the i-th entry is the state from the i-th
utterance in the batch.
Note:
It is the inverse of :func:`stack_states`.
Args:
batch_states: A list of cached tensors of all encoder layers. For layer-i,
states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
cached_conv1, cached_conv2).
state_list[-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
states[-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
Returns:
state_list: A list of list. Each element in state_list corresponding to the internal state
of the zipformer model for a single utterance.
"""
assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
tot_num_layers = (len(batch_states) - 2) // 6
processed_lens = batch_states[-1]
batch_size = processed_lens.shape[0]
state_list = [[] for _ in range(batch_size)]
for layer in range(tot_num_layers):
layer_offset = layer * 6
# cached_key: (left_context_len, batch_size, key_dim)
cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
# cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
chunks=batch_size, dim=1
)
# cached_val1: (left_context_len, batch_size, value_dim)
cached_val1_list = batch_states[layer_offset + 2].chunk(
chunks=batch_size, dim=1
)
# cached_val2: (left_context_len, batch_size, value_dim)
cached_val2_list = batch_states[layer_offset + 3].chunk(
chunks=batch_size, dim=1
)
# cached_conv1: (#batch, channels, left_pad)
cached_conv1_list = batch_states[layer_offset + 4].chunk(
chunks=batch_size, dim=0
)
# cached_conv2: (#batch, channels, left_pad)
cached_conv2_list = batch_states[layer_offset + 5].chunk(
chunks=batch_size, dim=0
)
for i in range(batch_size):
state_list[i] += [
cached_key_list[i],
cached_nonlin_attn_list[i],
cached_val1_list[i],
cached_val2_list[i],
cached_conv1_list[i],
cached_conv2_list[i],
]
cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
for i in range(batch_size):
state_list[i].append(cached_embed_left_pad_list[i])
processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
for i in range(batch_size):
state_list[i].append(processed_lens_list[i])
return state_list
def streaming_forward(
features: Tensor,
feature_lens: Tensor,
model: nn.Module,
states: List[Tensor],
chunk_size: int,
left_context_len: int,
) -> Tuple[Tensor, Tensor, List[Tensor]]:
"""
Returns encoder outputs, output lengths, and updated states.
"""
cached_embed_left_pad = states[-2]
(x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
x=features,
x_lens=feature_lens,
cached_left_pad=cached_embed_left_pad,
)
assert x.size(1) == chunk_size, (x.size(1), chunk_size)
src_key_padding_mask = make_pad_mask(x_lens)
# processed_mask is used to mask out initial states
processed_mask = torch.arange(left_context_len, device=x.device).expand(
x.size(0), left_context_len
)
processed_lens = states[-1] # (batch,)
# (batch, left_context_size)
processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
# Update processed lengths
new_processed_lens = processed_lens + x_lens
# (batch, left_context_size + chunk_size)
src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C)
encoder_states = states[:-2]
(
encoder_out,
encoder_out_lens,
new_encoder_states,
) = model.encoder.streaming_forward(
x=x,
x_lens=x_lens,
states=encoder_states,
src_key_padding_mask=src_key_padding_mask,
)
encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C)
new_states = new_encoder_states + [
new_cached_embed_left_pad,
new_processed_lens,
]
return encoder_out, encoder_out_lens, new_states
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
chunk_size = int(params.chunk_size)
left_context_len = int(params.left_context_frames)
features = []
feature_lens = []
states = []
processed_lens = [] # Used in fast-beam-search
for stream in decode_streams:
feat, feat_len = stream.get_feature_frames(chunk_size * 2)
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)
# Make sure the length after encoder_embed is at least 1.
# The encoder_embed subsample features (T - 7) // 2
# The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
tail_length = chunk_size * 2 + 7 + 2 * 3
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)
encoder_out, encoder_out_lens, new_states = streaming_forward(
features=features,
feature_lens=feature_lens,
model=model,
states=states,
chunk_size=chunk_size,
left_context_len=left_context_len,
)
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,
blank_penalty=params.blank_penalty,
)
elif params.decoding_method == "fast_beam_search":
processed_lens = torch.tensor(processed_lens, device=device)
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,
blank_penalty=params.blank_penalty,
)
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,
blank_penalty=params.blank_penalty,
)
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,
lexicon: Lexicon,
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.
lexicon:
The Lexicon.
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 = 100
decode_results = []
# Contain decode streams currently running.
decode_streams = []
for num, cut in enumerate(cuts):
# each utterance has a DecodeStream.
initial_states = get_init_states(model=model, batch_size=1, 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
if audio.max() > 1:
logging.warning(
f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}."
f"Clipping to [-1, 1]."
)
audio = np.clip(audio, -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=30)
decode_stream.ground_truth = cut.supervisions[0].text
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,
list(decode_streams[i].ground_truth.strip()),
[
lexicon.token_table[idx]
for idx in 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,
decode_streams[i].ground_truth.split(),
[
lexicon.token_table[idx]
for idx in decode_streams[i].decoding_result()
],
)
)
del decode_streams[i]
key = f"blank_penalty_{params.blank_penalty}"
if params.decoding_method == "greedy_search":
key = f"greedy_search_{key}"
elif params.decoding_method == "fast_beam_search":
key = (
f"beam_{params.beam}_"
f"max_contexts_{params.max_contexts}_"
f"max_states_{params.max_states}_{key}"
)
elif params.decoding_method == "modified_beam_search":
key = f"num_active_paths_{params.num_active_paths}_{key}"
else:
raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
return {key: decode_results}
def save_results(
params: AttributeDict,
test_set_name: str,
results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
):
test_set_wers = dict()
for key, results in results_dict.items():
recog_path = (
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
)
results = sorted(results)
store_transcripts(filename=recog_path, texts=results)
logging.info(f"The transcripts are stored in {recog_path}")
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = (
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_filename, "w") as f:
wer = write_error_stats(
f, f"{test_set_name}-{key}", results, 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}-{key}-{params.suffix}.txt"
)
with open(errs_info, "w") as f:
print("settings\tWER", file=f)
for key, val in test_set_wers:
print("{}\t{}".format(key, val), file=f)
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
note = "\tbest for {}".format(test_set_name)
for key, val in test_set_wers:
s += "{}\t{}{}\n".format(key, val, note)
note = ""
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
Aidatatang_200zhAsrDataModule.add_arguments(parser)
args = parser.parse_args()
args.exp_dir = Path(args.exp_dir)
params = get_params()
params.update(vars(args))
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}"
assert params.causal, params.causal
assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
assert (
"," not in params.left_context_frames
), "left_context_frames should be one value in decoding."
params.suffix += f"-chunk-{params.chunk_size}"
params.suffix += f"-left-context-{params.left_context_frames}"
params.suffix += f"-blank-penalty-{params.blank_penalty}"
# 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", 0)
logging.info(f"Device: {device}")
lexicon = Lexicon(params.lang_dir)
params.blank_id = lexicon.token_table["<blk>"]
params.vocab_size = max(lexicon.tokens) + 1
logging.info(params)
logging.info("About to create model")
model = get_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_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}")
aidatatang_200zh = Aidatatang_200zhAsrDataModule(args)
dev_cuts = aidatatang_200zh.valid_cuts()
test_cuts = aidatatang_200zh.test_cuts()
test_sets = ["valid_cuts", "test_cuts"]
test_cuts = [dev_cuts, test_cuts]
for test_set, test_cut in zip(test_sets, test_cuts):
results_dict = decode_dataset(
cuts=test_cut,
params=params,
model=model,
lexicon=lexicon,
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/aidatatang_200zh/ASR/zipformer/subsampling.py
View File

