add streaming decode
This commit is contained in:
luomingshuang
parent
227f32f089
commit
6d77f4c239
File diff suppressed because it is too large
Load Diff
@ -53,6 +53,7 @@ When training with the L subset, usage:
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import math
|
||||
from collections import defaultdict
|
||||
from pathlib import Path
|
||||
from typing import Dict, List, Optional, Tuple
|
||||
@ -68,10 +69,11 @@ from beam_search import (
|
||||
greedy_search_batch,
|
||||
modified_beam_search,
|
||||
)
|
||||
from train import get_params, get_transducer_model
|
||||
from train import add_model_arguments, get_params, get_transducer_model
|
||||
|
||||
from icefall.checkpoint import (
|
||||
average_checkpoints,
|
||||
average_checkpoints_with_averaged_model,
|
||||
find_checkpoints,
|
||||
load_checkpoint,
|
||||
)
|
||||
@ -80,9 +82,12 @@ from icefall.utils import (
|
||||
AttributeDict,
|
||||
setup_logger,
|
||||
store_transcripts,
|
||||
str2bool,
|
||||
write_error_stats,
|
||||
)
|
||||
|
||||
LOG_EPS = math.log(1e-10)
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
@ -92,17 +97,20 @@ def get_parser():
|
||||
parser.add_argument(
|
||||
"--epoch",
|
||||
type=int,
|
||||
default=28,
|
||||
help="It specifies the checkpoint to use for decoding."
|
||||
"Note: Epoch counts from 0.",
|
||||
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(
|
||||
"--batch",
|
||||
"--iter",
|
||||
type=int,
|
||||
default=None,
|
||||
help="It specifies the batch checkpoint to use for decoding."
|
||||
"Note: Epoch counts from 0.",
|
||||
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(
|
||||
@ -111,24 +119,24 @@ def get_parser():
|
||||
default=15,
|
||||
help="Number of checkpoints to average. Automatically select "
|
||||
"consecutive checkpoints before the checkpoint specified by "
|
||||
"'--epoch'. ",
|
||||
"'--epoch' and '--iter'",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--avg-last-n",
|
||||
type=int,
|
||||
default=0,
|
||||
help="""If positive, --epoch and --avg are ignored and it
|
||||
will use the last n checkpoints exp_dir/checkpoint-xxx.pt
|
||||
where xxx is the number of processed batches while
|
||||
saving that checkpoint.
|
||||
""",
|
||||
"--use-averaged-model",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="Whether to load averaged model. Currently it only supports "
|
||||
"using --epoch. If True, it would decode with the averaged model "
|
||||
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
|
||||
"Actually only the models with epoch number of `epoch-avg` and "
|
||||
"`epoch` are loaded for averaging. ",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--exp-dir",
|
||||
type=str,
|
||||
default="pruned_transducer_stateless2/exp",
|
||||
default="pruned_transducer_stateless5/exp",
|
||||
help="The experiment dir",
|
||||
)
|
||||
|
||||
@ -204,6 +212,31 @@ def get_parser():
|
||||
Used only when --decoding_method is greedy_search""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--simulate-streaming",
|
||||
type=str2bool,
|
||||
default=False,
|
||||
help="""Whether to simulate streaming in decoding, this is a good way to
|
||||
test a streaming model.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--decode-chunk-size",
|
||||
type=int,
|
||||
default=16,
|
||||
help="The chunk size for decoding (in frames after subsampling)",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--left-context",
|
||||
type=int,
|
||||
default=64,
|
||||
help="left context can be seen during decoding (in frames after subsampling)",
|
||||
)
|
||||
|
||||
add_model_arguments(parser)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
@ -240,7 +273,7 @@ def decode_one_batch(
|
||||
Return the decoding result. See above description for the format of
|
||||
the returned dict.
