mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-12-08 21:45:27 +00:00
Streaming Zipformer with multi-dataset (#984)
* modify train.py * add right padding option in decode.py * update RESULTS.md
This commit is contained in:
marcoyang1998
GitHub
parent
0efed1cec5
commit
57d6482a79
@ -387,14 +387,16 @@ class GigaSpeechAsrDataModule:
|
||||
@lru_cache()
|
||||
def train_cuts(self) -> CutSet:
|
||||
logging.info(f"About to get train_{self.args.subset} cuts")
|
||||
path = self.args.manifest_dir / f"cuts_{self.args.subset}.jsonl.gz"
|
||||
path = self.args.manifest_dir / f"gigaspeech_cuts_{self.args.subset}.jsonl.gz"
|
||||
cuts_train = CutSet.from_jsonl_lazy(path)
|
||||
return cuts_train
|
||||
|
||||
@lru_cache()
|
||||
def dev_cuts(self) -> CutSet:
|
||||
logging.info("About to get dev cuts")
|
||||
cuts_valid = load_manifest_lazy(self.args.manifest_dir / "cuts_DEV.jsonl.gz")
|
||||
cuts_valid = load_manifest_lazy(
|
||||
self.args.manifest_dir / "gigaspeech_cuts_DEV.jsonl.gz"
|
||||
)
|
||||
if self.args.small_dev:
|
||||
return cuts_valid.subset(first=1000)
|
||||
else:
|
||||
@ -403,4 +405,6 @@ class GigaSpeechAsrDataModule:
|
||||
@lru_cache()
|
||||
def test_cuts(self) -> CutSet:
|
||||
logging.info("About to get test cuts")
|
||||
return load_manifest_lazy(self.args.manifest_dir / "cuts_TEST.jsonl.gz")
|
||||
return load_manifest_lazy(
|
||||
self.args.manifest_dir / "gigaspeech_cuts_TEST.jsonl.gz"
|
||||
)
|
||||
|
||||
3
egs/librispeech/ASR/.gitignore
vendored
3
egs/librispeech/ASR/.gitignore
vendored
@ -1,2 +1,3 @@
|
||||
log-*
|
||||
.DS_Store
|
||||
.DS_Store
|
||||
run*.sh
|
||||
|
||||
@ -26,6 +26,7 @@ The following table lists the differences among them.
|
||||
| `pruned_transducer_stateless7_ctc` | Zipformer | Embedding + Conv1d | Same as pruned_transducer_stateless7, but with extra CTC head|
|
||||
| `pruned_transducer_stateless7_ctc_bs` | Zipformer | Embedding + Conv1d | pruned_transducer_stateless7_ctc + blank skip |
|
||||
| `pruned_transducer_stateless7_streaming` | Streaming Zipformer | Embedding + Conv1d | streaming version of pruned_transducer_stateless7 |
|
||||
| `pruned_transducer_stateless7_streaming_multi` | Streaming Zipformer | Embedding + Conv1d | same as pruned_transducer_stateless7_streaming, trained on LibriSpeech + GigaSpeech |
|
||||
| `pruned_transducer_stateless8` | Zipformer | Embedding + Conv1d | Same as pruned_transducer_stateless7, but using extra data from GigaSpeech|
|
||||
| `pruned_stateless_emformer_rnnt2` | Emformer(from torchaudio) | Embedding + Conv1d | Using Emformer from torchaudio for streaming ASR|
|
||||
| `conv_emformer_transducer_stateless` | ConvEmformer | Embedding + Conv1d | Using ConvEmformer for streaming ASR + mechanisms in reworked model |
|
||||
|
||||
@ -1,5 +1,141 @@
|
||||
## Results
|
||||
|
||||
### Streaming Zipformer-Transducer (Pruned Stateless Transducer + Streaming Zipformer + Multi-Dataset)
|
||||
|
||||
#### [pruned_transducer_stateless7_streaming_multi](./pruned_transducer_stateless7_streaming_multi)
|
||||
|
||||
See <https://github.com/k2-fsa/icefall/pull/984> for more details.
|
||||
|
||||
You can find a pretrained model, training logs, decoding logs, and decoding
|
||||
results at: <https://huggingface.co/marcoyang/icefall-libri-giga-pruned-transducer-stateless7-streaming-2023-04-04>
|
||||
|
||||
Number of model parameters: 70369391, i.e., 70.37 M
|
||||
|
||||
##### training on full librispeech + full gigaspeech (with giga_prob=0.9)
|
||||
|
||||
The WERs are:
|
||||
|
||||
|
||||
| decoding method | chunk size | test-clean | test-other | comment | decoding mode |
|
||||
|----------------------|------------|------------|------------|---------------------|----------------------|
|
||||
| greedy search | 320ms | 2.43 | 6.0 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| greedy search | 320ms | 2.47 | 6.13 | --epoch 20 --avg 4 | chunk-wise |
|
||||
| fast beam search | 320ms | 2.43 | 5.99 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| fast beam search | 320ms | 2.8 | 6.46 | --epoch 20 --avg 4 | chunk-wise |
|
||||
| modified beam search | 320ms | 2.4 | 5.96 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| modified beam search | 320ms | 2.42 | 6.03 | --epoch 20 --avg 4 | chunk-size |
|
||||
| greedy search | 640ms | 2.26 | 5.58 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| greedy search | 640ms | 2.33 | 5.76 | --epoch 20 --avg 4 | chunk-wise |
|
||||
| fast beam search | 640ms | 2.27 | 5.54 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| fast beam search | 640ms | 2.37 | 5.75 | --epoch 20 --avg 4 | chunk-wise |
|
||||
| modified beam search | 640ms | 2.22 | 5.5 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| modified beam search | 640ms | 2.25 | 5.69 | --epoch 20 --avg 4 | chunk-size |
|
||||
|
||||
The model also has good WERs on GigaSpeech. The following WERs are achieved on GigaSpeech test and dev sets:
|
||||
|
||||
| decoding method | chunk size | dev | test | comment | decoding mode |
|
||||
|----------------------|------------|-----|------|------------|---------------------|
|
||||
| greedy search | 320ms | 12.08 | 11.98 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| greedy search | 640ms | 11.66 | 11.71 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| modified beam search | 320ms | 11.95 | 11.83 | --epoch 20 --avg 4 | simulated streaming |
|
||||
| modified beam search | 320ms | 11.65 | 11.56 | --epoch 20 --avg 4 | simulated streaming |
|
||||
|
||||
|
||||
Note: `simulated streaming` indicates feeding full utterance during decoding using `decode.py`,
|
||||
while `chunk-size` indicates feeding certain number of frames at each time using `streaming_decode.py`.
|
||||
|
||||
The training command is:
|
||||
|
||||
```bash
|
||||
./pruned_transducer_stateless7_streaming_multi/train.py \
|
||||
--world-size 4 \
|
||||
--num-epochs 20 \
|
||||
--start-epoch 1 \
|
||||
--use-fp16 1 \
|
||||
--exp-dir pruned_transducer_stateless7_streaming_multi/exp \
|
||||
--full-libri 1 \
|
||||
--giga-prob 0.9 \
|
||||
--max-duration 750 \
|
||||
--master-port 12345
|
||||
```
|
||||
|
||||
The tensorboard log can be found at
|
||||
<https://tensorboard.dev/experiment/G4yDMLXGQXexf41i4MA2Tg/#scalars>
|
||||
|
||||
The simulated streaming decoding command (e.g., chunk-size=320ms) is:
|
||||
```bash
|
||||
for m in greedy_search fast_beam_search modified_beam_search; do
|
||||
./pruned_transducer_stateless7_streaming_multi/decode.py \
|
||||
--epoch 20 \
|
||||
--avg 4 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming_multi/exp \
|
||||
--max-duration 600 \
|
||||
--decode-chunk-len 32 \
|
||||
--right-padding 64 \
|
||||
--decoding-method $m
|
||||
done
|
||||
```
|
||||
|
||||
The streaming chunk-size decoding command (e.g., chunk-size=320ms) is:
|
||||
```bash
|
||||
for m in greedy_search modified_beam_search fast_beam_search; do
|
||||
./pruned_transducer_stateless7_streaming_multi/streaming_decode.py \
|
||||
--epoch 20 \
|
||||
--avg 4 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming_multi/exp \
|
||||
--decoding-method $m \
|
||||
--decode-chunk-len 32 \
|
||||
--num-decode-streams 2000
|
||||
done
|
||||
```
|
||||
|
||||
|
||||
#### Smaller model
|
||||
|
||||
We also provide a very small version (only 6.1M parameters) of this setup. The training command for the small model is:
|
||||
|
||||
```bash
|
||||
./pruned_transducer_stateless7_streaming_multi/train.py \
|
||||
--world-size 4 \
|
||||
--num-epochs 30 \
|
||||
--start-epoch 1 \
|
||||
--use-fp16 1 \
|
||||
--exp-dir pruned_transducer_stateless7_streaming_multi/exp \
|
||||
--full-libri 1 \
|
||||
--giga-prob 0.9 \
|
||||
--num-encoder-layers "2,2,2,2,2" \
|
||||
--feedforward-dims "256,256,512,512,256" \
|
||||
--nhead "4,4,4,4,4" \
|
||||
--encoder-dims "128,128,128,128,128" \
|
||||
--attention-dims "96,96,96,96,96" \
|
||||
--encoder-unmasked-dims "96,96,96,96,96" \
|
||||
--max-duration 1200 \
|
||||
--master-port 12345
|
||||
```
|
||||
|
||||
You can find this pretrained small model and its training logs, decoding logs, and decoding
|
||||
results at:
|
||||
<https://huggingface.co/marcoyang/icefall-libri-giga-pruned-transducer-stateless7-streaming-6M-2023-04-03>
|
||||
|
||||
|
||||
| decoding method | chunk size | test-clean | test-other | comment | decoding mode |
|
||||
|----------------------|------------|------------|------------|---------------------|----------------------|
|
||||
| greedy search | 320ms | 5.95 | 15.03 | --epoch 30 --avg 1 | simulated streaming |
|
||||
| greedy search | 640ms | 5.61 | 13.86 | --epoch 30 --avg 1 | simulated streaming |
|
||||
| modified beam search | 320ms | 5.72 | 14.34 | --epoch 30 --avg 1 | simulated streaming |
|
||||
| modified beam search | 640ms | 5.43 | 13.16 | --epoch 30 --avg 1 | simulated streaming |
|
||||
| fast beam search | 320ms | 5.88 | 14.45 | --epoch 30 --avg 1 | simulated streaming |
|
||||
| fast beam search | 640ms | 5.48 | 13.31 | --epoch 30 --avg 1 | simulated streaming |
|
||||
|
||||
This small model achieves the following WERs on GigaSpeech test and dev sets:
|
||||
|
||||
| decoding method | chunk size | dev | test | comment | decoding mode |
|
||||
|----------------------|------------|------------|------------|---------------------|----------------------|
|
||||
| greedy search | 320ms | 17.57 | 17.2 | --epoch 30 --avg 1 | simulated streaming |
|
||||
| modified beam search | 320ms | 16.98 | 11.98 | --epoch 30 --avg 1 | simulated streaming |
|
||||
|
||||
You can find the tensorboard logs at <https://tensorboard.dev/experiment/tAc5iXxTQrCQxky5O5OLyw/#scalars>.
