From 4ace5a95b052d338947fc88809a440ccd55cf6b4 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 16 十一月 2023 16:39:52 +0800
Subject: [PATCH] funasr pages
---
egs_modelscope/asr/TEMPLATE/README.md | 163 +++++++++++++++++++++++++++++++++++++++++++++--------
1 files changed, 137 insertions(+), 26 deletions(-)
diff --git a/egs_modelscope/asr/TEMPLATE/README.md b/egs_modelscope/asr/TEMPLATE/README.md
index 30ae8c9..bf6d30b 100644
--- a/egs_modelscope/asr/TEMPLATE/README.md
+++ b/egs_modelscope/asr/TEMPLATE/README.md
@@ -1,7 +1,9 @@
+([绠�浣撲腑鏂嘳(./README_zh.md)|English)
+
# Speech Recognition
-> **Note**:
-> The modelscope pipeline supports all the models in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/modelscope_models.html#pretrained-models-on-modelscope) to inference and finetine. Here we take the typic models as examples to demonstrate the usage.
+> **Note**:
+> The modelscope pipeline supports all the models in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/model_zoo/modelscope_models.html#pretrained-models-on-modelscope) to inference and finetine. Here we take the typic models as examples to demonstrate the usage.
## Inference
@@ -19,21 +21,55 @@
rec_result = inference_pipeline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav')
print(rec_result)
```
+#### [Paraformer-long Model](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary)
+```python
+from modelscope.pipelines import pipeline
+from modelscope.utils.constant import Tasks
+
+inference_pipeline = pipeline(
+ task=Tasks.auto_speech_recognition,
+ model='damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
+ vad_model='damo/speech_fsmn_vad_zh-cn-16k-common-pytorch',
+ #punc_model='damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch',
+ punc_model='damo/punc_ct-transformer_cn-en-common-vocab471067-large',
+)
+
+rec_result = inference_pipeline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/vad_example.wav',
+ batch_size_token=5000, batch_size_token_threshold_s=40, max_single_segment_time=6000)
+print(rec_result)
+```
+
+Where,
+- `batch_size_token` refs to dynamic batch_size and the total tokens of batch is `batch_size_token`, 1 token = 60 ms.
+- `batch_size_token_threshold_s`: The batch_size is set to 1, when the audio duration exceeds the threshold value of `batch_size_token_threshold_s`, specified in `s`.
+- `max_single_segment_time`: The maximum length for audio segmentation in VAD, specified in `ms`.
+
+Suggestion: When encountering OOM (Out of Memory) issues with long audio inputs, as the GPU memory usage increases with the square of the audio duration, there are three possible scenarios:
+- a) In the initial inference stage, GPU memory usage primarily depends on `batch_size_token`. Reducing this value appropriately can help reduce memory usage.
+- b) In the middle of the inference process, when encountering long audio segments segmented by VAD, if the total number of tokens is still smaller than `batch_size_token` but OOM issues persist, reducing `batch_size_token_threshold_s` can help. If the threshold is exceeded, forcing the batch size to 1 can be considered.
+- c) Towards the end of the inference process, when encountering long audio segments segmented by VAD and the total number of tokens is smaller than `batch_size_token` but exceeds the threshold `batch_size_token_threshold_s`, forcing the batch size to 1 may still result in OOM errors. In such cases, reducing `max_single_segment_time` can be considered to shorten the duration of audio segments generated by VAD.