|
||||
"""
|
||||
device = model.device
|
||||
device = next(model.parameters()).device
|
||||
feature = batch["inputs"]
|
||||
assert feature.ndim == 3
|
||||
|
||||
@ -250,9 +283,26 @@ def decode_one_batch(
|
||||
supervisions = batch["supervisions"]
|
||||
feature_lens = supervisions["num_frames"].to(device)
|
||||
|
||||
encoder_out, encoder_out_lens = model.encoder(
|
||||
x=feature, x_lens=feature_lens
|
||||
feature_lens += params.left_context
|
||||
feature = torch.nn.functional.pad(
|
||||
feature,
|
||||
pad=(0, 0, 0, params.left_context),
|
||||
value=LOG_EPS,
|
||||
)
|
||||
|
||||
if params.simulate_streaming:
|
||||
encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
|
||||
x=feature,
|
||||
x_lens=feature_lens,
|
||||
chunk_size=params.decode_chunk_size,
|
||||
left_context=params.left_context,
|
||||
simulate_streaming=True,
|
||||
)
|
||||
else:
|
||||
encoder_out, encoder_out_lens = model.encoder(
|
||||
x=feature, x_lens=feature_lens
|
||||
)
|
||||
|
||||
hyps = []
|
||||
|
||||
if params.decoding_method == "fast_beam_search":
|
||||
@ -482,32 +532,92 @@ def main():
|
||||
params.blank_id = lexicon.token_table["<blk>"]
|
||||
params.vocab_size = max(lexicon.tokens) + 1
|
||||
|
||||
if params.simulate_streaming:
|
||||
assert (
|
||||
params.causal_convolution
|
||||
), "Decoding in streaming requires causal convolution"
|
||||
|
||||
logging.info(params)
|
||||
|
||||
logging.info("About to create model")
|
||||
model = get_transducer_model(params)
|
||||
|
||||
if params.avg_last_n > 0:
|
||||
filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
elif params.avg == 1:
|
||||
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
|
||||
elif params.batch is not None:
|
||||
filenames = f"{params.exp_dir}/checkpoint-{params.batch}.pt"
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints([filenames], device=device))
|
||||
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:
|
||||
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))
|
||||
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()
|
||||
|
||||
@ -0,0 +1,126 @@
|
||||
# Copyright 2022 Xiaomi Corp. (authors: 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.
|
||||
|
||||
import math
|
||||
from typing import List, Optional, Tuple
|
||||
|
||||
import k2
|
||||
import torch
|
||||
|
||||
from icefall.utils import AttributeDict
|
||||
|
||||
|
||||
class DecodeStream(object):
|
||||
def __init__(
|
||||
self,
|
||||
params: AttributeDict,
|
||||
initial_states: List[torch.Tensor],
|
||||
decoding_graph: Optional[k2.Fsa] = None,
|
||||
device: torch.device = torch.device("cpu"),
|
||||
) -> None:
|
||||
"""
|
||||
Args:
|
||||
initial_states:
|
||||
Initial decode states of the model, e.g. the return value of
|
||||
`get_init_state` in conformer.py
|
||||
decoding_graph:
|
||||
Decoding graph used for decoding, may be a TrivialGraph or a HLG.
|
||||
Used only when decoding_method is fast_beam_search.
|
||||
device:
|
||||
The device to run this stream.
|
||||
"""
|
||||
if decoding_graph is not None:
|
||||
assert device == decoding_graph.device
|
||||
|
||||
self.params = params
|
||||
self.LOG_EPS = math.log(1e-10)
|
||||
|
||||
self.states = initial_states
|
||||
|
||||
# It contains a 2-D tensors representing the feature frames.
|
||||
self.features: torch.Tensor = None
|
||||
|
||||
self.num_frames: int = 0
|
||||
# how many frames have been processed. (before subsampling).
|
||||
# we only modify this value in `func:get_feature_frames`.
|
||||
self.num_processed_frames: int = 0
|
||||
|
||||
self._done: bool = False
|
||||
|
||||
# The transcript of current utterance.
|
||||
self.ground_truth: str = ""
|
||||
|
||||
# The decoding result (partial or final) of current utterance.
|
||||
self.hyp: List = []
|
||||
|
||||
# how many frames have been processed, after subsampling (i.e. a
|
||||
# cumulative sum of the second return value of
|
||||
# encoder.streaming_forward
|
||||
self.done_frames: int = 0
|
||||
|
||||
self.pad_length = (
|
||||
params.right_context + 2
|
||||
) * params.subsampling_factor + 3
|
||||
|
||||
if params.decoding_method == "greedy_search":
|
||||
self.hyp = [params.blank_id] * params.context_size
|
||||
elif params.decoding_method == "fast_beam_search":
|
||||
# The rnnt_decoding_stream for fast_beam_search.
|
||||
self.rnnt_decoding_stream: k2.RnntDecodingStream = (
|
||||
k2.RnntDecodingStream(decoding_graph)
|
||||
)
|
||||
else:
|
||||
assert (
|
||||
False
|
||||
), f"Decoding method :{params.decoding_method} do not support."