|
||||
|
||||
### Streaming Zipformer-Transducer (Pruned Stateless Transducer + Streaming Zipformer)
|
||||
|
||||
#### [pruned_transducer_stateless7_streaming](./pruned_transducer_stateless7_streaming)
|
||||
@ -53,7 +189,7 @@ The tensorboard log can be found at
|
||||
|
||||
The simulated streaming decoding command (e.g., chunk-size=320ms) is:
|
||||
```bash
|
||||
for $m in greedy_search fast_beam_search modified_beam_search; do
|
||||
for m in greedy_search fast_beam_search modified_beam_search; do
|
||||
./pruned_transducer_stateless7_streaming/decode.py \
|
||||
--epoch 30 \
|
||||
--avg 9 \
|
||||
@ -599,7 +735,7 @@ done
|
||||
```
|
||||
|
||||
Note that a small change is made to the `pruned_transducer_stateless7/decoder.py` in
|
||||
this [PR](/ceph-data4/yangxiaoyu/softwares/icefall_development/icefall_random_padding/egs/librispeech/ASR/pruned_transducer_stateless7/exp_960h_no_paddingidx_ngpu4/tensorboard) to address the
|
||||
this [PR](https://github.com/k2-fsa/icefall/pull/942) to address the
|
||||
problem of emitting the first symbol at the very beginning. If you need a
|
||||
model without this issue, please download the model from here: <https://huggingface.co/marcoyang/icefall-asr-librispeech-pruned-transducer-stateless7-2023-03-10>
|
||||
|
||||
|
||||
@ -777,8 +777,9 @@ class WithLoss(torch.autograd.Function):
|
||||
|
||||
@staticmethod
|
||||
def backward(ctx, ans_grad: Tensor):
|
||||
return ans_grad, torch.ones(
|
||||
ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device
|
||||
return (
|
||||
ans_grad,
|
||||
torch.ones(ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device),
|
||||
)
|
||||
|
||||
|
||||
|
||||
@ -33,7 +33,6 @@ popd
|
||||
--use-averaged-model 0 \
|
||||
--epoch 99 \
|
||||
--avg 1 \
|
||||
\
|
||||
--decode-chunk-len 32 \
|
||||
--num-encoder-layers "2,4,3,2,4" \
|
||||
--feedforward-dims "1024,1024,2048,2048,1024" \
|
||||
|
||||
@ -71,13 +71,12 @@ It will generate the following 3 files inside $repo/exp:
|
||||
import argparse
|
||||
import logging
|
||||
|
||||
import torch
|
||||
from onnx_pretrained import OnnxModel
|
||||
from zipformer import stack_states
|
||||
|
||||
from icefall import is_module_available
|
||||
|
||||
import torch
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
|
||||
@ -109,6 +109,29 @@ class Model:
|
||||
self.init_joiner(args)
|
||||
|
||||
# Please change the parameters according to your model
|
||||
|
||||
# 20M
|
||||
# self.num_encoder_layers = to_int_tuple("2,2,2,2,2")
|
||||
# self.encoder_dims = to_int_tuple("256,256,256,256,256") # also known as d_model
|
||||
# self.attention_dims = to_int_tuple("192,192,192,192,192")
|
||||
# self.zipformer_downsampling_factors = to_int_tuple("1,2,4,8,2")
|
||||
# self.cnn_module_kernels = to_int_tuple("31,31,31,31,31")
|
||||
|
||||
# 9.6M
|
||||
# self.num_encoder_layers = to_int_tuple("2,3,2,2,3")
|
||||
# self.encoder_dims = to_int_tuple("160,160,160,160,160") # also known as d_model
|
||||
# self.attention_dims = to_int_tuple("96,96,96,96,96")
|
||||
# self.zipformer_downsampling_factors = to_int_tuple("1,2,4,8,2")
|
||||
# self.cnn_module_kernels = to_int_tuple("31,31,31,31,31")
|
||||
|
||||
# 5.5M or 6M
|
||||
|
||||
# self.num_encoder_layers = to_int_tuple("2,2,2,2,2")
|
||||
# self.encoder_dims = to_int_tuple("128,128,128,128,128") # also known as d_model
|
||||
# self.attention_dims = to_int_tuple("96,96,96,96,96")
|
||||
# self.zipformer_downsampling_factors = to_int_tuple("1,2,4,8,2")
|
||||
# self.cnn_module_kernels = to_int_tuple("31,31,31,31,31")
|
||||
|
||||
self.num_encoder_layers = to_int_tuple("2,4,3,2,4")
|
||||
self.encoder_dims = to_int_tuple("384,384,384,384,384") # also known as d_model
|
||||
self.attention_dims = to_int_tuple("192,192,192,192,192")
|
||||
|
||||
@ -62,6 +62,43 @@ def test_model():
|
||||
model = torch.jit.script(model)
|
||||
|
||||
|
||||
def test_model_small():
|
||||
params = get_params()
|
||||
params.vocab_size = 500
|
||||
params.blank_id = 0
|
||||
params.context_size = 2
|
||||
params.num_encoder_layers = "2,2,2,2,2"
|
||||
params.feedforward_dims = "256,256,512,512,256"
|
||||
params.nhead = "4,4,4,4,4"
|
||||
params.encoder_dims = "128,128,128,128,128"
|
||||
params.attention_dims = "96,96,96,96,96"
|
||||
params.encoder_unmasked_dims = "96,96,96,96,96"
|
||||
params.zipformer_downsampling_factors = "1,2,4,8,2"
|
||||
params.cnn_module_kernels = "31,31,31,31,31"
|
||||
params.decoder_dim = 320
|
||||
params.joiner_dim = 320
|
||||
params.num_left_chunks = 4
|
||||
params.short_chunk_size = 50
|
||||
params.decode_chunk_len = 32
|
||||
model = get_transducer_model(params)
|
||||
|
||||
num_param = sum([p.numel() for p in model.parameters()])
|
||||
print(f"Number of model parameters: {num_param}")
|
||||
import pdb
|
||||
|
||||
pdb.set_trace()
|
||||
|
||||
# Test jit script
|
||||
convert_scaled_to_non_scaled(model, inplace=True)
|
||||
# We won't use the forward() method of the model in C++, so just ignore
|
||||
# it here.
|
||||
# Otherwise, one of its arguments is a ragged tensor and is not
|
||||
# torch scriptabe.
|
||||
model.__class__.forward = torch.jit.ignore(model.__class__.forward)
|
||||
print("Using torch.jit.script")
|
||||
model = torch.jit.script(model)
|
||||
|
||||
|
||||
def test_model_jit_trace():
|
||||
params = get_params()
|
||||
params.vocab_size = 500
|
||||
@ -142,7 +179,7 @@ def test_model_jit_trace():
|
||||
|
||||
|
||||
def main():
|
||||
test_model()
|
||||
test_model_small()
|
||||
test_model_jit_trace()
|
||||
|
||||
|
||||
|
||||
@ -1049,10 +1049,10 @@ def run(rank, world_size, args):
|
||||
|
||||
librispeech = LibriSpeechAsrDataModule(args)
|
||||
|
||||
train_cuts = librispeech.train_clean_100_cuts()
|
||||
if params.full_libri:
|
||||
train_cuts = librispeech.train_all_shuf_cuts()
|
||||
else:
|
||||
train_cuts = librispeech.train_clean_100_cuts()
|
||||
train_cuts += librispeech.train_clean_360_cuts()
|
||||
train_cuts += librispeech.train_other_500_cuts()
|
||||
|
||||
def remove_short_and_long_utt(c: Cut):
|
||||
# Keep only utterances with duration between 1 second and 20 seconds
|
||||
@ -1091,7 +1091,7 @@ def run(rank, world_size, args):
|
||||
|
||||
return True
|
||||
|
||||
train_cuts = train_cuts.filter(remove_short_and_long_utt)
|
||||
# train_cuts = train_cuts.filter(remove_short_and_long_utt)
|
||||
|
||||
if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
|
||||
# We only load the sampler's state dict when it loads a checkpoint
|
||||
@ -1108,14 +1108,14 @@ def run(rank, world_size, args):
|
||||
valid_cuts += librispeech.dev_other_cuts()
|
||||
valid_dl = librispeech.valid_dataloaders(valid_cuts)
|
||||
|
||||
if not params.print_diagnostics:
|
||||
scan_pessimistic_batches_for_oom(
|
||||
model=model,
|
||||
train_dl=train_dl,
|
||||
optimizer=optimizer,
|
||||
sp=sp,
|
||||
params=params,
|
||||
)
|
||||
# if not params.print_diagnostics:
|
||||
# scan_pessimistic_batches_for_oom(
|
||||
# model=model,
|
||||
# train_dl=train_dl,
|
||||
# optimizer=optimizer,
|
||||
# sp=sp,
|
||||
# params=params,
|
||||
# )
|
||||
|
||||
scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
|
||||
if checkpoints and "grad_scaler" in checkpoints:
|
||||
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless8/asr_datamodule.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless2/beam_search.py
|
||||
818
egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode.py
Executable file
818
egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode.py
Executable file
@ -0,0 +1,818 @@
|
||||
#!/usr/bin/env python3
|
||||
#
|
||||
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
|
||||
# Zengwei Yao)
|
||||
#
|
||||
# See ../../../../LICENSE for clarification regarding multiple authors
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""
|
||||
Usage:
|
||||
(1) greedy search
|
||||
./pruned_transducer_stateless7_streaming/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
|
||||
--max-duration 600 \
|
||||
--decode-chunk-len 32 \
|
||||
--decoding-method greedy_search
|
||||
|
||||
(2) beam search (not recommended)
|
||||
./pruned_transducer_stateless7_streaming/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
|
||||
--max-duration 600 \
|
||||
--decode-chunk-len 32 \
|
||||
--decoding-method beam_search \
|
||||
--beam-size 4
|
||||
|
||||
(3) modified beam search
|
||||
./pruned_transducer_stateless7_streaming/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
|
||||
--max-duration 600 \
|
||||
--decode-chunk-len 32 \
|
||||
--decoding-method modified_beam_search \
|
||||
--beam-size 4
|
||||
|
||||
(4) fast beam search (one best)
|
||||
./pruned_transducer_stateless7_streaming/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
|
||||
--max-duration 600 \
|
||||
--decode-chunk-len 32 \
|
||||
--decoding-method fast_beam_search \
|
||||
--beam 20.0 \
|
||||
--max-contexts 8 \
|
||||
--max-states 64
|
||||
|
||||
(5) fast beam search (nbest)
|
||||
./pruned_transducer_stateless7_streaming/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
|
||||
--max-duration 600 \
|
||||
--decode-chunk-len 32 \
|
||||
--decoding-method fast_beam_search_nbest \
|
||||
--beam 20.0 \
|
||||
--max-contexts 8 \
|
||||
--max-states 64 \
|
||||
--num-paths 200 \
|
||||
--nbest-scale 0.5
|
||||
|
||||
(6) fast beam search (nbest oracle WER)
|
||||
./pruned_transducer_stateless7_streaming/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
|
||||
--max-duration 600 \
|
||||
--decode-chunk-len 32 \
|
||||
--decoding-method fast_beam_search_nbest_oracle \
|
||||
--beam 20.0 \
|
||||
--max-contexts 8 \
|
||||
--max-states 64 \
|
||||
--num-paths 200 \
|
||||
--nbest-scale 0.5
|
||||
|
||||
(7) fast beam search (with LG)
|
||||
./pruned_transducer_stateless7_streaming/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
|
||||
--max-duration 600 \
|
||||
--decode-chunk-len 32 \
|
||||
--decoding-method fast_beam_search_nbest_LG \
|
||||
--beam 20.0 \
|
||||
--max-contexts 8 \
|
||||
--max-states 64
|
||||
"""
|
||||
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import math
|
||||
from collections import defaultdict
|
||||
from pathlib import Path
|
||||
from typing import Dict, List, Optional, Tuple
|
||||
|
||||
import k2
|
||||
import sentencepiece as spm
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from asr_datamodule import AsrDataModule
|
||||
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 librispeech import LibriSpeech
|
||||
from train import add_model_arguments, get_params, get_transducer_model
|
||||
|
||||
from icefall.checkpoint import (
|
||||
average_checkpoints,
|
||||
average_checkpoints_with_averaged_model,
|
||||
find_checkpoints,
|
||||
load_checkpoint,
|
||||
)
|
||||
from icefall.lexicon import Lexicon
|
||||
from icefall.utils import (
|
||||
AttributeDict,
|
||||
setup_logger,
|
||||
store_transcripts,
|
||||
str2bool,
|
||||
write_error_stats,
|
||||
)
|
||||
|
||||
LOG_EPS = math.log(1e-10)
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--epoch",
|
||||
type=int,
|
||||
default=30,
|
||||
help="""It specifies the checkpoint to use for decoding.