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

1350
egs/aidatatang_200zh/ASR/zipformer/train.py
View File

File diff suppressed because it is too large Load Diff

1
egs/aidatatang_200zh/ASR/zipformer/zipformer.py
View File

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

0
egs/aishell2/ASR/zipformer/__init__.py
View File

1
egs/aishell2/ASR/zipformer/asr_datamodule.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless5/asr_datamodule.py

1
egs/aishell2/ASR/zipformer/beam_search.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless5/beam_search.py

814
egs/aishell2/ASR/zipformer/decode.py
View File

@ -1,814 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao
# 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
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method greedy_search
(2) modified beam search
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method modified_beam_search \
--beam-size 4
(3) fast beam search (trivial_graph)
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method fast_beam_search \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(4) fast beam search (LG)
./zipformer/decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method fast_beam_search_LG \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(5) fast beam search (nbest oracle WER)
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/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
"""
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 AiShell2AsrDataModule
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 lhotse.cut import Cut
from train import add_model_arguments, get_model, get_params
from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
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,
make_pad_mask,
setup_logger,
store_transcripts,
str2bool,
write_error_stats,
)
LOG_EPS = math.log(1e-10)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=30,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="zipformer/exp",
help="The experiment dir",
)
parser.add_argument(
"--lang-dir",
type=Path,
default="data/lang_char",
help="The lang dir containing word table and LG graph",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_search
- modified_beam_search
- fast_beam_search
- fast_beam_search_LG
- fast_beam_search_nbest_oracle
If you use fast_beam_search_LG, you have to specify
`--lang-dir`, which should contain `LG.pt`.
""",
)
parser.add_argument(
"--beam-size",
type=int,
default=4,
help="""An integer indicating how many candidates we will keep for each
frame. Used only when --decoding-method is beam_search or
modified_beam_search.""",
)
parser.add_argument(
"--beam",
type=float,
default=20.0,
help="""A floating point value to calculate the cutoff score during beam
search (i.e., `cutoff = max-score - beam`), which is the same as the
`beam` in Kaldi.
Used only when --decoding-method is fast_beam_search,
fast_beam_search, fast_beam_search_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_LG.
It specifies the scale for n-gram LM scores.
""",
)
parser.add_argument(
"--ilme-scale",
type=float,
default=0.2,
help="""
Used only when --decoding_method is fast_beam_search_LG.
It specifies the scale for the internal language model estimation.
""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=8,
help="""Used only when --decoding-method is
fast_beam_search, fast_beam_search, fast_beam_search_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, fast_beam_search_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_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 and fast_beam_search_nbest_oracle""",
)
parser.add_argument(
"--blank-penalty",
type=float,
default=0.0,
help="""
The penalty applied on blank symbol during decoding.
Note: It is a positive value that would be applied to logits like
this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
[batch_size, vocab] and blank id is 0).
""",
)
add_model_arguments(parser)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
lexicon: Lexicon,
graph_compiler: CharCtcTrainingGraphCompiler,
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.
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 LG, 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.causal:
# this seems to cause insertions at the end of the utterance if used with zipformer.
pad_len = 30
feature_lens += pad_len
feature = torch.nn.functional.pad(
feature,
pad=(0, 0, 0, pad_len),
value=LOG_EPS,
)
x, x_lens = model.encoder_embed(feature, feature_lens)
src_key_padding_mask = make_pad_mask(x_lens)
x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C)
encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask)
encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C)
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,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
elif params.decoding_method == "fast_beam_search_LG":
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,
blank_penalty=params.blank_penalty,
ilme_scale=params.ilme_scale,
)
for hyp in hyp_tokens:
sentence = "".join([lexicon.word_table[i] for i in hyp])
hyps.append(list(sentence))
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=graph_compiler.texts_to_ids(supervisions["text"]),
nbest_scale=params.nbest_scale,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
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,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
elif params.decoding_method == "modified_beam_search":
hyp_tokens = modified_beam_search(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
blank_penalty=params.blank_penalty,
beam=params.beam_size,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
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,
blank_penalty=params.blank_penalty,
)
elif params.decoding_method == "beam_search":
hyp = beam_search(
model=model,
encoder_out=encoder_out_i,
beam=params.beam_size,
blank_penalty=params.blank_penalty,
)
else:
raise ValueError(
f"Unsupported decoding method: {params.decoding_method}"
)
hyps.append([lexicon.token_table[idx] for idx in hyp])
key = f"blank_penalty_{params.blank_penalty}"
if params.decoding_method == "greedy_search":
return {"greedy_search_" + key: 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"_ilme_scale_{params.ilme_scale}"
key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
return {key: hyps}
else:
return {f"beam_size_{params.beam_size}_" + key: hyps}
def decode_dataset(
dl: torch.utils.data.DataLoader,
params: AttributeDict,
model: nn.Module,
lexicon: Lexicon,
graph_compiler: CharCtcTrainingGraphCompiler,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[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.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or LG, 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"]
texts = [list("".join(text.split())) for text in texts]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
hyps_dict = decode_one_batch(
params=params,
model=model,
lexicon=lexicon,
graph_compiler=graph_compiler,
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):
this_batch.append((cut_id, ref_text, 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[List[int], List[int]]]],
):
test_set_wers = dict()
for key, results in results_dict.items():
recog_path = (
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
)
results = sorted(results)
store_transcripts(filename=recog_path, texts=results)
logging.info(f"The transcripts are stored in {recog_path}")
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = (
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_filename, "w") as f:
wer = write_error_stats(
f, f"{test_set_name}-{key}", results, 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}-{key}-{params.suffix}.txt"
)
with open(errs_info, "w") as f:
print("settings\tWER", file=f)
for key, val in test_set_wers:
print("{}\t{}".format(key, val), file=f)
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
note = "\tbest for {}".format(test_set_name)
for key, val in test_set_wers:
s += "{}\t{}{}\n".format(key, val, note)
note = ""
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
AiShell2AsrDataModule.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",
"modified_beam_search",
"fast_beam_search",
"fast_beam_search_LG",
"fast_beam_search_nbest_oracle",
)
params.res_dir = params.exp_dir / params.decoding_method
if params.iter > 0:
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
else:
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
if params.causal:
assert (
"," not in params.chunk_size
), "chunk_size should be one value in decoding."
assert (
"," not in params.left_context_frames
), "left_context_frames should be one value in decoding."
params.suffix += f"-chunk-{params.chunk_size}"
params.suffix += f"-left-context-{params.left_context_frames}"
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"_ilme_scale_{params.ilme_scale}"
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}"
params.suffix += f"-blank-penalty-{params.blank_penalty}"
if params.use_averaged_model:
params.suffix += "-use-averaged-model"
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
logging.info("Decoding started")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"Device: {device}")
lexicon = Lexicon(params.lang_dir)
params.blank_id = lexicon.token_table["<blk>"]
params.vocab_size = max(lexicon.tokens) + 1
graph_compiler = CharCtcTrainingGraphCompiler(
lexicon=lexicon,
device=device,
)
logging.info(params)
logging.info("About to create model")
model = get_model(params)
if not params.use_averaged_model:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
elif params.avg == 1:
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if i >= 1:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
else:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg + 1
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg + 1:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
filename_start = filenames[-1]
filename_end = filenames[0]
logging.info(
"Calculating the averaged model over iteration checkpoints"
f" from {filename_start} (excluded) to {filename_end}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
else:
assert params.avg > 0, params.avg
start = params.epoch - params.avg
assert start >= 1, start
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
logging.info(
f"Calculating the averaged model over epoch range from "
f"{start} (excluded) to {params.epoch}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
model.to(device)
model.eval()
if "fast_beam_search" in params.decoding_method:
if "LG" in params.decoding_method:
lexicon = Lexicon(params.lang_dir)
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
aishell2 = AiShell2AsrDataModule(args)
def remove_short_utt(c: Cut):
T = ((c.num_frames - 7) // 2 + 1) // 2
if T <= 0:
logging.warning(
f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
)
return T > 0
dev_cuts = aishell2.valid_cuts()
dev_cuts = dev_cuts.filter(remove_short_utt)
dev_dl = aishell2.valid_dataloaders(dev_cuts)
test_cuts = aishell2.test_cuts()
test_cuts = test_cuts.filter(remove_short_utt)
test_dl = aishell2.test_dataloaders(test_cuts)
test_sets = ["dev", "test"]
test_dls = [dev_dl, test_dl]
for test_set, test_dl in zip(test_sets, test_dls):
results_dict = decode_dataset(
dl=test_dl,
params=params,
model=model,
lexicon=lexicon,
graph_compiler=graph_compiler,
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/aishell2/ASR/zipformer/decode_stream.py
View File