|
||||
|
||||
@property
|
||||
def done(self) -> bool:
|
||||
"""Return True if all the features are processed."""
|
||||
return self._done
|
||||
|
||||
def set_features(
|
||||
self,
|
||||
features: torch.Tensor,
|
||||
) -> None:
|
||||
"""Set features tensor of current utterance."""
|
||||
assert features.dim() == 2, features.dim()
|
||||
self.features = torch.nn.functional.pad(
|
||||
features,
|
||||
(0, 0, 0, self.pad_length),
|
||||
mode="constant",
|
||||
value=self.LOG_EPS,
|
||||
)
|
||||
self.num_frames = self.features.size(0)
|
||||
|
||||
def get_feature_frames(self, chunk_size: int) -> Tuple[torch.Tensor, int]:
|
||||
"""Consume chunk_size frames of features"""
|
||||
chunk_length = chunk_size + self.pad_length
|
||||
|
||||
ret_length = min(
|
||||
self.num_frames - self.num_processed_frames, chunk_length
|
||||
)
|
||||
|
||||
ret_features = self.features[
|
||||
self.num_processed_frames : self.num_processed_frames # noqa
|
||||
+ ret_length
|
||||
]
|
||||
|
||||
self.num_processed_frames += chunk_size
|
||||
if self.num_processed_frames >= self.num_frames:
|
||||
self._done = True
|
||||
|
||||
return ret_features, ret_length
|
||||
@ -0,0 +1,739 @@
|
||||
#!/usr/bin/env python3
|
||||
# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
|
||||
#
|
||||
# See ../../../../LICENSE for clarification regarding multiple authors
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
"""
|
||||
Usage:
|
||||
python pruned_transducer_stateless5/streaming_decode.py \
|
||||
--epoch 5 \
|
||||
--avg 1 \
|
||||
--decode-chunk-size 16 \
|
||||
--left-context 64 \
|
||||
--right-context 0 \
|
||||
--exp-dir ./pruned_transducer_stateless5/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
|
||||
import torch.nn as nn
|
||||
from asr_datamodule import WenetSpeechAsrDataModule
|
||||
from decode_stream import DecodeStream
|
||||
from kaldifeat import Fbank, FbankOptions
|
||||
from lhotse import CutSet
|
||||
from torch.nn.utils.rnn import pad_sequence
|
||||
from train import add_model_arguments, get_params, get_transducer_model
|
||||
|
||||
from icefall.checkpoint import (
|
||||
average_checkpoints,
|
||||
average_checkpoints_with_averaged_model,
|
||||
find_checkpoints,
|
||||
load_checkpoint,
|
||||
)
|
||||
from icefall.decode import one_best_decoding
|
||||
from icefall.lexicon import Lexicon
|
||||
from icefall.utils import (
|
||||
AttributeDict,
|
||||
get_texts,
|
||||
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 0.
|
||||
You can specify --avg to use more checkpoints for model averaging.""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--iter",
|
||||
type=int,
|
||||
default=0,
|
||||
help="""If positive, --epoch is ignored and it
|
||||
will use the checkpoint exp_dir/checkpoint-iter.pt.
|
||||
You can specify --avg to use more checkpoints for model averaging.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--avg",
|
||||
type=int,
|
||||
default=15,
|
||||
help="Number of checkpoints to average. Automatically select "
|
||||
"consecutive checkpoints before the checkpoint specified by "
|
||||
"'--epoch' and '--iter'",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--use-averaged-model",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="Whether to load averaged model. Currently it only supports "
|
||||
"using --epoch. If True, it would decode with the averaged model "
|
||||
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
|
||||
"Actually only the models with epoch number of `epoch-avg` and "
|
||||
"`epoch` are loaded for averaging. ",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--exp-dir",
|
||||
type=str,
|
||||
default="pruned_transducer_stateless5/exp",
|
||||
help="The experiment dir",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--lang-dir",
|
||||
type=str,
|
||||
default="data/lang_char",
|
||||
help="""The lang dir
|
||||
It contains language related input files such as
|
||||
"lexicon.txt"
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--decoding-method",
|
||||
type=str,
|
||||
default="greedy_search",
|
||||
help="""Support only greedy_search and fast_beam_search now.