|
||||
Note: Epoch counts from 1.
|
||||
You can specify --avg to use more checkpoints for model averaging.""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--iter",
|
||||
type=int,
|
||||
default=0,
|
||||
help="""If positive, --epoch is ignored and it
|
||||
will use the checkpoint exp_dir/checkpoint-iter.pt.
|
||||
You can specify --avg to use more checkpoints for model averaging.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--avg",
|
||||
type=int,
|
||||
default=9,
|
||||
help="Number of checkpoints to average. Automatically select "
|
||||
"consecutive checkpoints before the checkpoint specified by "
|
||||
"'--epoch' and '--iter'",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--use-averaged-model",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="Whether to load averaged model. Currently it only supports "
|
||||
"using --epoch. If True, it would decode with the averaged model "
|
||||
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
|
||||
"Actually only the models with epoch number of `epoch-avg` and "
|
||||
"`epoch` are loaded for averaging. ",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--exp-dir",
|
||||
type=str,
|
||||
default="pruned_transducer_stateless7_streaming/exp",
|
||||
help="The experiment dir",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--bpe-model",
|
||||
type=str,
|
||||
default="data/lang_bpe_500/bpe.model",
|
||||
help="Path to the BPE model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--lang-dir",
|
||||
type=Path,
|
||||
default="data/lang_bpe_500",
|
||||
help="The lang dir containing word table and LG graph",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--decoding-method",
|
||||
type=str,
|
||||
default="greedy_search",
|
||||
help="""Possible values are:
|
||||
- greedy_search
|
||||
- beam_search
|
||||
- modified_beam_search
|
||||
- fast_beam_search
|
||||
- fast_beam_search_nbest
|
||||
- fast_beam_search_nbest_oracle
|
||||
- fast_beam_search_nbest_LG
|
||||
If you use fast_beam_search_nbest_LG, you have to specify
|
||||
`--lang-dir`, which should contain `LG.pt`.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--beam-size",
|
||||
type=int,
|
||||
default=4,
|
||||
help="""An integer indicating how many candidates we will keep for each
|
||||
frame. Used only when --decoding-method is beam_search or
|
||||
modified_beam_search.""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--beam",
|
||||
type=float,
|
||||
default=20.0,
|
||||
help="""A floating point value to calculate the cutoff score during beam
|
||||
search (i.e., `cutoff = max-score - beam`), which is the same as the
|
||||
`beam` in Kaldi.
|
||||
Used only when --decoding-method is fast_beam_search,
|
||||
fast_beam_search_nbest, fast_beam_search_nbest_LG,
|
||||
and fast_beam_search_nbest_oracle
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--ngram-lm-scale",
|
||||
type=float,
|
||||
default=0.01,
|
||||
help="""
|
||||
Used only when --decoding_method is fast_beam_search_nbest_LG.
|
||||
It specifies the scale for n-gram LM scores.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--max-contexts",
|
||||
type=int,
|
||||
default=8,
|
||||
help="""Used only when --decoding-method is
|
||||
fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
|
||||
and fast_beam_search_nbest_oracle""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--max-states",
|
||||
type=int,
|
||||
default=64,
|
||||
help="""Used only when --decoding-method is
|
||||
fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
|
||||
and fast_beam_search_nbest_oracle""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--context-size",
|
||||
type=int,
|
||||
default=2,
|
||||
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--right-padding",
|
||||
type=int,
|
||||
default=64,
|
||||
help="Padding frames at the end of features",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--max-sym-per-frame",
|
||||
type=int,
|
||||
default=1,
|
||||
help="""Maximum number of symbols per frame.
|
||||
Used only when --decoding_method is greedy_search""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--num-paths",
|
||||
type=int,
|
||||
default=200,
|
||||
help="""Number of paths for nbest decoding.
|
||||
Used only when the decoding method is fast_beam_search_nbest,
|
||||
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--nbest-scale",
|
||||
type=float,
|
||||
default=0.5,
|
||||
help="""Scale applied to lattice scores when computing nbest paths.
|
||||
Used only when the decoding method is fast_beam_search_nbest,
|
||||
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
|
||||
)
|
||||
|
||||
add_model_arguments(parser)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
def decode_one_batch(
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
sp: spm.SentencePieceProcessor,
|
||||
batch: dict,
|
||||
word_table: Optional[k2.SymbolTable] = None,
|
||||
decoding_graph: Optional[k2.Fsa] = None,
|
||||
) -> Dict[str, List[List[str]]]:
|
||||
"""Decode one batch and return the result in a dict. The dict has the
|
||||
following format:
|
||||
|
||||
- key: It indicates the setting used for decoding. For example,
|
||||
if greedy_search is used, it would be "greedy_search"
|
||||
If beam search with a beam size of 7 is used, it would be
|
||||
"beam_7"
|
||||
- value: It contains the decoding result. `len(value)` equals to
|
||||
batch size. `value[i]` is the decoding result for the i-th
|
||||
utterance in the given batch.
|
||||
Args:
|
||||
params:
|
||||
It's the return value of :func:`get_params`.
|
||||
model:
|
||||
The neural model.
|
||||
sp:
|
||||
The BPE model.
|
||||
batch:
|
||||
It is the return value from iterating
|
||||
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
|
||||
for the format of the `batch`.
|
||||
word_table:
|
||||
The word symbol table.
|
||||
decoding_graph:
|
||||
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
|
||||
only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
|
||||
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
|
||||
Returns:
|
||||
Return the decoding result. See above description for the format of
|
||||
the returned dict.
|
||||
"""
|
||||
device = next(model.parameters()).device
|
||||
feature = batch["inputs"]
|
||||
assert feature.ndim == 3
|
||||
|
||||
feature = feature.to(device)
|
||||
# at entry, feature is (N, T, C)
|
||||
|
||||
supervisions = batch["supervisions"]
|
||||
feature_lens = supervisions["num_frames"].to(device)
|
||||
|
||||
feature_lens += params.right_padding
|
||||
feature = torch.nn.functional.pad(
|
||||
feature,
|
||||
pad=(0, 0, 0, params.right_padding),
|
||||
value=LOG_EPS,
|
||||
)
|
||||
encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
|
||||
|
||||
hyps = []
|
||||
|
||||
if params.decoding_method == "fast_beam_search":
|
||||
hyp_tokens = fast_beam_search_one_best(
|
||||
model=model,
|
||||
decoding_graph=decoding_graph,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
elif params.decoding_method == "fast_beam_search_nbest_LG":
|
||||
hyp_tokens = fast_beam_search_nbest_LG(
|
||||
model=model,
|
||||
decoding_graph=decoding_graph,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
num_paths=params.num_paths,
|
||||
nbest_scale=params.nbest_scale,
|
||||
)
|
||||
for hyp in hyp_tokens:
|
||||
hyps.append([word_table[i] for i in hyp])
|
||||
elif params.decoding_method == "fast_beam_search_nbest":
|
||||
hyp_tokens = fast_beam_search_nbest(
|
||||
model=model,
|
||||
decoding_graph=decoding_graph,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
num_paths=params.num_paths,
|
||||
nbest_scale=params.nbest_scale,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
elif params.decoding_method == "fast_beam_search_nbest_oracle":
|
||||
hyp_tokens = fast_beam_search_nbest_oracle(
|
||||
model=model,
|
||||
decoding_graph=decoding_graph,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
num_paths=params.num_paths,
|
||||
ref_texts=sp.encode(supervisions["text"]),
|
||||
nbest_scale=params.nbest_scale,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
|
||||
hyp_tokens = greedy_search_batch(
|
||||
model=model,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
elif params.decoding_method == "modified_beam_search":
|
||||
hyp_tokens = modified_beam_search(
|
||||
model=model,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam_size,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
else:
|
||||
batch_size = encoder_out.size(0)
|
||||
|
||||
for i in range(batch_size):
|
||||
# fmt: off
|
||||
encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
|
||||
# fmt: on
|
||||
if params.decoding_method == "greedy_search":
|
||||
hyp = greedy_search(
|
||||
model=model,
|
||||
encoder_out=encoder_out_i,
|
||||
max_sym_per_frame=params.max_sym_per_frame,
|
||||
)
|
||||
elif params.decoding_method == "beam_search":
|
||||
hyp = beam_search(
|
||||
model=model,
|
||||
encoder_out=encoder_out_i,
|
||||
beam=params.beam_size,
|
||||
)
|
||||
else:
|
||||
raise ValueError(
|
||||
f"Unsupported decoding method: {params.decoding_method}"
|
||||
)
|
||||
hyps.append(sp.decode(hyp).split())
|
||||
|
||||
if params.decoding_method == "greedy_search":
|
||||
return {"greedy_search": hyps}
|
||||
elif "fast_beam_search" in params.decoding_method:
|
||||
key = f"beam_{params.beam}_"
|
||||
key += f"max_contexts_{params.max_contexts}_"
|
||||
key += f"max_states_{params.max_states}"
|
||||
if "nbest" in params.decoding_method:
|
||||
key += f"_num_paths_{params.num_paths}_"
|
||||
key += f"nbest_scale_{params.nbest_scale}"
|
||||
if "LG" in params.decoding_method:
|
||||
key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
|
||||
|
||||
return {key: hyps}
|
||||
else:
|
||||
return {f"beam_size_{params.beam_size}": hyps}
|
||||
|
||||
|
||||
def decode_dataset(
|
||||
dl: torch.utils.data.DataLoader,
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
sp: spm.SentencePieceProcessor,
|
||||
word_table: Optional[k2.SymbolTable] = None,
|
||||
decoding_graph: Optional[k2.Fsa] = None,
|
||||
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
|
||||
"""Decode dataset.
|
||||
|
||||
Args:
|
||||
dl:
|
||||
PyTorch's dataloader containing the dataset to decode.