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

1
egs/aishell2/ASR/zipformer/decoder.py
View File

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

1
egs/aishell2/ASR/zipformer/encoder_interface.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless5/encoder_interface.py

1
egs/aishell2/ASR/zipformer/export-onnx-streaming.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/zipformer/export-onnx-streaming.py

1
egs/aishell2/ASR/zipformer/export-onnx.py
View File

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

1
egs/aishell2/ASR/zipformer/export.py
View File

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

1
egs/aishell2/ASR/zipformer/jit_pretrained.py
View File

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

1
egs/aishell2/ASR/zipformer/jit_pretrained_streaming.py
View File

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

1
egs/aishell2/ASR/zipformer/joiner.py
View File

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

1
egs/aishell2/ASR/zipformer/model.py
View File

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

1
egs/aishell2/ASR/zipformer/onnx_check.py
View File

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

286
egs/aishell2/ASR/zipformer/onnx_decode.py
View File

@ -1,286 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao,
# Xiaoyu Yang,
# 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.
"""
This script loads ONNX exported models and uses them to decode the test sets.
"""
import argparse
import logging
import time
from pathlib import Path
from typing import List, Tuple
import k2
import torch
import torch.nn as nn
from asr_datamodule import AiShell2AsrDataModule
from lhotse.cut import Cut
from onnx_pretrained import OnnxModel, greedy_search
from icefall.utils import setup_logger, store_transcripts, write_error_stats
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(
"--exp-dir",
type=str,
default="pruned_transducer_stateless7/exp",
help="The experiment dir",
)
parser.add_argument(
"--tokens",
type=str,
default="data/lang_char/tokens.txt",
help="Path to the tokens.txt",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="Valid values are greedy_search and modified_beam_search",
)
return parser
def decode_one_batch(
model: OnnxModel, token_table: k2.SymbolTable, batch: dict
) -> List[List[str]]:
"""Decode one batch and return the result.
Currently it only greedy_search is supported.
Args:
model:
The neural model.
token_table:
Mapping ids to tokens.
batch:
It is the return value from iterating
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
for the format of the `batch`.
Returns:
Return the decoded results for each utterance.
"""
feature = batch["inputs"]
assert feature.ndim == 3
# at entry, feature is (N, T, C)
supervisions = batch["supervisions"]
feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
hyps = greedy_search(
model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
)
hyps = [[token_table[h] for h in hyp] for hyp in hyps]
return hyps
def decode_dataset(
dl: torch.utils.data.DataLoader,
model: nn.Module,
token_table: k2.SymbolTable,
) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
"""Decode dataset.
Args:
dl:
PyTorch's dataloader containing the dataset to decode.
model:
The neural model.
token_table:
Mapping ids to tokens.
Returns:
- A list of tuples. Each tuple contains three elements:
- cut_id,
- reference transcript,
- predicted result.
- The total duration (in seconds) of the dataset.
"""
num_cuts = 0
try:
num_batches = len(dl)
except TypeError:
num_batches = "?"
log_interval = 10
total_duration = 0
results = []
for batch_idx, batch in enumerate(dl):
texts = batch["supervisions"]["text"]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
hyps = decode_one_batch(model=model, token_table=token_table, batch=batch)
this_batch = []
assert len(hyps) == len(texts)
for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
ref_words = list(ref_text)
this_batch.append((cut_id, ref_words, hyp_words))
results.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, total_duration
def save_results(
res_dir: Path,
test_set_name: str,
results: List[Tuple[str, List[str], List[str]]],
):
recog_path = res_dir / f"recogs-{test_set_name}.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 = res_dir / f"errs-{test_set_name}.txt"
with open(errs_filename, "w") as f:
wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
logging.info("Wrote detailed error stats to {}".format(errs_filename))
errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
with open(errs_info, "w") as f:
print("WER", file=f)
print(wer, file=f)
s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
AiShell2AsrDataModule.add_arguments(parser)
args = parser.parse_args()
assert (
args.decoding_method == "greedy_search"
), "Only supports greedy_search currently."
res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
setup_logger(f"{res_dir}/log-decode")
logging.info("Decoding started")
device = torch.device("cpu")
logging.info(f"Device: {device}")
token_table = k2.SymbolTable.from_file(args.tokens)
assert token_table[0] == "<blk>"
logging.info(vars(args))
logging.info("About to create model")
model = OnnxModel(
encoder_model_filename=args.encoder_model_filename,
decoder_model_filename=args.decoder_model_filename,
joiner_model_filename=args.joiner_model_filename,
)
# we need cut ids to display recognition results.
args.return_cuts = True
aishell2 = AiShell2AsrDataModule(args)
def remove_short_utt(c: Cut):
T = ((c.num_frames - 7) // 2 + 1) // 2
if T <= 0:
logging.warning(
f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
)
return T > 0
dev_cuts = aishell2.valid_cuts()
dev_cuts = dev_cuts.filter(remove_short_utt)
dev_dl = aishell2.valid_dataloaders(dev_cuts)
test_cuts = aishell2.test_net_cuts()
test_cuts = test_cuts.filter(remove_short_utt)
test_dl = aishell2.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):
start_time = time.time()
results, total_duration = decode_dataset(
dl=test_dl, model=model, token_table=token_table
)
end_time = time.time()
elapsed_seconds = end_time - start_time
rtf = elapsed_seconds / total_duration
logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
logging.info(f"Wave duration: {total_duration:.3f} s")
logging.info(
f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
)
save_results(res_dir=res_dir, test_set_name=test_set, results=results)
logging.info("Done!")
if __name__ == "__main__":
main()