|
||||
""",
|
||||
)
|
||||
|
||||
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(
|
||||
"--decode-chunk-size",
|
||||
type=int,
|
||||
default=16,
|
||||
help="The chunk size for decoding (in frames after subsampling)",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--left-context",
|
||||
type=int,
|
||||
default=64,
|
||||
help="left context can be seen during decoding (in frames after subsampling)",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--right-context",
|
||||
type=int,
|
||||
default=0,
|
||||
help="right context can be seen during decoding (in frames after subsampling)",
|
||||
)
|
||||
|
||||
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 greedy_search(
|
||||
model: nn.Module,
|
||||
encoder_out: torch.Tensor,
|
||||
streams: List[DecodeStream],
|
||||
) -> List[List[int]]:
|
||||
|
||||
assert len(streams) == encoder_out.size(0)
|
||||
assert encoder_out.ndim == 3
|
||||
|
||||
blank_id = model.decoder.blank_id
|
||||
context_size = model.decoder.context_size
|
||||
device = model.device
|
||||
T = encoder_out.size(1)
|
||||
|
||||
decoder_input = torch.tensor(
|
||||
[stream.hyp[-context_size:] for stream in streams],
|
||||
device=device,
|
||||
dtype=torch.int64,
|
||||
)
|
||||
# decoder_out is of shape (N, decoder_out_dim)
|
||||
decoder_out = model.decoder(decoder_input, need_pad=False)
|
||||
decoder_out = model.joiner.decoder_proj(decoder_out)
|
||||
|
||||
for t in range(T):
|
||||
# current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
|
||||
current_encoder_out = encoder_out[:, t : t + 1, :] # noqa
|
||||
# print("encoder_out shape: ", current_encoder_out.shape, "decoder_out shape: ", decoder_out.shape)
|
||||
logits = model.joiner(
|
||||
current_encoder_out.unsqueeze(2),
|
||||
decoder_out.unsqueeze(1),
|
||||
project_input=False,
|
||||
)
|
||||
# logits'shape (batch_size, vocab_size)
|
||||
logits = logits.squeeze(1).squeeze(1)
|
||||
|
||||
assert logits.ndim == 2, logits.shape
|
||||
y = logits.argmax(dim=1).tolist()
|
||||
emitted = False
|
||||
for i, v in enumerate(y):
|
||||
if v != blank_id:
|
||||
streams[i].hyp.append(v)
|
||||
emitted = True
|
||||
if emitted:
|
||||
# update decoder output
|
||||
decoder_input = torch.tensor(
|
||||
[stream.hyp[-context_size:] for stream in streams],
|
||||
device=device,
|
||||
dtype=torch.int64,
|
||||
)
|
||||
decoder_out = model.decoder(
|
||||
decoder_input,
|
||||
need_pad=False,
|
||||
)
|
||||
decoder_out = model.joiner.decoder_proj(decoder_out)
|
||||
|
||||
hyp_tokens = []
|
||||
for stream in streams:
|
||||
hyp_tokens.append(stream.hyp)
|
||||
return hyp_tokens
|
||||
|
||||
|
||||
def fast_beam_search(
|
||||
model: nn.Module,
|
||||
encoder_out: torch.Tensor,
|
||||
processed_lens: torch.Tensor,
|
||||
decoding_streams: k2.RnntDecodingStreams,
|
||||
) -> List[List[int]]:
|
||||
|
||||
B, T, C = encoder_out.shape
|
||||
for t in range(T):
|
||||
# shape is a RaggedShape of shape (B, context)
|
||||
# contexts is a Tensor of shape (shape.NumElements(), context_size)
|
||||
shape, contexts = decoding_streams.get_contexts()
|
||||
# `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
|
||||
contexts = contexts.to(torch.int64)
|
||||
# decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
|
||||
decoder_out = model.decoder(contexts, need_pad=False)
|
||||
# current_encoder_out is of shape
|
||||
# (shape.NumElements(), 1, joiner_dim)
|
||||
# fmt: off
|
||||
current_encoder_out = torch.index_select(
|
||||
encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
|
||||
)
|
||||
# fmt: on
|
||||
logits = model.joiner(
|
||||
current_encoder_out.unsqueeze(2),
|
||||
decoder_out.unsqueeze(1),
|
||||
)
|
||||
logits = logits.squeeze(1).squeeze(1)
|
||||
log_probs = logits.log_softmax(dim=-1)
|
||||
decoding_streams.advance(log_probs)
|
||||
|
||||
decoding_streams.terminate_and_flush_to_streams()
|
||||
|
||||
lattice = decoding_streams.format_output(processed_lens.tolist())
|
||||
best_path = one_best_decoding(lattice)
|
||||
hyp_tokens = get_texts(best_path)
|
||||
return hyp_tokens
|
||||
|
||||
|
||||
def decode_one_chunk(
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
decode_streams: List[DecodeStream],
|
||||
) -> List[int]:
|
||||
"""Decode one chunk frames of features for each decode_streams and
|
||||
return the indexes of finished streams in a List.