|
||||
params:
|
||||
It is returned by :func:`get_params`.
|
||||
model:
|
||||
The neural model.
|
||||
sp:
|
||||
The BPE model.
|
||||
word_table:
|
||||
The word symbol table.
|
||||
decoding_graph:
|
||||
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
|
||||
only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
|
||||
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
|
||||
Returns:
|
||||
Return a dict, whose key may be "greedy_search" if greedy search
|
||||
is used, or it may be "beam_7" if beam size of 7 is used.
|
||||
Its value is a list of tuples. Each tuple contains two elements:
|
||||
The first is the reference transcript, and the second is the
|
||||
predicted result.
|
||||
"""
|
||||
num_cuts = 0
|
||||
|
||||
try:
|
||||
num_batches = len(dl)
|
||||
except TypeError:
|
||||
num_batches = "?"
|
||||
|
||||
if params.decoding_method == "greedy_search":
|
||||
log_interval = 50
|
||||
else:
|
||||
log_interval = 20
|
||||
|
||||
results = defaultdict(list)
|
||||
for batch_idx, batch in enumerate(dl):
|
||||
texts = batch["supervisions"]["text"]
|
||||
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
|
||||
|
||||
hyps_dict = decode_one_batch(
|
||||
params=params,
|
||||
model=model,
|
||||
sp=sp,
|
||||
decoding_graph=decoding_graph,
|
||||
word_table=word_table,
|
||||
batch=batch,
|
||||
)
|
||||
|
||||
for name, hyps in hyps_dict.items():
|
||||
this_batch = []
|
||||
assert len(hyps) == len(texts)
|
||||
for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
|
||||
ref_words = ref_text.split()
|
||||
this_batch.append((cut_id, ref_words, hyp_words))
|
||||
|
||||
results[name].extend(this_batch)
|
||||
|
||||
num_cuts += len(texts)
|
||||
|
||||
if batch_idx % log_interval == 0:
|
||||
batch_str = f"{batch_idx}/{num_batches}"
|
||||
|
||||
logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
|
||||
return results
|
||||
|
||||
|
||||
def save_results(
|
||||
params: AttributeDict,
|
||||
test_set_name: str,
|
||||
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
|
||||
):
|
||||
test_set_wers = dict()
|
||||
for key, results in results_dict.items():
|
||||
recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
|
||||
results = sorted(results)
|
||||
store_transcripts(filename=recog_path, texts=results)
|
||||
logging.info(f"The transcripts are stored in {recog_path}")
|
||||
|
||||
# The following prints out WERs, per-word error statistics and aligned
|
||||
# ref/hyp pairs.
|
||||
errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
|
||||
with open(errs_filename, "w") as f:
|
||||
wer = write_error_stats(
|
||||
f, f"{test_set_name}-{key}", results, enable_log=True
|
||||
)
|
||||
test_set_wers[key] = wer
|
||||
|
||||
logging.info("Wrote detailed error stats to {}".format(errs_filename))
|
||||
|
||||
test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
|
||||
errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
|
||||
with open(errs_info, "w") as f:
|
||||
print("settings\tWER", file=f)
|
||||
for key, val in test_set_wers:
|
||||
print("{}\t{}".format(key, val), file=f)
|
||||
|
||||
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
|
||||
note = "\tbest for {}".format(test_set_name)
|
||||
for key, val in test_set_wers:
|
||||
s += "{}\t{}{}\n".format(key, val, note)
|
||||
note = ""
|
||||
logging.info(s)
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
parser = get_parser()
|
||||
AsrDataModule.add_arguments(parser)
|
||||
args = parser.parse_args()
|
||||
args.exp_dir = Path(args.exp_dir)
|
||||
|
||||
params = get_params()
|
||||
params.update(vars(args))
|
||||
|
||||
assert params.decoding_method in (
|
||||
"greedy_search",
|
||||
"beam_search",
|
||||
"fast_beam_search",
|
||||
"fast_beam_search_nbest",
|
||||
"fast_beam_search_nbest_LG",
|
||||
"fast_beam_search_nbest_oracle",
|
||||
"modified_beam_search",
|
||||
)
|
||||
params.res_dir = params.exp_dir / params.decoding_method
|
||||
|
||||
if params.iter > 0:
|
||||
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
|
||||
else:
|
||||
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
|
||||
|
||||
params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
|
||||
|
||||
if "fast_beam_search" in params.decoding_method:
|
||||
params.suffix += f"-beam-{params.beam}"
|
||||
params.suffix += f"-max-contexts-{params.max_contexts}"
|
||||
params.suffix += f"-max-states-{params.max_states}"
|
||||
params.suffix += f"-right-padding-{params.right_padding}"
|
||||
if "nbest" in params.decoding_method:
|
||||
params.suffix += f"-nbest-scale-{params.nbest_scale}"
|
||||
params.suffix += f"-num-paths-{params.num_paths}"
|
||||
if "LG" in params.decoding_method:
|
||||
params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
|
||||
elif "beam_search" in params.decoding_method:
|
||||
params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
|
||||
else:
|
||||
params.suffix += f"-context-{params.context_size}"
|
||||
params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
|
||||
|
||||
if params.use_averaged_model:
|
||||
params.suffix += "-use-averaged-model"
|
||||
|
||||
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
|
||||
logging.info("Decoding started")
|
||||
|
||||
device = torch.device("cpu")
|
||||
if torch.cuda.is_available():
|
||||
device = torch.device("cuda", 0)
|
||||
|
||||
logging.info(f"Device: {device}")
|
||||
|
||||
sp = spm.SentencePieceProcessor()
|
||||
sp.load(params.bpe_model)
|
||||
|
||||
# <blk> and <unk> are defined in local/train_bpe_model.py
|
||||
params.blank_id = sp.piece_to_id("<blk>")
|
||||
params.unk_id = sp.piece_to_id("<unk>")
|
||||
params.vocab_size = sp.get_piece_size()
|
||||
|
||||
logging.info(params)
|
||||
|
||||
logging.info("About to create model")
|
||||
model = get_transducer_model(params)
|
||||
assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
|
||||
model.encoder.decode_chunk_size,
|
||||
params.decode_chunk_len,
|
||||
)
|
||||
|
||||
if not params.use_averaged_model:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
elif params.avg == 1:
|
||||
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
|
||||
else:
|
||||
start = params.epoch - params.avg + 1
|
||||
filenames = []
|
||||
for i in range(start, params.epoch + 1):
|
||||
if i >= 1:
|
||||
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
else:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg + 1
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg + 1:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
filename_start = filenames[-1]
|
||||
filename_end = filenames[0]
|
||||
logging.info(
|
||||
"Calculating the averaged model over iteration checkpoints"
|
||||
f" from {filename_start} (excluded) to {filename_end}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
else:
|
||||
assert params.avg > 0, params.avg
|
||||
start = params.epoch - params.avg
|
||||
assert start >= 1, start
|
||||
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
|
||||
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
|
||||
logging.info(
|
||||
f"Calculating the averaged model over epoch range from "
|
||||
f"{start} (excluded) to {params.epoch}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
|
||||
model.to(device)
|
||||
model.eval()
|
||||
|
||||
if "fast_beam_search" in params.decoding_method:
|
||||
if params.decoding_method == "fast_beam_search_nbest_LG":
|
||||
lexicon = Lexicon(params.lang_dir)
|
||||
word_table = lexicon.word_table
|
||||
lg_filename = params.lang_dir / "LG.pt"
|
||||
logging.info(f"Loading {lg_filename}")
|
||||
decoding_graph = k2.Fsa.from_dict(
|
||||
torch.load(lg_filename, map_location=device)
|
||||
)
|
||||
decoding_graph.scores *= params.ngram_lm_scale
|
||||
else:
|
||||
word_table = None
|
||||
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
|
||||
else:
|
||||
decoding_graph = None
|
||||
word_table = None
|
||||
|
||||
num_param = sum([p.numel() for p in model.parameters()])
|
||||
logging.info(f"Number of model parameters: {num_param}")
|
||||
|
||||
# we need cut ids to display recognition results.
|
||||
args.return_cuts = True
|
||||
asr_datamodule = AsrDataModule(args)
|
||||
librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
|
||||
|
||||
test_clean_cuts = librispeech.test_clean_cuts()
|
||||
test_other_cuts = librispeech.test_other_cuts()
|
||||
|
||||
test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
|
||||
test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts)
|
||||
|
||||
test_sets = ["test-clean", "test-other"]
|
||||
test_dl = [test_clean_dl, test_other_dl]
|
||||
|
||||
for test_set, test_dl in zip(test_sets, test_dl):
|
||||
results_dict = decode_dataset(
|
||||
dl=test_dl,
|
||||
params=params,
|
||||
model=model,
|
||||
sp=sp,
|
||||
word_table=word_table,
|
||||
decoding_graph=decoding_graph,
|
||||
)
|
||||
|
||||
save_results(
|
||||
params=params,
|
||||
test_set_name=test_set,
|
||||
results_dict=results_dict,
|
||||
)
|
||||
|
||||
logging.info("Done!")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1,837 @@
|
||||
#!/usr/bin/env python3
|
||||
#
|
||||
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
|
||||
# Zengwei Yao,
|
||||
# Xiaoyu Yang)
|
||||
#
|
||||
# See ../../../../LICENSE for clarification regarding multiple authors
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""
|
||||
Usage:
|
||||
(1) greedy search
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--max-duration 600 \
|
||||
--decoding-method greedy_search
|
||||
|
||||
(2) beam search (not recommended)
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--max-duration 600 \
|
||||
--decoding-method beam_search \
|
||||
--beam-size 4
|
||||
|
||||
(3) modified beam search
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--max-duration 600 \
|
||||
--decoding-method modified_beam_search \
|
||||
--beam-size 4
|
||||
|
||||
(4) fast beam search (one best)
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--max-duration 600 \
|
||||
--decoding-method fast_beam_search \
|
||||
--beam 20.0 \
|
||||
--max-contexts 8 \
|
||||
--max-states 64
|
||||
|
||||
(5) fast beam search (nbest)
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--max-duration 600 \
|
||||
--decoding-method fast_beam_search_nbest \
|
||||
--beam 20.0 \
|
||||
--max-contexts 8 \
|
||||
--max-states 64 \
|
||||
--num-paths 200 \
|
||||
--nbest-scale 0.5
|
||||
|
||||
(6) fast beam search (nbest oracle WER)
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--max-duration 600 \
|
||||
--decoding-method fast_beam_search_nbest_oracle \
|
||||
--beam 20.0 \
|
||||
--max-contexts 8 \
|
||||
--max-states 64 \
|
||||
--num-paths 200 \
|
||||
--nbest-scale 0.5
|
||||
|
||||
(7) fast beam search (with LG)
|
||||
./pruned_transducer_stateless7/decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--exp-dir ./pruned_transducer_stateless7/exp \
|
||||
--max-duration 600 \
|
||||
--decoding-method fast_beam_search_nbest_LG \
|
||||
--beam 20.0 \
|
||||
--max-contexts 8 \
|
||||
--max-states 64
|
||||
"""
|
||||
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import math
|
||||
from collections import defaultdict
|
||||
from pathlib import Path
|
||||
from typing import Dict, List, Optional, Tuple
|
||||
|
||||
import k2
|
||||
import sentencepiece as spm
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from 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 asr_datamodule import LibriSpeechAsrDataModule
|
||||
from gigaspeech_asrmodule import GigaSpeechAsrDataModule
|
||||
from gigaspeech_scoring import asr_text_post_processing
|
||||
from train import add_model_arguments, get_params, get_transducer_model
|
||||
|
||||
from icefall.checkpoint import (
|
||||
average_checkpoints,
|
||||
average_checkpoints_with_averaged_model,
|
||||
find_checkpoints,
|
||||
load_checkpoint,
|
||||
)
|
||||
from icefall.lexicon import Lexicon
|
||||
from icefall.utils import (
|
||||
AttributeDict,
|
||||
setup_logger,
|
||||
store_transcripts,
|
||||
str2bool,
|
||||
write_error_stats,
|
||||
)
|
||||
|
||||
LOG_EPS = math.log(1e-10)
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--epoch",
|
||||
type=int,
|
||||
default=30,
|
||||
help="""It specifies the checkpoint to use for decoding.