1
egs/aishell2/ASR/zipformer/onnx_pretrained-streaming.py
View File

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

1
egs/aishell2/ASR/zipformer/onnx_pretrained.py
View File

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

1
egs/aishell2/ASR/zipformer/optim.py
View File

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

1
egs/aishell2/ASR/zipformer/pretrained.py
View File

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

1
egs/aishell2/ASR/zipformer/scaling.py
View File

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

1
egs/aishell2/ASR/zipformer/scaling_converter.py
View File

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

1
egs/aishell2/ASR/zipformer/streaming_beam_search.py
View File

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

880
egs/aishell2/ASR/zipformer/streaming_decode.py
View File

@ -1,880 +0,0 @@
#!/usr/bin/env python3
# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang,
# 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:
./zipformer/streaming_decode.py \
--epoch 28 \
--avg 15 \
--causal 1 \
--chunk-size 16 \
--left-context-frames 256 \
--exp-dir ./zipformer/exp \
--decoding-method greedy_search \
--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
from asr_datamodule import AiShell2AsrDataModule
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 torch import Tensor, nn
from torch.nn.utils.rnn import pad_sequence
from train import add_model_arguments, get_model, get_params
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,
make_pad_mask,
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=28,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="zipformer/exp",
help="The experiment dir",
)
parser.add_argument(
"--lang-dir",
type=str,
default="data/lang_char",
help="Path to the lang dir(containing lexicon, tokens, etc.)",
)
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(
"--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,
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(
"--blank-penalty",
type=float,
default=0.0,
help="""
The penalty applied on blank symbol during decoding.
Note: It is a positive value that would be applied to logits like
this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
[batch_size, vocab] and blank id is 0).
""",
)
parser.add_argument(
"--num-decode-streams",
type=int,
default=2000,
help="The number of streams that can be decoded parallel.",
)
add_model_arguments(parser)
return parser
def get_init_states(
model: nn.Module,
batch_size: int = 1,
device: torch.device = torch.device("cpu"),
) -> List[torch.Tensor]:
"""
Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
states[-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
states[-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
"""
states = model.encoder.get_init_states(batch_size, device)
embed_states = model.encoder_embed.get_init_states(batch_size, device)
states.append(embed_states)
processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
states.append(processed_lens)
return states
def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
"""Stack list of zipformer states that correspond to separate utterances
into a single emformer state, so that it can be used as an input for
zipformer when those utterances are formed into a batch.
Args:
state_list:
Each element in state_list corresponding to the internal state
of the zipformer model for a single utterance. For element-n,
state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
cached_val2, cached_conv1, cached_conv2).
state_list[n][-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
state_list[n][-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
Note:
It is the inverse of :func:`unstack_states`.
"""
batch_size = len(state_list)
assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
tot_num_layers = (len(state_list[0]) - 2) // 6
batch_states = []
for layer in range(tot_num_layers):
layer_offset = layer * 6
# cached_key: (left_context_len, batch_size, key_dim)
cached_key = torch.cat(
[state_list[i][layer_offset] for i in range(batch_size)], dim=1
)
# cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
cached_nonlin_attn = torch.cat(
[state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
)
# cached_val1: (left_context_len, batch_size, value_dim)
cached_val1 = torch.cat(
[state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
)
# cached_val2: (left_context_len, batch_size, value_dim)
cached_val2 = torch.cat(
[state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
)
# cached_conv1: (#batch, channels, left_pad)
cached_conv1 = torch.cat(
[state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
)
# cached_conv2: (#batch, channels, left_pad)
cached_conv2 = torch.cat(
[state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
)
batch_states += [
cached_key,
cached_nonlin_attn,
cached_val1,
cached_val2,
cached_conv1,
cached_conv2,
]
cached_embed_left_pad = torch.cat(
[state_list[i][-2] for i in range(batch_size)], dim=0
)
batch_states.append(cached_embed_left_pad)
processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
batch_states.append(processed_lens)
return batch_states
def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
"""Unstack the zipformer state corresponding to a batch of utterances
into a list of states, where the i-th entry is the state from the i-th
utterance in the batch.
Note:
It is the inverse of :func:`stack_states`.
Args:
batch_states: A list of cached tensors of all encoder layers. For layer-i,
states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
cached_conv1, cached_conv2).
state_list[-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
states[-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
Returns:
state_list: A list of list. Each element in state_list corresponding to the internal state
of the zipformer model for a single utterance.
"""
assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
tot_num_layers = (len(batch_states) - 2) // 6
processed_lens = batch_states[-1]
batch_size = processed_lens.shape[0]
state_list = [[] for _ in range(batch_size)]
for layer in range(tot_num_layers):
layer_offset = layer * 6
# cached_key: (left_context_len, batch_size, key_dim)
cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
# cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
chunks=batch_size, dim=1
)
# cached_val1: (left_context_len, batch_size, value_dim)
cached_val1_list = batch_states[layer_offset + 2].chunk(
chunks=batch_size, dim=1
)
# cached_val2: (left_context_len, batch_size, value_dim)
cached_val2_list = batch_states[layer_offset + 3].chunk(
chunks=batch_size, dim=1
)
# cached_conv1: (#batch, channels, left_pad)
cached_conv1_list = batch_states[layer_offset + 4].chunk(
chunks=batch_size, dim=0
)
# cached_conv2: (#batch, channels, left_pad)
cached_conv2_list = batch_states[layer_offset + 5].chunk(
chunks=batch_size, dim=0
)
for i in range(batch_size):
state_list[i] += [
cached_key_list[i],
cached_nonlin_attn_list[i],
cached_val1_list[i],
cached_val2_list[i],
cached_conv1_list[i],
cached_conv2_list[i],
]
cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
for i in range(batch_size):
state_list[i].append(cached_embed_left_pad_list[i])
processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
for i in range(batch_size):
state_list[i].append(processed_lens_list[i])
return state_list
def streaming_forward(
features: Tensor,
feature_lens: Tensor,
model: nn.Module,
states: List[Tensor],
chunk_size: int,
left_context_len: int,
) -> Tuple[Tensor, Tensor, List[Tensor]]:
"""
Returns encoder outputs, output lengths, and updated states.
"""
cached_embed_left_pad = states[-2]
(x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
x=features,
x_lens=feature_lens,
cached_left_pad=cached_embed_left_pad,
)
assert x.size(1) == chunk_size, (x.size(1), chunk_size)
src_key_padding_mask = make_pad_mask(x_lens)
# processed_mask is used to mask out initial states
processed_mask = torch.arange(left_context_len, device=x.device).expand(
x.size(0), left_context_len
)
processed_lens = states[-1] # (batch,)
# (batch, left_context_size)
processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
# Update processed lengths
new_processed_lens = processed_lens + x_lens
# (batch, left_context_size + chunk_size)
src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C)
encoder_states = states[:-2]
(
encoder_out,
encoder_out_lens,
new_encoder_states,
) = model.encoder.streaming_forward(
x=x,
x_lens=x_lens,
states=encoder_states,
src_key_padding_mask=src_key_padding_mask,
)
encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C)
new_states = new_encoder_states + [
new_cached_embed_left_pad,
new_processed_lens,
]
return encoder_out, encoder_out_lens, new_states
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
chunk_size = int(params.chunk_size)
left_context_len = int(params.left_context_frames)
features = []
feature_lens = []
states = []
processed_lens = [] # Used in fast-beam-search
for stream in decode_streams:
feat, feat_len = stream.get_feature_frames(chunk_size * 2)
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)
# Make sure the length after encoder_embed is at least 1.
# The encoder_embed subsample features (T - 7) // 2
# The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
tail_length = chunk_size * 2 + 7 + 2 * 3
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)
encoder_out, encoder_out_lens, new_states = streaming_forward(
features=features,
feature_lens=feature_lens,
model=model,
states=states,
chunk_size=chunk_size,
left_context_len=left_context_len,
)
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,
blank_penalty=params.blank_penalty,
)
elif params.decoding_method == "fast_beam_search":
processed_lens = torch.tensor(processed_lens, device=device)
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,
blank_penalty=params.blank_penalty,
)
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,
blank_penalty=params.blank_penalty,
)
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,
lexicon: Lexicon,
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.
lexicon:
The Lexicon.
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 = 100
decode_results = []
# Contain decode streams currently running.
decode_streams = []
for num, cut in enumerate(cuts):
# each utterance has a DecodeStream.
initial_states = get_init_states(model=model, batch_size=1, 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
if audio.max() > 1:
logging.warning(
f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}."
f"Clipping to [-1, 1]."
)
audio = np.clip(audio, -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=30)
decode_stream.ground_truth = cut.supervisions[0].text
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,
list(decode_streams[i].ground_truth.strip()),
[
lexicon.token_table[idx]
for idx in 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,
decode_streams[i].ground_truth.split(),
[
lexicon.token_table[idx]
for idx in decode_streams[i].decoding_result()
],
)
)
del decode_streams[i]
key = f"blank_penalty_{params.blank_penalty}"
if params.decoding_method == "greedy_search":
key = f"greedy_search_{key}"
elif params.decoding_method == "fast_beam_search":
key = (
f"beam_{params.beam}_"
f"max_contexts_{params.max_contexts}_"
f"max_states_{params.max_states}_{key}"
)
elif params.decoding_method == "modified_beam_search":
key = f"num_active_paths_{params.num_active_paths}_{key}"
else:
raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
return {key: decode_results}
def save_results(
params: AttributeDict,
test_set_name: str,
results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
):
test_set_wers = dict()
for key, results in results_dict.items():
recog_path = (
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
)
results = sorted(results)
store_transcripts(filename=recog_path, texts=results)
logging.info(f"The transcripts are stored in {recog_path}")
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = (
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_filename, "w") as f:
wer = write_error_stats(
f, f"{test_set_name}-{key}", results, 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}-{key}-{params.suffix}.txt"
)
with open(errs_info, "w") as f:
print("settings\tWER", file=f)
for key, val in test_set_wers:
print("{}\t{}".format(key, val), file=f)
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
note = "\tbest for {}".format(test_set_name)
for key, val in test_set_wers:
s += "{}\t{}{}\n".format(key, val, note)
note = ""
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
AiShell2AsrDataModule.add_arguments(parser)
args = parser.parse_args()
args.exp_dir = Path(args.exp_dir)
params = get_params()
params.update(vars(args))
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}"
assert params.causal, params.causal
assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
assert (
"," not in params.left_context_frames
), "left_context_frames should be one value in decoding."
params.suffix += f"-chunk-{params.chunk_size}"
params.suffix += f"-left-context-{params.left_context_frames}"
params.suffix += f"-blank-penalty-{params.blank_penalty}"
# 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", 0)
logging.info(f"Device: {device}")
lexicon = Lexicon(params.lang_dir)
params.blank_id = lexicon.token_table["<blk>"]
params.vocab_size = max(lexicon.tokens) + 1
logging.info(params)
logging.info("About to create model")
model = get_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_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}")
aishell2 = AiShell2AsrDataModule(args)
dev_cuts = aishell2.valid_cuts()
test_cuts = aishell2.test_cuts()
test_sets = ["dev", "test"]
test_cuts = [dev_cuts, test_cuts]
for test_set, test_cut in zip(test_sets, test_cuts):
results_dict = decode_dataset(
cuts=test_cut,
params=params,
model=model,
lexicon=lexicon,
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/aishell2/ASR/zipformer/subsampling.py
View File