|
||||
|
||||
Args:
|
||||
params:
|
||||
It's the return value of :func:`get_params`.
|
||||
model:
|
||||
The neural model.
|
||||
decode_streams:
|
||||
A List of DecodeStream, each belonging to a utterance.
|
||||
Returns:
|
||||
Return a List containing which DecodeStreams are finished.
|
||||
"""
|
||||
device = model.device
|
||||
|
||||
features = []
|
||||
feature_lens = []
|
||||
states = []
|
||||
|
||||
rnnt_stream_list = []
|
||||
processed_lens = []
|
||||
|
||||
for stream in decode_streams:
|
||||
feat, feat_len = stream.get_feature_frames(
|
||||
params.decode_chunk_size * params.subsampling_factor
|
||||
)
|
||||
features.append(feat)
|
||||
feature_lens.append(feat_len)
|
||||
states.append(stream.states)
|
||||
processed_lens.append(stream.done_frames)
|
||||
if params.decoding_method == "fast_beam_search":
|
||||
rnnt_stream_list.append(stream.rnnt_decoding_stream)
|
||||
|
||||
feature_lens = torch.tensor(feature_lens, device=device)
|
||||
features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
|
||||
|
||||
# if T is less than 7 there will be an error in time reduction layer,
|
||||
# because we subsample features with ((x_len - 1) // 2 - 1) // 2
|
||||
# we plus 2 here because we will cut off one frame on each size of
|
||||
# encoder_embed output as they see invalid paddings. so we need extra 2
|
||||
# frames.
|
||||
tail_length = 7 + (2 + params.right_context) * params.subsampling_factor
|
||||
if features.size(1) < tail_length:
|
||||
feature_lens += tail_length - features.size(1)
|
||||
features = torch.cat(
|
||||
[
|
||||
features,
|
||||
torch.tensor(
|
||||
LOG_EPS, dtype=features.dtype, device=device
|
||||
).expand(
|
||||
features.size(0),
|
||||
tail_length - features.size(1),
|
||||
features.size(2),
|
||||
),
|
||||
],
|
||||
dim=1,
|
||||
)
|
||||
|
||||
states = [
|
||||
torch.stack([x[0] for x in states], dim=2),
|
||||
torch.stack([x[1] for x in states], dim=2),
|
||||
]
|
||||
|
||||
processed_lens = torch.tensor(processed_lens, device=device)
|
||||
|
||||
encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
|
||||
x=features,
|
||||
x_lens=feature_lens,
|
||||
states=states,
|
||||
left_context=params.left_context,
|
||||
right_context=params.right_context,
|
||||
processed_lens=processed_lens,
|
||||
)
|
||||
|
||||
encoder_out = model.joiner.encoder_proj(encoder_out)
|
||||
|
||||
if params.decoding_method == "greedy_search":
|
||||
hyp_tokens = greedy_search(model, encoder_out, decode_streams)
|
||||
elif params.decoding_method == "fast_beam_search":
|
||||
config = k2.RnntDecodingConfig(
|
||||
vocab_size=params.vocab_size,
|
||||
decoder_history_len=params.context_size,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
)
|
||||
decoding_streams = k2.RnntDecodingStreams(rnnt_stream_list, config)
|
||||
processed_lens = processed_lens + encoder_out_lens
|
||||
hyp_tokens = fast_beam_search(
|
||||
model, encoder_out, processed_lens, decoding_streams
|
||||
)
|
||||
else:
|
||||
assert False
|
||||
|
||||
states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)]
|
||||
|
||||
finished_streams = []
|
||||
for i in range(len(decode_streams)):
|
||||
decode_streams[i].states = [states[0][i], states[1][i]]
|
||||
decode_streams[i].done_frames += encoder_out_lens[i]
|
||||
if params.decoding_method == "fast_beam_search":
|
||||
decode_streams[i].hyp = hyp_tokens[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.