|
||||
Note: Epoch counts from 1.
|
||||
You can specify --avg to use more checkpoints for model averaging.""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--iter",
|
||||
type=int,
|
||||
default=0,
|
||||
help="""If positive, --epoch is ignored and it
|
||||
will use the checkpoint exp_dir/checkpoint-iter.pt.
|
||||
You can specify --avg to use more checkpoints for model averaging.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--avg",
|
||||
type=int,
|
||||
default=9,
|
||||
help="Number of checkpoints to average. Automatically select "
|
||||
"consecutive checkpoints before the checkpoint specified by "
|
||||
"'--epoch' and '--iter'",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--use-averaged-model",
|
||||
type=str2bool,
|
||||
default=True,
|
||||
help="Whether to load averaged model. Currently it only supports "
|
||||
"using --epoch. If True, it would decode with the averaged model "
|
||||
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
|
||||
"Actually only the models with epoch number of `epoch-avg` and "
|
||||
"`epoch` are loaded for averaging. ",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--exp-dir",
|
||||
type=str,
|
||||
default="pruned_transducer_stateless7/exp",
|
||||
help="The experiment dir",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--bpe-model",
|
||||
type=str,
|
||||
default="data/lang_bpe_500/bpe.model",
|
||||
help="Path to the BPE model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--lang-dir",
|
||||
type=Path,
|
||||
default="data/lang_bpe_500",
|
||||
help="The lang dir containing word table and LG graph",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--decoding-method",
|
||||
type=str,
|
||||
default="greedy_search",
|
||||
help="""Possible values are:
|
||||
- greedy_search
|
||||
- beam_search
|
||||
- modified_beam_search
|
||||
- fast_beam_search
|
||||
- fast_beam_search_nbest
|
||||
- fast_beam_search_nbest_oracle
|
||||
- fast_beam_search_nbest_LG
|
||||
If you use fast_beam_search_nbest_LG, you have to specify
|
||||
`--lang-dir`, which should contain `LG.pt`.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--beam-size",
|
||||
type=int,
|
||||
default=4,
|
||||
help="""An integer indicating how many candidates we will keep for each
|
||||
frame. Used only when --decoding-method is beam_search or
|
||||
modified_beam_search.""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--beam",
|
||||
type=float,
|
||||
default=20.0,
|
||||
help="""A floating point value to calculate the cutoff score during beam
|
||||
search (i.e., `cutoff = max-score - beam`), which is the same as the
|
||||
`beam` in Kaldi.
|
||||
Used only when --decoding-method is fast_beam_search,
|
||||
fast_beam_search_nbest, fast_beam_search_nbest_LG,
|
||||
and fast_beam_search_nbest_oracle
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--ngram-lm-scale",
|
||||
type=float,
|
||||
default=0.01,
|
||||
help="""
|
||||
Used only when --decoding_method is fast_beam_search_nbest_LG.
|
||||
It specifies the scale for n-gram LM scores.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--max-contexts",
|
||||
type=int,
|
||||
default=8,
|
||||
help="""Used only when --decoding-method is
|
||||
fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
|
||||
and fast_beam_search_nbest_oracle""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--max-states",
|
||||
type=int,
|
||||
default=64,
|
||||
help="""Used only when --decoding-method is
|
||||
fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
|
||||
and fast_beam_search_nbest_oracle""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--context-size",
|
||||
type=int,
|
||||
default=2,
|
||||
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--max-sym-per-frame",
|
||||
type=int,
|
||||
default=1,
|
||||
help="""Maximum number of symbols per frame.
|
||||
Used only when --decoding_method is greedy_search""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--num-paths",
|
||||
type=int,
|
||||
default=200,
|
||||
help="""Number of paths for nbest decoding.
|
||||
Used only when the decoding method is fast_beam_search_nbest,
|
||||
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--nbest-scale",
|
||||
type=float,
|
||||
default=0.5,
|
||||
help="""Scale applied to lattice scores when computing nbest paths.
|
||||
Used only when the decoding method is fast_beam_search_nbest,
|
||||
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--decode-chunk-size",
|
||||
type=int,
|
||||
default=16,
|
||||
help="The chunk size for decoding (in frames after subsampling)",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--right-padding",
|
||||
type=int,
|
||||
default=64,
|
||||
help="Padding frames at the end of features",
|
||||
)
|
||||
|
||||
add_model_arguments(parser)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
def post_processing(
|
||||
results: List[Tuple[str, List[str], List[str]]],
|
||||
) -> List[Tuple[str, List[str], List[str]]]:
|
||||
new_results = []
|
||||
for key, ref, hyp in results:
|
||||
new_ref = asr_text_post_processing(" ".join(ref)).split()
|
||||
new_hyp = asr_text_post_processing(" ".join(hyp)).split()
|
||||
new_results.append((key, new_ref, new_hyp))
|
||||
return new_results
|
||||
|
||||
|
||||
def decode_one_batch(
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
sp: spm.SentencePieceProcessor,
|
||||
batch: dict,
|
||||
word_table: Optional[k2.SymbolTable] = None,
|
||||
decoding_graph: Optional[k2.Fsa] = None,
|
||||
) -> Dict[str, List[List[str]]]:
|
||||
"""Decode one batch and return the result in a dict. The dict has the
|
||||
following format:
|
||||
|
||||
- key: It indicates the setting used for decoding. For example,
|
||||
if greedy_search is used, it would be "greedy_search"
|
||||
If beam search with a beam size of 7 is used, it would be
|
||||
"beam_7"
|
||||
- value: It contains the decoding result. `len(value)` equals to
|
||||
batch size. `value[i]` is the decoding result for the i-th
|
||||
utterance in the given batch.
|
||||
Args:
|
||||
params:
|
||||
It's the return value of :func:`get_params`.
|
||||
model:
|
||||
The neural model.
|
||||
sp:
|
||||
The BPE model.
|
||||
batch:
|
||||
It is the return value from iterating
|
||||
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
|
||||
for the format of the `batch`.
|
||||
word_table:
|
||||
The word symbol table.
|
||||
decoding_graph:
|
||||
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
|
||||
only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
|
||||
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
|
||||
Returns:
|
||||
Return the decoding result. See above description for the format of
|
||||
the returned dict.
|
||||
"""
|
||||
device = next(model.parameters()).device
|
||||
feature = batch["inputs"]
|
||||
assert feature.ndim == 3
|
||||
|
||||
feature = feature.to(device)
|
||||
# at entry, feature is (N, T, C)
|
||||
|
||||
supervisions = batch["supervisions"]
|
||||
feature_lens = supervisions["num_frames"].to(device)
|
||||
|
||||
feature_lens += params.right_padding
|
||||
feature = torch.nn.functional.pad(
|
||||
feature,
|
||||
pad=(0, 0, 0, params.right_padding),
|
||||
value=LOG_EPS,
|
||||
)
|
||||
encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
|
||||
|
||||
hyps = []
|
||||
|
||||
if params.decoding_method == "fast_beam_search":
|
||||
hyp_tokens = fast_beam_search_one_best(
|
||||
model=model,
|
||||
decoding_graph=decoding_graph,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
elif params.decoding_method == "fast_beam_search_nbest_LG":
|
||||
hyp_tokens = fast_beam_search_nbest_LG(
|
||||
model=model,
|
||||
decoding_graph=decoding_graph,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
num_paths=params.num_paths,
|
||||
nbest_scale=params.nbest_scale,
|
||||
)
|
||||
for hyp in hyp_tokens:
|
||||
hyps.append([word_table[i] for i in hyp])
|
||||
elif params.decoding_method == "fast_beam_search_nbest":
|
||||
hyp_tokens = fast_beam_search_nbest(
|
||||
model=model,
|
||||
decoding_graph=decoding_graph,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
num_paths=params.num_paths,
|
||||
nbest_scale=params.nbest_scale,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
elif params.decoding_method == "fast_beam_search_nbest_oracle":
|
||||
hyp_tokens = fast_beam_search_nbest_oracle(
|
||||
model=model,
|
||||
decoding_graph=decoding_graph,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam,
|
||||
max_contexts=params.max_contexts,
|
||||
max_states=params.max_states,
|
||||
num_paths=params.num_paths,
|
||||
ref_texts=sp.encode(supervisions["text"]),
|
||||
nbest_scale=params.nbest_scale,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
|
||||
hyp_tokens = greedy_search_batch(
|
||||
model=model,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
elif params.decoding_method == "modified_beam_search":
|
||||
hyp_tokens = modified_beam_search(
|
||||
model=model,
|
||||
encoder_out=encoder_out,
|
||||
encoder_out_lens=encoder_out_lens,
|
||||
beam=params.beam_size,
|
||||
)
|
||||
for hyp in sp.decode(hyp_tokens):
|
||||
hyps.append(hyp.split())
|
||||
else:
|
||||
batch_size = encoder_out.size(0)
|
||||
|
||||
for i in range(batch_size):
|
||||
# fmt: off
|
||||
encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
|
||||
# fmt: on
|
||||
if params.decoding_method == "greedy_search":
|
||||
hyp = greedy_search(
|
||||
model=model,
|
||||
encoder_out=encoder_out_i,
|
||||
max_sym_per_frame=params.max_sym_per_frame,
|
||||
)
|
||||
elif params.decoding_method == "beam_search":
|
||||
hyp = beam_search(
|
||||
model=model,
|
||||
encoder_out=encoder_out_i,
|
||||
beam=params.beam_size,
|
||||
)
|
||||
else:
|
||||
raise ValueError(
|
||||
f"Unsupported decoding method: {params.decoding_method}"
|
||||
)
|
||||
hyps.append(sp.decode(hyp).split())
|
||||
|
||||
if params.decoding_method == "greedy_search":
|
||||
return {"greedy_search": hyps}
|
||||
elif "fast_beam_search" in params.decoding_method:
|
||||
key = f"beam_{params.beam}_"
|
||||
key += f"max_contexts_{params.max_contexts}_"
|
||||
key += f"max_states_{params.max_states}"
|
||||
if "nbest" in params.decoding_method:
|
||||
key += f"_num_paths_{params.num_paths}_"
|
||||
key += f"nbest_scale_{params.nbest_scale}"
|
||||
if "LG" in params.decoding_method:
|
||||
key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
|
||||
|
||||
return {key: hyps}
|
||||
else:
|
||||
return {f"beam_size_{params.beam_size}": hyps}
|
||||
|
||||
|
||||
def decode_dataset(
|
||||
dl: torch.utils.data.DataLoader,
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
sp: spm.SentencePieceProcessor,
|
||||
word_table: Optional[k2.SymbolTable] = None,
|
||||
decoding_graph: Optional[k2.Fsa] = None,
|
||||
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
|
||||
"""Decode dataset.