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

1350
egs/aishell2/ASR/zipformer/train.py
View File

File diff suppressed because it is too large Load Diff

1
egs/aishell2/ASR/zipformer/zipformer.py
View File

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

0
egs/aishell4/ASR/zipformer/__init__.py
View File

1
egs/aishell4/ASR/zipformer/asr_datamodule.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless5/asr_datamodule.py

1
egs/aishell4/ASR/zipformer/beam_search.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless5/beam_search.py

819
egs/aishell4/ASR/zipformer/decode.py
View File

@ -1,819 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao
# 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
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method greedy_search
(2) modified beam search
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method modified_beam_search \
--beam-size 4
(3) fast beam search (trivial_graph)
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method fast_beam_search \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(4) fast beam search (LG)
./zipformer/decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./zipformer/exp \
--lang-dir data/lang_char \
--max-duration 600 \
--decoding-method fast_beam_search_LG \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(5) fast beam search (nbest oracle WER)
./zipformer/decode.py \
--epoch 35 \
--avg 15 \
--exp-dir ./zipformer/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
"""
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 Aishell4AsrDataModule
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 lhotse.cut import Cut
from local.text_normalize import text_normalize
from train import add_model_arguments, get_model, get_params
from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
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,
make_pad_mask,
setup_logger,
store_transcripts,
str2bool,
write_error_stats,
)
LOG_EPS = math.log(1e-10)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=30,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="zipformer/exp",
help="The experiment dir",
)
parser.add_argument(
"--lang-dir",
type=Path,
default="data/lang_char",
help="The lang dir containing word table and LG graph",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_search
- modified_beam_search
- fast_beam_search
- fast_beam_search_LG
- fast_beam_search_nbest_oracle
If you use fast_beam_search_LG, you have to specify
`--lang-dir`, which should contain `LG.pt`.
""",
)
parser.add_argument(
"--beam-size",
type=int,
default=4,
help="""An integer indicating how many candidates we will keep for each
frame. Used only when --decoding-method is beam_search or
modified_beam_search.""",
)
parser.add_argument(
"--beam",
type=float,
default=20.0,
help="""A floating point value to calculate the cutoff score during beam
search (i.e., `cutoff = max-score - beam`), which is the same as the
`beam` in Kaldi.
Used only when --decoding-method is fast_beam_search,
fast_beam_search, fast_beam_search_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_LG.
It specifies the scale for n-gram LM scores.
""",
)
parser.add_argument(
"--ilme-scale",
type=float,
default=0.2,
help="""
Used only when --decoding_method is fast_beam_search_LG.
It specifies the scale for the internal language model estimation.
""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=8,
help="""Used only when --decoding-method is
fast_beam_search, fast_beam_search, fast_beam_search_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, fast_beam_search_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_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 and fast_beam_search_nbest_oracle""",
)
parser.add_argument(
"--blank-penalty",
type=float,
default=0.0,
help="""
The penalty applied on blank symbol during decoding.
Note: It is a positive value that would be applied to logits like
this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
[batch_size, vocab] and blank id is 0).
""",
)
add_model_arguments(parser)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
lexicon: Lexicon,
graph_compiler: CharCtcTrainingGraphCompiler,
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.
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 LG, 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.causal:
# this seems to cause insertions at the end of the utterance if used with zipformer.
pad_len = 30
feature_lens += pad_len
feature = torch.nn.functional.pad(
feature,
pad=(0, 0, 0, pad_len),
value=LOG_EPS,
)
x, x_lens = model.encoder_embed(feature, feature_lens)
src_key_padding_mask = make_pad_mask(x_lens)
x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C)
encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask)
encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C)
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,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
elif params.decoding_method == "fast_beam_search_LG":
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,
blank_penalty=params.blank_penalty,
ilme_scale=params.ilme_scale,
)
for hyp in hyp_tokens:
sentence = "".join([lexicon.word_table[i] for i in hyp])
hyps.append(list(sentence))
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=graph_compiler.texts_to_ids(supervisions["text"]),
nbest_scale=params.nbest_scale,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
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,
blank_penalty=params.blank_penalty,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
elif params.decoding_method == "modified_beam_search":
hyp_tokens = modified_beam_search(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
blank_penalty=params.blank_penalty,
beam=params.beam_size,
)
for i in range(encoder_out.size(0)):
hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
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,
blank_penalty=params.blank_penalty,
)
elif params.decoding_method == "beam_search":
hyp = beam_search(
model=model,
encoder_out=encoder_out_i,
beam=params.beam_size,
blank_penalty=params.blank_penalty,
)
else:
raise ValueError(
f"Unsupported decoding method: {params.decoding_method}"
)
hyps.append([lexicon.token_table[idx] for idx in hyp])
key = f"blank_penalty_{params.blank_penalty}"
if params.decoding_method == "greedy_search":
return {"greedy_search_" + key: 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"_ilme_scale_{params.ilme_scale}"
key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
return {key: hyps}
else:
return {f"beam_size_{params.beam_size}_" + key: hyps}
def decode_dataset(
dl: torch.utils.data.DataLoader,
params: AttributeDict,
model: nn.Module,
lexicon: Lexicon,
graph_compiler: CharCtcTrainingGraphCompiler,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[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.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or LG, 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"]
texts = [list("".join(text.split())) for text in texts]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
hyps_dict = decode_one_batch(
params=params,
model=model,
lexicon=lexicon,
graph_compiler=graph_compiler,
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):
this_batch.append((cut_id, ref_text, 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[List[int], List[int]]]],
):
test_set_wers = dict()
for key, results in results_dict.items():
recog_path = (
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
)
results = sorted(results)
store_transcripts(filename=recog_path, texts=results)
logging.info(f"The transcripts are stored in {recog_path}")
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = (
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_filename, "w") as f:
wer = write_error_stats(
f, f"{test_set_name}-{key}", results, 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}-{key}-{params.suffix}.txt"
)
with open(errs_info, "w") as f:
print("settings\tWER", file=f)
for key, val in test_set_wers:
print("{}\t{}".format(key, val), file=f)
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
note = "\tbest for {}".format(test_set_name)
for key, val in test_set_wers:
s += "{}\t{}{}\n".format(key, val, note)
note = ""
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
Aishell4AsrDataModule.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",
"modified_beam_search",
"fast_beam_search",
"fast_beam_search_LG",
"fast_beam_search_nbest_oracle",
)
params.res_dir = params.exp_dir / params.decoding_method
if params.iter > 0:
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
else:
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
if params.causal:
assert (
"," not in params.chunk_size
), "chunk_size should be one value in decoding."
assert (
"," not in params.left_context_frames
), "left_context_frames should be one value in decoding."
params.suffix += f"-chunk-{params.chunk_size}"
params.suffix += f"-left-context-{params.left_context_frames}"
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"_ilme_scale_{params.ilme_scale}"
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}"
params.suffix += f"-blank-penalty-{params.blank_penalty}"
if params.use_averaged_model:
params.suffix += "-use-averaged-model"
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
logging.info("Decoding started")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"Device: {device}")
lexicon = Lexicon(params.lang_dir)
params.blank_id = lexicon.token_table["<blk>"]
params.vocab_size = max(lexicon.tokens) + 1
graph_compiler = CharCtcTrainingGraphCompiler(
lexicon=lexicon,
device=device,
)
logging.info(params)
logging.info("About to create model")
model = get_model(params)
if not params.use_averaged_model:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
elif params.avg == 1:
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if i >= 1:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
else:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg + 1
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg + 1:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
filename_start = filenames[-1]
filename_end = filenames[0]
logging.info(
"Calculating the averaged model over iteration checkpoints"
f" from {filename_start} (excluded) to {filename_end}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
else:
assert params.avg > 0, params.avg
start = params.epoch - params.avg
assert start >= 1, start
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
logging.info(
f"Calculating the averaged model over epoch range from "
f"{start} (excluded) to {params.epoch}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
model.to(device)
model.eval()
if "fast_beam_search" in params.decoding_method:
if "LG" in params.decoding_method:
lexicon = Lexicon(params.lang_dir)
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
aishell4 = Aishell4AsrDataModule(args)
def remove_short_utt(c: Cut):
T = ((c.num_frames - 7) // 2 + 1) // 2
if T <= 0:
logging.warning(
f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
)
return T > 0
def text_normalize_for_cut(c: Cut):
# Text normalize for each sample
text = c.supervisions[0].text
text = text.strip("\n").strip("\t")
c.supervisions[0].text = text_normalize(text)
return c
test_cuts = aishell4.test_cuts()
test_cuts = test_cuts.map(text_normalize_for_cut)
test_cuts = test_cuts.filter(remove_short_utt)
test_dl = aishell4.test_dataloaders(test_cuts)
test_sets = ["test"]
test_dls = [test_dl]
for test_set, test_dl in zip(test_sets, test_dls):
results_dict = decode_dataset(
dl=test_dl,
params=params,
model=model,
lexicon=lexicon,
graph_compiler=graph_compiler,
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/aishell4/ASR/zipformer/decode_stream.py
View File