|
||||
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 = []
|
||||
initial_states = model.encoder.get_init_state(
|
||||
params.left_context, device=device
|
||||
)
|
||||
for num, cut in enumerate(cuts):
|
||||
# each utterance has a DecodeStream.
|
||||
decode_stream = DecodeStream(
|
||||
params=params,
|
||||
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
|
||||
|
||||
samples = torch.from_numpy(audio).squeeze(0)
|
||||
|
||||
fbank = Fbank(opts)
|
||||
decode_stream.set_features(fbank(samples.to(device)))
|
||||
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):
|
||||
hyp = decode_streams[i].hyp
|
||||
if params.decoding_method == "greedy_search":
|
||||
hyp = hyp[params.context_size :] # noqa
|
||||
decode_results.append(
|
||||
(
|
||||
list(decode_streams[i].ground_truth),
|
||||
[lexicon.token_table[idx] for idx in hyp],
|
||||
)
|
||||
)
|
||||
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):
|
||||
hyp = decode_streams[i].hyp
|
||||
if params.decoding_method == "greedy_search":
|
||||
hyp = hyp[params.context_size :] # noqa
|
||||
decode_results.append(
|
||||
(
|
||||
list(decode_streams[i].ground_truth),
|
||||
[lexicon.token_table[idx] for idx in hyp],
|
||||
)
|
||||
)
|
||||
del decode_streams[i]
|
||||
|
||||
key = "greedy_search"
|
||||
if params.decoding_method == "fast_beam_search":
|
||||
key = (
|
||||
f"beam_{params.beam}_"
|
||||
f"max_contexts_{params.max_contexts}_"
|
||||
f"max_states_{params.max_states}"
|
||||
)
|
||||
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"
|
||||
)
|
||||
# sort results so we can easily compare the difference between two
|
||||
# recognition results
|
||||
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()
|
||||
WenetSpeechAsrDataModule.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}"
|
||||
|
||||
# for streaming
|
||||
params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
|
||||
params.suffix += f"-left-context-{params.left_context}"
|
||||
params.suffix += f"-right-context-{params.right_context}"
|
||||
|
||||
# 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}"
|
||||
|
||||
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
|
||||
|
||||
params.causal_convolution = True
|
||||
|
||||
logging.info(params)
|
||||
|
||||
logging.info("About to create model")
|
||||
model = get_transducer_model(params)
|
||||
|
||||
if not params.use_averaged_model:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(
|
||||
params.exp_dir, iteration=-params.iter
|
||||
)[: params.avg]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
elif params.avg == 1:
|
||||
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
|
||||
else:
|
||||
start = params.epoch - params.avg + 1
|
||||
filenames = []
|
||||
for i in range(start, params.epoch + 1):
|
||||
if i >= 1:
|
||||
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
else:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(
|
||||
params.exp_dir, iteration=-params.iter
|
||||
)[: params.avg + 1]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg + 1:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
filename_start = filenames[-1]
|
||||
filename_end = filenames[0]
|
||||
logging.info(
|
||||
"Calculating the averaged model over iteration checkpoints"
|
||||
f" from {filename_start} (excluded) to {filename_end}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
else:
|
||||
assert params.avg > 0, params.avg
|
||||
start = params.epoch - params.avg
|
||||
assert start >= 1, start
|
||||
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
|
||||
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
|
||||
logging.info(
|
||||
f"Calculating the averaged model over epoch range from "
|
||||
f"{start} (excluded) to {params.epoch}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
|
||||
model.to(device)
|
||||
model.eval()
|
||||
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}")
|
||||
|
||||
wenetspeech = WenetSpeechAsrDataModule(args)
|
||||
|
||||
dev_cuts = wenetspeech.valid_cuts()
|
||||
test_net_cuts = wenetspeech.test_net_cuts()
|
||||
test_meeting_cuts = wenetspeech.test_meeting_cuts()
|
||||
|
||||
test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
|
||||
test_cuts = [dev_cuts, test_net_cuts, test_meeting_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()
|
||||
Loading…
x
Reference in New Issue
Block a user