|
||||
|
||||
Args:
|
||||
dl:
|
||||
PyTorch's dataloader containing the dataset to decode.
|
||||
params:
|
||||
It is returned by :func:`get_params`.
|
||||
model:
|
||||
The neural model.
|
||||
sp:
|
||||
The BPE model.
|
||||
word_table:
|
||||
The word symbol table.
|
||||
decoding_graph:
|
||||
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
|
||||
only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
|
||||
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
|
||||
Returns:
|
||||
Return a dict, whose key may be "greedy_search" if greedy search
|
||||
is used, or it may be "beam_7" if beam size of 7 is used.
|
||||
Its value is a list of tuples. Each tuple contains two elements:
|
||||
The first is the reference transcript, and the second is the
|
||||
predicted result.
|
||||
"""
|
||||
num_cuts = 0
|
||||
|
||||
try:
|
||||
num_batches = len(dl)
|
||||
except TypeError:
|
||||
num_batches = "?"
|
||||
|
||||
if params.decoding_method == "greedy_search":
|
||||
log_interval = 50
|
||||
else:
|
||||
log_interval = 20
|
||||
|
||||
results = defaultdict(list)
|
||||
for batch_idx, batch in enumerate(dl):
|
||||
texts = batch["supervisions"]["text"]
|
||||
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
|
||||
|
||||
hyps_dict = decode_one_batch(
|
||||
params=params,
|
||||
model=model,
|
||||
sp=sp,
|
||||
decoding_graph=decoding_graph,
|
||||
word_table=word_table,
|
||||
batch=batch,
|
||||
)
|
||||
|
||||
for name, hyps in hyps_dict.items():
|
||||
this_batch = []
|
||||
assert len(hyps) == len(texts)
|
||||
for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
|
||||
ref_words = ref_text.split()
|
||||
this_batch.append((cut_id, ref_words, hyp_words))
|
||||
|
||||
results[name].extend(this_batch)
|
||||
|
||||
num_cuts += len(texts)
|
||||
|
||||
if batch_idx % log_interval == 0:
|
||||
batch_str = f"{batch_idx}/{num_batches}"
|
||||
|
||||
logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
|
||||
return results
|
||||
|
||||
|
||||
def save_results(
|
||||
params: AttributeDict,
|
||||
test_set_name: str,
|
||||
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
|
||||
):
|
||||
test_set_wers = dict()
|
||||
for key, results in results_dict.items():
|
||||
recog_path = (
|
||||
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
|
||||
)
|
||||
results = post_processing(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():
|
||||
"""
|
||||
This scripts test a libri model with libri BPE
|
||||
on Gigaspeech.
|
||||
"""
|
||||
parser = get_parser()
|
||||
GigaSpeechAsrDataModule.add_arguments(parser)
|
||||
args = parser.parse_args()
|
||||
args.exp_dir = Path(args.exp_dir)
|
||||
|
||||
params = get_params()
|
||||
params.update(vars(args))
|
||||
|
||||
assert params.decoding_method in (
|
||||
"greedy_search",
|
||||
"beam_search",
|
||||
"fast_beam_search",
|
||||
"fast_beam_search_nbest",
|
||||
"fast_beam_search_nbest_LG",
|
||||
"fast_beam_search_nbest_oracle",
|
||||
"modified_beam_search",
|
||||
)
|
||||
params.res_dir = params.exp_dir / (params.decoding_method + "_gigaspeech")
|
||||
|
||||
if params.iter > 0:
|
||||
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
|
||||
else:
|
||||
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
|
||||
|
||||
params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
|
||||
|
||||
if "fast_beam_search" in params.decoding_method:
|
||||
params.suffix += f"-beam-{params.beam}"
|
||||
params.suffix += f"-max-contexts-{params.max_contexts}"
|
||||
params.suffix += f"-max-states-{params.max_states}"
|
||||
params.suffix += f"-right-padding-{params.right_padding}"
|
||||
if "nbest" in params.decoding_method:
|
||||
params.suffix += f"-nbest-scale-{params.nbest_scale}"
|
||||
params.suffix += f"-num-paths-{params.num_paths}"
|
||||
if "LG" in params.decoding_method:
|
||||
params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
|
||||
elif "beam_search" in params.decoding_method:
|
||||
params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
|
||||
else:
|
||||
params.suffix += f"-context-{params.context_size}"
|
||||
params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
|
||||
|
||||
if params.use_averaged_model:
|
||||
params.suffix += "-use-averaged-model"
|
||||
|
||||
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
|
||||
logging.info("Decoding started")
|
||||
|
||||
device = torch.device("cpu")
|
||||
if torch.cuda.is_available():
|
||||
device = torch.device("cuda", 0)
|
||||
|
||||
logging.info(f"Device: {device}")
|
||||
|
||||
sp = spm.SentencePieceProcessor()
|
||||
sp.load(params.bpe_model)
|
||||
|
||||
# <blk> and <unk> are defined in local/train_bpe_model.py
|
||||
params.blank_id = sp.piece_to_id("<blk>")
|
||||
params.unk_id = sp.piece_to_id("<unk>")
|
||||
params.vocab_size = sp.get_piece_size()
|
||||
|
||||
logging.info(params)
|
||||
|
||||
logging.info("About to create model")
|
||||
model = get_transducer_model(params)
|
||||
|
||||
if not params.use_averaged_model:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
elif params.avg == 1:
|
||||
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
|
||||
else:
|
||||
start = params.epoch - params.avg + 1
|
||||
filenames = []
|
||||
for i in range(start, params.epoch + 1):
|
||||
if i >= 1:
|
||||
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
else:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg + 1
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg + 1:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
filename_start = filenames[-1]
|
||||
filename_end = filenames[0]
|
||||
logging.info(
|
||||
"Calculating the averaged model over iteration checkpoints"
|
||||
f" from {filename_start} (excluded) to {filename_end}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
else:
|
||||
assert params.avg > 0, params.avg
|
||||
start = params.epoch - params.avg
|
||||
assert start >= 1, start
|
||||
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
|
||||
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
|
||||
logging.info(
|
||||
f"Calculating the averaged model over epoch range from "
|
||||
f"{start} (excluded) to {params.epoch}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
|
||||
model.to(device)
|
||||
model.eval()
|
||||
|
||||
if "fast_beam_search" in params.decoding_method:
|
||||
if params.decoding_method == "fast_beam_search_nbest_LG":
|
||||
lexicon = Lexicon(params.lang_dir)
|
||||
word_table = lexicon.word_table
|
||||
lg_filename = params.lang_dir / "LG.pt"
|
||||
logging.info(f"Loading {lg_filename}")
|
||||
decoding_graph = k2.Fsa.from_dict(
|
||||
torch.load(lg_filename, map_location=device)
|
||||
)
|
||||
decoding_graph.scores *= params.ngram_lm_scale
|
||||
else:
|
||||
word_table = None
|
||||
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
|
||||
else:
|
||||
decoding_graph = None
|
||||
word_table = None
|
||||
|
||||
num_param = sum([p.numel() for p in model.parameters()])
|
||||
logging.info(f"Number of model parameters: {num_param}")
|
||||
|
||||
# we need cut ids to display recognition results.
|
||||
args.return_cuts = True
|
||||
gigaspeech = GigaSpeechAsrDataModule(args)
|
||||
|
||||
dev_cuts = gigaspeech.dev_cuts()
|
||||
test_cuts = gigaspeech.test_cuts()
|
||||
|
||||
dev_dl = gigaspeech.test_dataloaders(dev_cuts)
|
||||
test_dl = gigaspeech.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,
|
||||
sp=sp,
|
||||
word_table=word_table,
|
||||
decoding_graph=decoding_graph,
|
||||
)
|
||||
|
||||
save_results(
|
||||
params=params,
|
||||
test_set_name=test_set,
|
||||
results_dict=results_dict,
|
||||
)
|
||||
|
||||
logging.info("Done!")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/decode_stream.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/decoder.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/encoder_interface.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
|
||||
@ -0,0 +1,368 @@
|
||||
#!/usr/bin/env python3
|
||||
|
||||
"""
|
||||
Please see
|
||||
https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
|
||||
for more details about how to use this file.
|
||||
|
||||
We use
|
||||
https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
|
||||
to demonstrate the usage of this file.
|
||||
|
||||
1. Download the pre-trained model
|
||||
|
||||
cd egs/librispeech/ASR
|
||||
|
||||
repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
|
||||
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
|
||||
repo=$(basename $repo_url)
|
||||
|
||||
pushd $repo
|
||||
git lfs pull --include "data/lang_bpe/bpe.model"
|
||||
git lfs pull --include "exp/pretrained.pt"
|
||||
|
||||
cd exp
|
||||
ln -s pretrained.pt epoch-99.pt
|
||||
popd
|
||||
|
||||
2. Export to ncnn
|
||||
|
||||
./pruned_transducer_stateless7_streaming/export-for-ncnn.py \
|
||||
--bpe-model $repo/data/lang_bpe_500/bpe.model \
|
||||
--exp-dir $repo/exp \
|
||||
--use-averaged-model 0 \
|
||||
--epoch 99 \
|
||||
--avg 1 \
|
||||
--decode-chunk-len 32 \
|
||||
--num-encoder-layers "2,4,3,2,4" \
|
||||
--feedforward-dims "1024,1024,2048,2048,1024" \
|
||||
--nhead "8,8,8,8,8" \
|
||||
--encoder-dims "384,384,384,384,384" \
|
||||
--attention-dims "192,192,192,192,192" \
|
||||
--encoder-unmasked-dims "256,256,256,256,256" \
|
||||
--zipformer-downsampling-factors "1,2,4,8,2" \
|
||||
--cnn-module-kernels "31,31,31,31,31" \
|
||||
--decoder-dim 512 \
|
||||
--joiner-dim 512
|
||||
|
||||
cd $repo/exp
|
||||
|
||||
pnnx encoder_jit_trace-pnnx.pt
|
||||
pnnx decoder_jit_trace-pnnx.pt
|
||||
pnnx joiner_jit_trace-pnnx.pt
|
||||
|
||||
You can find converted models at
|
||||
https://huggingface.co/csukuangfj/sherpa-ncnn-streaming-zipformer-en-2023-02-13
|
||||
|
||||
See ./streaming-ncnn-decode.py
|
||||
and
|
||||
https://github.com/k2-fsa/sherpa-ncnn
|
||||
for usage.