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

1
egs/aishell4/ASR/zipformer/decoder.py
View File

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

1
egs/aishell4/ASR/zipformer/encoder_interface.py
View File

@ -1 +0,0 @@
../pruned_transducer_stateless5/encoder_interface.py

1
egs/aishell4/ASR/zipformer/export-onnx-streaming.py
View File

@ -1 +0,0 @@
../../../librispeech/ASR/zipformer/export-onnx-streaming.py

1
egs/aishell4/ASR/zipformer/export-onnx.py
View File

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

1
egs/aishell4/ASR/zipformer/export.py
View File

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

1
egs/aishell4/ASR/zipformer/jit_pretrained.py
View File

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

1
egs/aishell4/ASR/zipformer/jit_pretrained_streaming.py
View File

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

1
egs/aishell4/ASR/zipformer/joiner.py
View File

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

1
egs/aishell4/ASR/zipformer/local
View File

@ -1 +0,0 @@
../local

1
egs/aishell4/ASR/zipformer/model.py
View File

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

1
egs/aishell4/ASR/zipformer/onnx_check.py
View File

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

291
egs/aishell4/ASR/zipformer/onnx_decode.py
View File

@ -1,291 +0,0 @@
#!/usr/bin/env python3
#
# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
# Zengwei Yao,
# Xiaoyu Yang,
# 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.
"""
This script loads ONNX exported models and uses them to decode the test sets.
"""
import argparse
import logging
import time
from pathlib import Path
from typing import List, Tuple
import k2
import torch
import torch.nn as nn
from asr_datamodule import Aishell4AsrDataModule
from lhotse.cut import Cut
from local.text_normalize import text_normalize
from onnx_pretrained import OnnxModel, greedy_search
from icefall.utils import setup_logger, store_transcripts, write_error_stats
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(
"--exp-dir",
type=str,
default="pruned_transducer_stateless7/exp",
help="The experiment dir",
)
parser.add_argument(
"--tokens",
type=str,
default="data/lang_char/tokens.txt",
help="Path to the tokens.txt",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="Valid values are greedy_search and modified_beam_search",
)
return parser
def decode_one_batch(
model: OnnxModel, token_table: k2.SymbolTable, batch: dict
) -> List[List[str]]:
"""Decode one batch and return the result.
Currently it only greedy_search is supported.
Args:
model:
The neural model.
token_table:
Mapping ids to tokens.
batch:
It is the return value from iterating
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
for the format of the `batch`.
Returns:
Return the decoded results for each utterance.
"""
feature = batch["inputs"]
assert feature.ndim == 3
# at entry, feature is (N, T, C)
supervisions = batch["supervisions"]
feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
hyps = greedy_search(
model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
)
hyps = [[token_table[h] for h in hyp] for hyp in hyps]
return hyps
def decode_dataset(
dl: torch.utils.data.DataLoader,
model: nn.Module,
token_table: k2.SymbolTable,
) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
"""Decode dataset.
Args:
dl:
PyTorch's dataloader containing the dataset to decode.
model:
The neural model.
token_table:
Mapping ids to tokens.
Returns:
- A list of tuples. Each tuple contains three elements:
- cut_id,
- reference transcript,
- predicted result.
- The total duration (in seconds) of the dataset.
"""
num_cuts = 0
try:
num_batches = len(dl)
except TypeError:
num_batches = "?"
log_interval = 10
total_duration = 0
results = []
for batch_idx, batch in enumerate(dl):
texts = batch["supervisions"]["text"]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
hyps = decode_one_batch(model=model, token_table=token_table, batch=batch)
this_batch = []
assert len(hyps) == len(texts)
for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
ref_words = list(ref_text)
this_batch.append((cut_id, ref_words, hyp_words))
results.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, total_duration
def save_results(
res_dir: Path,
test_set_name: str,
results: List[Tuple[str, List[str], List[str]]],
):
recog_path = res_dir / f"recogs-{test_set_name}.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 = res_dir / f"errs-{test_set_name}.txt"
with open(errs_filename, "w") as f:
wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
logging.info("Wrote detailed error stats to {}".format(errs_filename))
errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
with open(errs_info, "w") as f:
print("WER", file=f)
print(wer, file=f)
s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
Aishell4AsrDataModule.add_arguments(parser)
args = parser.parse_args()
assert (
args.decoding_method == "greedy_search"
), "Only supports greedy_search currently."
res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
setup_logger(f"{res_dir}/log-decode")
logging.info("Decoding started")
device = torch.device("cpu")
logging.info(f"Device: {device}")
token_table = k2.SymbolTable.from_file(args.tokens)
assert token_table[0] == "<blk>"
logging.info(vars(args))
logging.info("About to create model")
model = OnnxModel(
encoder_model_filename=args.encoder_model_filename,
decoder_model_filename=args.decoder_model_filename,
joiner_model_filename=args.joiner_model_filename,
)
# we need cut ids to display recognition results.
args.return_cuts = True
aishell4 = Aishell4AsrDataModule(args)
def remove_short_utt(c: Cut):
T = ((c.num_frames - 7) // 2 + 1) // 2
if T <= 0:
logging.warning(
f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
)
return T > 0
def text_normalize_for_cut(c: Cut):
# Text normalize for each sample
text = c.supervisions[0].text
text = text.strip("\n").strip("\t")
c.supervisions[0].text = text_normalize(text)
return c
test_cuts = aishell4.test_cuts()
test_cuts = test_cuts.map(text_normalize_for_cut)
test_cuts = test_cuts.filter(remove_short_utt)
test_dl = aishell4.test_dataloaders(test_cuts)
test_sets = ["test"]
test_dls = [test_dl]
for test_set, test_dl in zip(test_sets, test_dls):
start_time = time.time()
results, total_duration = decode_dataset(
dl=test_dl, model=model, token_table=token_table
)
end_time = time.time()
elapsed_seconds = end_time - start_time
rtf = elapsed_seconds / total_duration
logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
logging.info(f"Wave duration: {total_duration:.3f} s")
logging.info(
f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
)
save_results(res_dir=res_dir, test_set_name=test_set, results=results)
logging.info("Done!")
if __name__ == "__main__":
main()