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
from pathlib import Path
|
||||
|
||||
import sentencepiece as spm
|
||||
import torch
|
||||
from scaling_converter import convert_scaled_to_non_scaled
|
||||
from train2 import add_model_arguments, get_params, get_transducer_model
|
||||
|
||||
from icefall.checkpoint import (
|
||||
average_checkpoints,
|
||||
average_checkpoints_with_averaged_model,
|
||||
find_checkpoints,
|
||||
load_checkpoint,
|
||||
)
|
||||
from icefall.utils import setup_logger, str2bool
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--epoch",
|
||||
type=int,
|
||||
default=28,
|
||||
help="""It specifies the checkpoint to use for averaging.
|
||||
Note: Epoch counts from 0.
|
||||
You can specify --avg to use more checkpoints for model averaging.""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--iter",
|
||||
type=int,
|
||||
default=0,
|
||||
help="""If positive, --epoch is ignored and it
|
||||
will use the checkpoint exp_dir/checkpoint-iter.pt.
|
||||
You can specify --avg to use more checkpoints for model averaging.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--avg",
|
||||
type=int,
|
||||
default=15,
|
||||
help="Number of checkpoints to average. Automatically select "
|
||||
"consecutive checkpoints before the checkpoint specified by "
|
||||
"'--epoch' and '--iter'",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--exp-dir",
|
||||
type=str,
|
||||
default="pruned_transducer_stateless7_streaming/exp",
|
||||
help="""It specifies the directory where all training related
|
||||
files, e.g., checkpoints, log, etc, are saved
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--bpe-model",
|
||||
type=str,
|
||||
default="data/lang_bpe_500/bpe.model",
|
||||
help="Path to the BPE model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--context-size",
|
||||
type=int,
|
||||
default=2,
|
||||
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
|
||||
)
|
||||
|
||||
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. ",
|
||||
)
|
||||
|
||||
add_model_arguments(parser)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
def export_encoder_model_jit_trace(
|
||||
encoder_model: torch.nn.Module,
|
||||
encoder_filename: str,
|
||||
) -> None:
|
||||
"""Export the given encoder model with torch.jit.trace()
|
||||
|
||||
Note: The warmup argument is fixed to 1.
|
||||
|
||||
Args:
|
||||
encoder_model:
|
||||
The input encoder model
|
||||
encoder_filename:
|
||||
The filename to save the exported model.
|
||||
"""
|
||||
encoder_model.__class__.forward = encoder_model.__class__.streaming_forward
|
||||
|
||||
decode_chunk_len = encoder_model.decode_chunk_size * 2
|
||||
pad_length = 7
|
||||
T = decode_chunk_len + pad_length # 32 + 7 = 39
|
||||
|
||||
logging.info(f"decode_chunk_len: {decode_chunk_len}")
|
||||
logging.info(f"T: {T}")
|
||||
|
||||
x = torch.zeros(1, T, 80, dtype=torch.float32)
|
||||
states = encoder_model.get_init_state()
|
||||
|
||||
traced_model = torch.jit.trace(encoder_model, (x, states))
|
||||
traced_model.save(encoder_filename)
|
||||
logging.info(f"Saved to {encoder_filename}")
|
||||
|
||||
|
||||
def export_decoder_model_jit_trace(
|
||||
decoder_model: torch.nn.Module,
|
||||
decoder_filename: str,
|
||||
) -> None:
|
||||
"""Export the given decoder model with torch.jit.trace()
|
||||
|
||||
Note: The argument need_pad is fixed to False.
|
||||
|
||||
Args:
|
||||
decoder_model:
|
||||
The input decoder model
|
||||
decoder_filename:
|
||||
The filename to save the exported model.
|
||||
"""
|
||||
y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
|
||||
need_pad = torch.tensor([False])
|
||||
|
||||
traced_model = torch.jit.trace(decoder_model, (y, need_pad))
|
||||
traced_model.save(decoder_filename)
|
||||
logging.info(f"Saved to {decoder_filename}")
|
||||
|
||||
|
||||
def export_joiner_model_jit_trace(
|
||||
joiner_model: torch.nn.Module,
|
||||
joiner_filename: str,
|
||||
) -> None:
|
||||
"""Export the given joiner model with torch.jit.trace()
|
||||
|
||||
Note: The argument project_input is fixed to True. A user should not
|
||||
project the encoder_out/decoder_out by himself/herself. The exported joiner
|
||||
will do that for the user.
|
||||
|
||||
Args:
|
||||
joiner_model:
|
||||
The input joiner model
|
||||
joiner_filename:
|
||||
The filename to save the exported model.
|
||||
|
||||
"""
|
||||
encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
|
||||
decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
|
||||
encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
|
||||
decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
|
||||
|
||||
traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
|
||||
traced_model.save(joiner_filename)
|
||||
logging.info(f"Saved to {joiner_filename}")
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
args = get_parser().parse_args()
|
||||
args.exp_dir = Path(args.exp_dir)
|
||||
|
||||
params = get_params()
|
||||
params.update(vars(args))
|
||||
|
||||
device = torch.device("cpu")
|
||||
|
||||
setup_logger(f"{params.exp_dir}/log-export/log-export-ncnn")
|
||||
|
||||
logging.info(f"device: {device}")
|
||||
|
||||
sp = spm.SentencePieceProcessor()
|
||||
sp.load(params.bpe_model)
|
||||
|
||||
# <blk> is defined in local/train_bpe_model.py
|
||||
params.blank_id = sp.piece_to_id("<blk>")
|
||||
params.vocab_size = sp.get_piece_size()
|
||||
|
||||
logging.info(params)
|
||||
|
||||
logging.info("About to create model")
|
||||
model = get_transducer_model(params)
|
||||
|
||||
if not params.use_averaged_model:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
elif params.avg == 1:
|
||||
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
|
||||
else:
|
||||
start = params.epoch - params.avg + 1
|
||||
filenames = []
|
||||
for i in range(start, params.epoch + 1):
|
||||
if i >= 1:
|
||||
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
else:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg + 1
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg + 1:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
filename_start = filenames[-1]
|
||||
filename_end = filenames[0]
|
||||
logging.info(
|
||||
"Calculating the averaged model over iteration checkpoints"
|
||||
f" from {filename_start} (excluded) to {filename_end}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
else:
|
||||
assert params.avg > 0, params.avg
|
||||
start = params.epoch - params.avg
|
||||
assert start >= 1, start
|
||||
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
|
||||
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
|
||||
logging.info(
|
||||
f"Calculating the averaged model over epoch range from "
|
||||
f"{start} (excluded) to {params.epoch}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
|
||||
model.to("cpu")
|
||||
model.eval()
|
||||
|
||||
convert_scaled_to_non_scaled(model, inplace=True, is_pnnx=True)
|
||||
|
||||
encoder_num_param = sum([p.numel() for p in model.encoder.parameters()])
|
||||
decoder_num_param = sum([p.numel() for p in model.decoder.parameters()])
|
||||
joiner_num_param = sum([p.numel() for p in model.joiner.parameters()])
|
||||
total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
|
||||
logging.info(f"encoder parameters: {encoder_num_param}")
|
||||
logging.info(f"decoder parameters: {decoder_num_param}")
|
||||
logging.info(f"joiner parameters: {joiner_num_param}")
|
||||
logging.info(f"total parameters: {total_num_param}")
|
||||
|
||||
logging.info("Using torch.jit.trace()")
|
||||
|
||||
logging.info("Exporting encoder")
|
||||
encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"
|
||||
export_encoder_model_jit_trace(model.encoder, encoder_filename)
|
||||
|
||||
logging.info("Exporting decoder")
|
||||
decoder_filename = params.exp_dir / "decoder_jit_trace-pnnx.pt"
|
||||
export_decoder_model_jit_trace(model.decoder, decoder_filename)
|
||||
|
||||
logging.info("Exporting joiner")
|
||||
joiner_filename = params.exp_dir / "joiner_jit_trace-pnnx.pt"
|
||||
export_joiner_model_jit_trace(model.joiner, joiner_filename)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
|
||||
|
||||
main()
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/export-onnx.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/export.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless8/gigaspeech.py
|
||||
@ -0,0 +1 @@
|
||||
../../../gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
|
||||
@ -0,0 +1 @@
|
||||
../../../gigaspeech/ASR/pruned_transducer_stateless2/gigaspeech_scoring.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/jit_pretrained.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/jit_trace_export.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/joiner.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless8/librispeech.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/model.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/onnx_check.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/onnx_pretrained.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/optim.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/pretrained.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/scaling.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/scaling_converter.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/streaming_beam_search.py
|
||||
@ -0,0 +1,610 @@
|
||||
#!/usr/bin/env python3
|
||||
# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
|
||||
#
|
||||
# See ../../../../LICENSE for clarification regarding multiple authors
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
"""
|
||||
Usage:
|
||||
./pruned_transducer_stateless7_streaming/streaming_decode.py \
|
||||
--epoch 28 \
|
||||
--avg 15 \
|
||||
--decode-chunk-len 32 \
|
||||
--exp-dir ./pruned_transducer_stateless7_streaming/exp \
|
||||
--decoding_method greedy_search \
|
||||
--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 sentencepiece as spm
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from asr_datamodule import AsrDataModule
|
||||
from decode_stream import DecodeStream
|
||||
from kaldifeat import Fbank, FbankOptions
|
||||
from lhotse import CutSet
|
||||
from librispeech import LibriSpeech
|
||||
from streaming_beam_search import (
|
||||
fast_beam_search_one_best,
|
||||
greedy_search,
|
||||
modified_beam_search,
|
||||
)
|
||||
from torch.nn.utils.rnn import pad_sequence
|
||||
from train import add_model_arguments, get_params, get_transducer_model
|
||||
from zipformer import stack_states, unstack_states
|
||||
|
||||
from icefall.checkpoint import (
|
||||
average_checkpoints,
|
||||
average_checkpoints_with_averaged_model,
|
||||
find_checkpoints,
|
||||
load_checkpoint,
|
||||
)
|
||||
from icefall.utils import (
|
||||
AttributeDict,
|
||||
setup_logger,
|
||||
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_stateless2/exp",
|
||||
help="The experiment dir",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--bpe-model",
|
||||
type=str,
|
||||
default="data/lang_bpe_500/bpe.model",
|
||||
help="Path to the BPE model",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--decoding-method",
|
||||
type=str,
|
||||
default="greedy_search",
|
||||
help="""Supported decoding methods are:
|
||||
greedy_search
|
||||
modified_beam_search
|
||||
fast_beam_search
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--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(
|
||||
"--num-decode-streams",
|
||||
type=int,
|
||||
default=2000,
|
||||
help="The number of streams that can be decoded parallel.",
|
||||
)
|
||||
|
||||
add_model_arguments(parser)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
def decode_one_chunk(
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
decode_streams: List[DecodeStream],
|
||||
) -> List[int]:
|
||||
"""Decode one chunk frames of features for each decode_streams and
|
||||
return the indexes of finished streams in a List.
|
||||
|
||||
Args:
|
||||
params:
|
||||
It's the return value of :func:`get_params`.
|
||||
model:
|
||||
The neural model.
|
||||
decode_streams:
|
||||
A List of DecodeStream, each belonging to a utterance.
|
||||
Returns:
|
||||
Return a List containing which DecodeStreams are finished.