1
egs/aishell4/ASR/zipformer/onnx_pretrained-streaming.py
View File

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

1
egs/aishell4/ASR/zipformer/onnx_pretrained.py
View File

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

1
egs/aishell4/ASR/zipformer/optim.py
View File

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

1
egs/aishell4/ASR/zipformer/pretrained.py
View File

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

1
egs/aishell4/ASR/zipformer/scaling.py
View File

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

1
egs/aishell4/ASR/zipformer/scaling_converter.py
View File

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

1
egs/aishell4/ASR/zipformer/streaming_beam_search.py
View File

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

888
egs/aishell4/ASR/zipformer/streaming_decode.py
View File

@ -1,888 +0,0 @@
#!/usr/bin/env python3
# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang,
# 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:
./zipformer/streaming_decode.py \
--epoch 28 \
--avg 15 \
--causal 1 \
--chunk-size 16 \
--left-context-frames 256 \
--exp-dir ./zipformer/exp \
--decoding-method greedy_search \
--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
from asr_datamodule import Aishell4AsrDataModule
from decode_stream import DecodeStream
from kaldifeat import Fbank, FbankOptions
from lhotse import CutSet
from lhotse.cut import Cut
from local.text_normalize import text_normalize
from streaming_beam_search import (
fast_beam_search_one_best,
greedy_search,
modified_beam_search,
)
from torch import Tensor, nn
from torch.nn.utils.rnn import pad_sequence
from train import add_model_arguments, get_model, get_params
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,
make_pad_mask,
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=28,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="zipformer/exp",
help="The experiment dir",
)
parser.add_argument(
"--lang-dir",
type=str,
default="data/lang_char",
help="Path to the lang dir(containing lexicon, tokens, etc.)",
)
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(
"--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,
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(
"--blank-penalty",
type=float,
default=0.0,
help="""
The penalty applied on blank symbol during decoding.
Note: It is a positive value that would be applied to logits like
this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
[batch_size, vocab] and blank id is 0).
""",
)
parser.add_argument(
"--num-decode-streams",
type=int,
default=2000,
help="The number of streams that can be decoded parallel.",
)
add_model_arguments(parser)
return parser
def get_init_states(
model: nn.Module,
batch_size: int = 1,
device: torch.device = torch.device("cpu"),
) -> List[torch.Tensor]:
"""
Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
states[-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
states[-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
"""
states = model.encoder.get_init_states(batch_size, device)
embed_states = model.encoder_embed.get_init_states(batch_size, device)
states.append(embed_states)
processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
states.append(processed_lens)
return states
def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
"""Stack list of zipformer states that correspond to separate utterances
into a single emformer state, so that it can be used as an input for
zipformer when those utterances are formed into a batch.
Args:
state_list:
Each element in state_list corresponding to the internal state
of the zipformer model for a single utterance. For element-n,
state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
cached_val2, cached_conv1, cached_conv2).
state_list[n][-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
state_list[n][-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
Note:
It is the inverse of :func:`unstack_states`.
"""
batch_size = len(state_list)
assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
tot_num_layers = (len(state_list[0]) - 2) // 6
batch_states = []
for layer in range(tot_num_layers):
layer_offset = layer * 6
# cached_key: (left_context_len, batch_size, key_dim)
cached_key = torch.cat(
[state_list[i][layer_offset] for i in range(batch_size)], dim=1
)
# cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
cached_nonlin_attn = torch.cat(
[state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
)
# cached_val1: (left_context_len, batch_size, value_dim)
cached_val1 = torch.cat(
[state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
)
# cached_val2: (left_context_len, batch_size, value_dim)
cached_val2 = torch.cat(
[state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
)
# cached_conv1: (#batch, channels, left_pad)
cached_conv1 = torch.cat(
[state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
)
# cached_conv2: (#batch, channels, left_pad)
cached_conv2 = torch.cat(
[state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
)
batch_states += [
cached_key,
cached_nonlin_attn,
cached_val1,
cached_val2,
cached_conv1,
cached_conv2,
]
cached_embed_left_pad = torch.cat(
[state_list[i][-2] for i in range(batch_size)], dim=0
)
batch_states.append(cached_embed_left_pad)
processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
batch_states.append(processed_lens)
return batch_states
def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
"""Unstack the zipformer state corresponding to a batch of utterances
into a list of states, where the i-th entry is the state from the i-th
utterance in the batch.
Note:
It is the inverse of :func:`stack_states`.
Args:
batch_states: A list of cached tensors of all encoder layers. For layer-i,
states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
cached_conv1, cached_conv2).
state_list[-2] is the cached left padding for ConvNeXt module,
of shape (batch_size, num_channels, left_pad, num_freqs)
states[-1] is processed_lens of shape (batch,), which records the number
of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
Returns:
state_list: A list of list. Each element in state_list corresponding to the internal state
of the zipformer model for a single utterance.
"""
assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
tot_num_layers = (len(batch_states) - 2) // 6
processed_lens = batch_states[-1]
batch_size = processed_lens.shape[0]
state_list = [[] for _ in range(batch_size)]
for layer in range(tot_num_layers):
layer_offset = layer * 6
# cached_key: (left_context_len, batch_size, key_dim)
cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
# cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
chunks=batch_size, dim=1
)
# cached_val1: (left_context_len, batch_size, value_dim)
cached_val1_list = batch_states[layer_offset + 2].chunk(
chunks=batch_size, dim=1
)
# cached_val2: (left_context_len, batch_size, value_dim)
cached_val2_list = batch_states[layer_offset + 3].chunk(
chunks=batch_size, dim=1
)
# cached_conv1: (#batch, channels, left_pad)
cached_conv1_list = batch_states[layer_offset + 4].chunk(
chunks=batch_size, dim=0
)
# cached_conv2: (#batch, channels, left_pad)
cached_conv2_list = batch_states[layer_offset + 5].chunk(
chunks=batch_size, dim=0
)
for i in range(batch_size):
state_list[i] += [
cached_key_list[i],
cached_nonlin_attn_list[i],
cached_val1_list[i],
cached_val2_list[i],
cached_conv1_list[i],