|
||||
"""
|
||||
device = model.device
|
||||
|
||||
features = []
|
||||
feature_lens = []
|
||||
states = []
|
||||
processed_lens = []
|
||||
|
||||
for stream in decode_streams:
|
||||
feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
|
||||
features.append(feat)
|
||||
feature_lens.append(feat_len)
|
||||
states.append(stream.states)
|
||||
processed_lens.append(stream.done_frames)
|
||||
|
||||
feature_lens = torch.tensor(feature_lens, device=device)
|
||||
features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
|
||||
|
||||
# We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
|
||||
# factor in encoders is 8.
|
||||
# After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
|
||||
tail_length = 23
|
||||
if features.size(1) < tail_length:
|
||||
pad_length = tail_length - features.size(1)
|
||||
feature_lens += pad_length
|
||||
features = torch.nn.functional.pad(
|
||||
features,
|
||||
(0, 0, 0, pad_length),
|
||||
mode="constant",
|
||||
value=LOG_EPS,
|
||||
)
|
||||
|
||||
states = stack_states(states)
|
||||
processed_lens = torch.tensor(processed_lens, device=device)
|
||||
|
||||
encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
|
||||
x=features,
|
||||
x_lens=feature_lens,
|
||||
states=states,
|
||||
)
|
||||
|
||||
encoder_out = model.joiner.encoder_proj(encoder_out)
|
||||
|
||||
if params.decoding_method == "greedy_search":
|
||||
greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
|
||||
elif params.decoding_method == "fast_beam_search":
|
||||
processed_lens = processed_lens + encoder_out_lens
|
||||
fast_beam_search_one_best(
|
||||
model=model,
|
||||
encoder_out=encoder_out,
|
||||
processed_lens=processed_lens,
|
||||
streams=decode_streams,
|
||||
beam=params.beam,
|
||||
max_states=params.max_states,
|
||||
max_contexts=params.max_contexts,
|
||||
)
|
||||
elif params.decoding_method == "modified_beam_search":
|
||||
modified_beam_search(
|
||||
model=model,
|
||||
streams=decode_streams,
|
||||
encoder_out=encoder_out,
|
||||
num_active_paths=params.num_active_paths,
|
||||
)
|
||||
else:
|
||||
raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
|
||||
|
||||
states = unstack_states(new_states)
|
||||
|
||||
finished_streams = []
|
||||
for i in range(len(decode_streams)):
|
||||
decode_streams[i].states = states[i]
|
||||
decode_streams[i].done_frames += encoder_out_lens[i]
|
||||
if decode_streams[i].done:
|
||||
finished_streams.append(i)
|
||||
|
||||
return finished_streams
|
||||
|
||||
|
||||
def decode_dataset(
|
||||
cuts: CutSet,
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
sp: spm.SentencePieceProcessor,
|
||||
decoding_graph: Optional[k2.Fsa] = None,
|
||||
) -> Dict[str, List[Tuple[List[str], List[str]]]]:
|
||||
"""Decode dataset.
|
||||
|
||||
Args:
|
||||
cuts:
|
||||
Lhotse Cutset containing the dataset to decode.
|
||||
params:
|
||||
It is returned by :func:`get_params`.
|
||||
model:
|
||||
The neural model.
|
||||
sp:
|
||||
The BPE model.
|
||||
decoding_graph:
|
||||
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
|
||||
only when --decoding_method is fast_beam_search.
|
||||
Returns:
|
||||
Return a dict, whose key may be "greedy_search" if greedy search
|
||||
is used, or it may be "beam_7" if beam size of 7 is used.
|
||||
Its value is a list of tuples. Each tuple contains two elements:
|
||||
The first is the reference transcript, and the second is the
|
||||
predicted result.
|
||||
"""
|
||||
device = model.device
|
||||
|
||||
opts = FbankOptions()
|
||||
opts.device = device
|
||||
opts.frame_opts.dither = 0
|
||||
opts.frame_opts.snip_edges = False
|
||||
opts.frame_opts.samp_freq = 16000
|
||||
opts.mel_opts.num_bins = 80
|
||||
|
||||
log_interval = 50
|
||||
|
||||
decode_results = []
|
||||
# Contain decode streams currently running.
|
||||
decode_streams = []
|
||||
for num, cut in enumerate(cuts):
|
||||
# each utterance has a DecodeStream.
|
||||
initial_states = model.encoder.get_init_state(device=device)
|
||||
decode_stream = DecodeStream(
|
||||
params=params,
|
||||
cut_id=cut.id,
|
||||
initial_states=initial_states,
|
||||
decoding_graph=decoding_graph,
|
||||
device=device,
|
||||
)
|
||||
|
||||
audio: np.ndarray = cut.load_audio()
|
||||
# audio.shape: (1, num_samples)
|
||||
assert len(audio.shape) == 2
|
||||
assert audio.shape[0] == 1, "Should be single channel"
|
||||
assert audio.dtype == np.float32, audio.dtype
|
||||
|
||||
# The trained model is using normalized samples
|
||||
assert audio.max() <= 1, "Should be normalized to [-1, 1])"
|
||||
|
||||
samples = torch.from_numpy(audio).squeeze(0)
|
||||
|
||||
fbank = Fbank(opts)
|
||||
feature = fbank(samples.to(device))
|
||||
decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
|
||||
decode_stream.ground_truth = cut.supervisions[0].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,
|
||||
decode_streams[i].ground_truth.split(),
|
||||
sp.decode(decode_streams[i].decoding_result()).split(),
|
||||
)
|
||||
)
|
||||
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(),
|
||||
sp.decode(decode_streams[i].decoding_result()).split(),
|
||||
)
|
||||
)
|
||||
del decode_streams[i]
|
||||
|
||||
if params.decoding_method == "greedy_search":
|
||||
key = "greedy_search"
|
||||
elif params.decoding_method == "fast_beam_search":
|
||||
key = (
|
||||
f"beam_{params.beam}_"
|
||||
f"max_contexts_{params.max_contexts}_"
|
||||
f"max_states_{params.max_states}"
|
||||
)
|
||||
elif params.decoding_method == "modified_beam_search":
|
||||
key = f"num_active_paths_{params.num_active_paths}"
|
||||
else:
|
||||
raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
|
||||
return {key: decode_results}
|
||||
|
||||
|
||||
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}-{params.suffix}.txt"
|
||||
results = sorted(results)
|
||||
store_transcripts(filename=recog_path, texts=results)
|
||||
logging.info(f"The transcripts are stored in {recog_path}")
|
||||
|
||||
# The following prints out WERs, per-word error statistics and aligned
|
||||
# ref/hyp pairs.
|
||||
errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
|
||||
with open(errs_filename, "w") as f:
|
||||
wer = write_error_stats(
|
||||
f, f"{test_set_name}-{key}", results, enable_log=True
|
||||
)
|
||||
test_set_wers[key] = wer
|
||||
|
||||
logging.info("Wrote detailed error stats to {}".format(errs_filename))
|
||||
|
||||
test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
|
||||
errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
|
||||
with open(errs_info, "w") as f:
|
||||
print("settings\tWER", file=f)
|
||||
for key, val in test_set_wers:
|
||||
print("{}\t{}".format(key, val), file=f)
|
||||
|
||||
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
|
||||
note = "\tbest for {}".format(test_set_name)
|
||||
for key, val in test_set_wers:
|
||||
s += "{}\t{}{}\n".format(key, val, note)
|
||||
note = ""
|
||||
logging.info(s)
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
parser = get_parser()
|
||||
AsrDataModule.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_len}"
|
||||
|
||||
# for fast_beam_search
|
||||
if params.decoding_method == "fast_beam_search":
|
||||
params.suffix += f"-beam-{params.beam}"
|
||||
params.suffix += f"-max-contexts-{params.max_contexts}"
|
||||
params.suffix += f"-max-states-{params.max_states}"
|
||||
|
||||
if params.use_averaged_model:
|
||||
params.suffix += "-use-averaged-model"
|
||||
|
||||
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
|
||||
logging.info("Decoding started")
|
||||
|
||||
device = torch.device("cpu")
|
||||
if torch.cuda.is_available():
|
||||
device = torch.device("cuda", 0)
|
||||
|
||||
logging.info(f"Device: {device}")
|
||||
|
||||
sp = spm.SentencePieceProcessor()
|
||||
sp.load(params.bpe_model)
|
||||
|
||||
# <blk> and <unk> is defined in local/train_bpe_model.py
|
||||
params.blank_id = sp.piece_to_id("<blk>")
|
||||
params.unk_id = sp.piece_to_id("<unk>")
|
||||
params.vocab_size = sp.get_piece_size()
|
||||
|
||||
logging.info(params)
|
||||
|
||||
logging.info("About to create model")
|
||||
model = get_transducer_model(params)
|
||||
|
||||
if not params.use_averaged_model:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
elif params.avg == 1:
|
||||
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
|
||||
else:
|
||||
start = params.epoch - params.avg + 1
|
||||
filenames = []
|
||||
for i in range(start, params.epoch + 1):
|
||||
if start >= 0:
|
||||
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
|
||||
logging.info(f"averaging {filenames}")
|
||||
model.to(device)
|
||||
model.load_state_dict(average_checkpoints(filenames, device=device))
|
||||
else:
|
||||
if params.iter > 0:
|
||||
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
|
||||
: params.avg + 1
|
||||
]
|
||||
if len(filenames) == 0:
|
||||
raise ValueError(
|
||||
f"No checkpoints found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
elif len(filenames) < params.avg + 1:
|
||||
raise ValueError(
|
||||
f"Not enough checkpoints ({len(filenames)}) found for"
|
||||
f" --iter {params.iter}, --avg {params.avg}"
|
||||
)
|
||||
filename_start = filenames[-1]
|
||||
filename_end = filenames[0]
|
||||
logging.info(
|
||||
"Calculating the averaged model over iteration checkpoints"
|
||||
f" from {filename_start} (excluded) to {filename_end}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
else:
|
||||
assert params.avg > 0, params.avg
|
||||
start = params.epoch - params.avg
|
||||
assert start >= 1, start
|
||||
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
|
||||
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
|
||||
logging.info(
|
||||
f"Calculating the averaged model over epoch range from "
|
||||
f"{start} (excluded) to {params.epoch}"
|
||||
)
|
||||
model.to(device)
|
||||
model.load_state_dict(
|
||||
average_checkpoints_with_averaged_model(
|
||||
filename_start=filename_start,
|
||||
filename_end=filename_end,
|
||||
device=device,
|
||||
)
|
||||
)
|
||||
|
||||
model.to(device)
|
||||
model.eval()
|
||||
model.device = device
|
||||
|
||||
decoding_graph = None
|
||||
if params.decoding_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}")
|
||||
|
||||
librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
|
||||
|
||||
test_clean_cuts = librispeech.test_clean_cuts()
|
||||
test_other_cuts = librispeech.test_other_cuts()
|
||||
|
||||
test_sets = ["test-clean", "test-other"]
|
||||
test_cuts = [test_clean_cuts, test_other_cuts]
|
||||
|
||||
for test_set, test_cut in zip(test_sets, test_cuts):
|
||||
results_dict = decode_dataset(
|
||||
cuts=test_cut,
|
||||
params=params,
|
||||
model=model,
|
||||
sp=sp,
|
||||
decoding_graph=decoding_graph,
|
||||
)
|
||||
|
||||
save_results(
|
||||
params=params,
|
||||
test_set_name=test_set,
|
||||
results_dict=results_dict,
|
||||
)
|
||||
|
||||
logging.info("Done!")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/test_model.py
|
||||
1370
egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py
Executable file
1370
egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py
Executable file
File diff suppressed because it is too large
Load Diff
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/train2.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/zipformer.py
|
||||
@ -0,0 +1 @@
|
||||
../pruned_transducer_stateless7_streaming/zipformer2.py
|
||||
Loading…
x
Reference in New Issue
Block a user