cached_conv2_list[i],
]
cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
for i in range(batch_size):
state_list[i].append(cached_embed_left_pad_list[i])
processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
for i in range(batch_size):
state_list[i].append(processed_lens_list[i])
return state_list
def streaming_forward(
features: Tensor,
feature_lens: Tensor,
model: nn.Module,
states: List[Tensor],
chunk_size: int,
left_context_len: int,
) -> Tuple[Tensor, Tensor, List[Tensor]]:
"""
Returns encoder outputs, output lengths, and updated states.
"""
cached_embed_left_pad = states[-2]
(x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
x=features,
x_lens=feature_lens,
cached_left_pad=cached_embed_left_pad,
)
assert x.size(1) == chunk_size, (x.size(1), chunk_size)
src_key_padding_mask = make_pad_mask(x_lens)
# processed_mask is used to mask out initial states
processed_mask = torch.arange(left_context_len, device=x.device).expand(
x.size(0), left_context_len
)
processed_lens = states[-1] # (batch,)
# (batch, left_context_size)
processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
# Update processed lengths
new_processed_lens = processed_lens + x_lens
# (batch, left_context_size + chunk_size)
src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C)
encoder_states = states[:-2]
(
encoder_out,
encoder_out_lens,
new_encoder_states,
) = model.encoder.streaming_forward(
x=x,
x_lens=x_lens,
states=encoder_states,
src_key_padding_mask=src_key_padding_mask,
)
encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C)
new_states = new_encoder_states + [
new_cached_embed_left_pad,
new_processed_lens,
]
return encoder_out, encoder_out_lens, new_states
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
chunk_size = int(params.chunk_size)
left_context_len = int(params.left_context_frames)
features = []
feature_lens = []
states = []
processed_lens = [] # Used in fast-beam-search
for stream in decode_streams:
feat, feat_len = stream.get_feature_frames(chunk_size * 2)
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)
# Make sure the length after encoder_embed is at least 1.
# The encoder_embed subsample features (T - 7) // 2
# The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
tail_length = chunk_size * 2 + 7 + 2 * 3
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)
encoder_out, encoder_out_lens, new_states = streaming_forward(
features=features,
feature_lens=feature_lens,
model=model,
states=states,
chunk_size=chunk_size,
left_context_len=left_context_len,
)
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,
blank_penalty=params.blank_penalty,
)
elif params.decoding_method == "fast_beam_search":
processed_lens = torch.tensor(processed_lens, device=device)
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,
blank_penalty=params.blank_penalty,
)
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,
blank_penalty=params.blank_penalty,
)
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,
lexicon: Lexicon,
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.
lexicon:
The Lexicon.
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 = 100
decode_results = []
# Contain decode streams currently running.
decode_streams = []
for num, cut in enumerate(cuts):
# each utterance has a DecodeStream.
initial_states = get_init_states(model=model, batch_size=1, 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
if audio.max() > 1:
logging.warning(
f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}."
f"Clipping to [-1, 1]."
)
audio = np.clip(audio, -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=30)
decode_stream.ground_truth = cut.supervisions[0].text
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,
list(decode_streams[i].ground_truth.strip()),
[
lexicon.token_table[idx]
for idx in 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,
decode_streams[i].ground_truth.split(),
[
lexicon.token_table[idx]
for idx in decode_streams[i].decoding_result()
],
)
)
del decode_streams[i]
key = f"blank_penalty_{params.blank_penalty}"
if params.decoding_method == "greedy_search":
key = f"greedy_search_{key}"
elif params.decoding_method == "fast_beam_search":
key = (
f"beam_{params.beam}_"
f"max_contexts_{params.max_contexts}_"
f"max_states_{params.max_states}_{key}"
)
elif params.decoding_method == "modified_beam_search":
key = f"num_active_paths_{params.num_active_paths}_{key}"
else:
raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
return {key: decode_results}
def save_results(
params: AttributeDict,
test_set_name: str,
results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
):
test_set_wers = dict()
for key, results in results_dict.items():
recog_path = (
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
)
results = sorted(results)
store_transcripts(filename=recog_path, texts=results)
logging.info(f"The transcripts are stored in {recog_path}")
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = (
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_filename, "w") as f:
wer = write_error_stats(
f, f"{test_set_name}-{key}", results, 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}-{key}-{params.suffix}.txt"
)
with open(errs_info, "w") as f:
print("settings\tWER", file=f)
for key, val in test_set_wers:
print("{}\t{}".format(key, val), file=f)
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
note = "\tbest for {}".format(test_set_name)
for key, val in test_set_wers:
s += "{}\t{}{}\n".format(key, val, note)
note = ""
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
Aishell4AsrDataModule.add_arguments(parser)
args = parser.parse_args()
args.exp_dir = Path(args.exp_dir)
params = get_params()
params.update(vars(args))
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}"
assert params.causal, params.causal
assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
assert (
"," not in params.left_context_frames
), "left_context_frames should be one value in decoding."
params.suffix += f"-chunk-{params.chunk_size}"
params.suffix += f"-left-context-{params.left_context_frames}"
params.suffix += f"-blank-penalty-{params.blank_penalty}"
# 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", 0)
logging.info(f"Device: {device}")
lexicon = Lexicon(params.lang_dir)
params.blank_id = lexicon.token_table["<blk>"]
params.vocab_size = max(lexicon.tokens) + 1
logging.info(params)
logging.info("About to create model")
model = get_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_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}")
aishell4 = Aishell4AsrDataModule(args)
def text_normalize_for_cut(c: Cut):
# Text normalize for each sample
text = c.supervisions[0].text
text = text.strip("\n").strip("\t")
c.supervisions[0].text = text_normalize(text)
return c
test_cuts = aishell4.test_cuts()
test_cuts = test_cuts.map(text_normalize_for_cut)
test_sets = ["test"]
test_cuts = [test_cuts]
for test_set, test_cut in zip(test_sets, test_cuts):
results_dict = decode_dataset(
cuts=test_cut,
params=params,
model=model,
lexicon=lexicon,
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/aishell4/ASR/zipformer/subsampling.py
View File

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

1362
egs/aishell4/ASR/zipformer/train.py
View File

File diff suppressed because it is too large Load Diff

1
egs/aishell4/ASR/zipformer/zipformer.py
View File

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