+
#### [Paraformer-online Model](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online/summary)
+##### Streaming Decoding
```python
inference_pipeline = pipeline(
task=Tasks.auto_speech_recognition,
model='damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online',
- model_revision='v1.0.4'
+ model_revision='v1.0.7',
+ update_model=False,
+ mode='paraformer_streaming'
)
import soundfile
speech, sample_rate = soundfile.read("example/asr_example.wav")
-chunk_size = [5, 10, 5] #[5, 10, 5] 600ms, [8, 8, 4] 480ms
-param_dict = {"cache": dict(), "is_final": False, "chunk_size": chunk_size}
+chunk_size = [0, 10, 5] #[0, 10, 5] 600ms, [0, 8, 4] 480ms
+encoder_chunk_look_back = 4 #number of chunks to lookback for encoder self-attention
+decoder_chunk_look_back = 1 #number of encoder chunks to lookback for decoder cross-attention
+param_dict = {"cache": dict(), "is_final": False, "chunk_size": chunk_size,
+ "encoder_chunk_look_back": encoder_chunk_look_back, "decoder_chunk_look_back": decoder_chunk_look_back}
chunk_stride = chunk_size[1] * 960 # 600ms銆�480ms
# first chunk, 600ms
-speech_chunk = speech[0:chunk_stride]
+speech_chunk = speech[0:chunk_stride]
rec_result = inference_pipeline(audio_in=speech_chunk, param_dict=param_dict)
print(rec_result)
# next chunk, 600ms
@@ -41,10 +77,44 @@
rec_result = inference_pipeline(audio_in=speech_chunk, param_dict=param_dict)
print(rec_result)
```
+
+##### Fake Streaming Decoding
+```python
+from modelscope.pipelines import pipeline
+from modelscope.utils.constant import Tasks
+
+inference_pipeline = pipeline(
+ task=Tasks.auto_speech_recognition,
+ model='damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online',
+ model_revision='v1.0.7',
+ update_model=False,
+ mode="paraformer_fake_streaming"
+)
+audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav'
+rec_result = inference_pipeline(audio_in=audio_in)
+print(rec_result)
+```
Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/FunASR/discussions/241)
+#### [Paraformer-contextual Model](https://www.modelscope.cn/models/damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404/summary)
+```python
+from modelscope.pipelines import pipeline
+from modelscope.utils.constant import Tasks
+
+param_dict = dict()
+# param_dict['hotword'] = "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/hotword.txt"
+param_dict['hotword']="閭撻儊鏉� 鐜嬮鏄� 鐜嬫檾鍚�"
+inference_pipeline = pipeline(
+ task=Tasks.auto_speech_recognition,
+ model="damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404",
+ param_dict=param_dict)
+
+rec_result = inference_pipeline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_hotword.wav')
+print(rec_result)
+```
+
#### [UniASR Model](https://www.modelscope.cn/models/damo/speech_UniASR_asr_2pass-zh-cn-8k-common-vocab3445-pytorch-online/summary)
-There are three decoding mode for UniASR model(`fast`銆乣normal`銆乣offline`), for more model detailes, please refer to [docs](https://www.modelscope.cn/models/damo/speech_UniASR_asr_2pass-zh-cn-8k-common-vocab3445-pytorch-online/summary)
+There are three decoding mode for UniASR model(`fast`銆乣normal`銆乣offline`), for more model details, please refer to [docs](https://www.modelscope.cn/models/damo/speech_UniASR_asr_2pass-zh-cn-8k-common-vocab3445-pytorch-online/summary)
```python
decoding_model = "fast" # "fast"銆�"normal"銆�"offline"
inference_pipeline = pipeline(
@@ -57,11 +127,33 @@
```
The decoding mode of `fast` and `normal` is fake streaming, which could be used for evaluating of recognition accuracy.
Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/FunASR/discussions/151)
+
+#### [Paraformer-Spk](https://modelscope.cn/models/damo/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn/summary)
+This model allows user to get recognition results which contain speaker info of each sentence. Refer to [CAM++](https://modelscope.cn/models/damo/speech_campplus_speaker-diarization_common/summary) for detailed information about speaker diarization model.
+```python
+from modelscope.pipelines import pipeline
+from modelscope.utils.constant import Tasks
+
+if __name__ == '__main__':
+ audio_in = 'https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_speaker_demo.wav'
+ output_dir = "./results"
+ inference_pipeline = pipeline(
+ task=Tasks.auto_speech_recognition,
+ model='damo/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn',
+ model_revision='v0.0.2',
+ vad_model='damo/speech_fsmn_vad_zh-cn-16k-common-pytorch',
+ punc_model='damo/punc_ct-transformer_cn-en-common-vocab471067-large',
+ output_dir=output_dir,
+ )
+ rec_result = inference_pipeline(audio_in=audio_in, batch_size_token=5000, batch_size_token_threshold_s=40, max_single_segment_time=6000)
+ print(rec_result)
+```
+
#### [RNN-T-online model]()
Undo
#### [MFCCA Model](https://www.modelscope.cn/models/NPU-ASLP/speech_mfcca_asr-zh-cn-16k-alimeeting-vocab4950/summary)
-For more model detailes, please refer to [docs](https://www.modelscope.cn/models/NPU-ASLP/speech_mfcca_asr-zh-cn-16k-alimeeting-vocab4950/summary)
+For more model details, please refer to [docs](https://www.modelscope.cn/models/NPU-ASLP/speech_mfcca_asr-zh-cn-16k-alimeeting-vocab4950/summary)
```python
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
@@ -79,18 +171,18 @@
### API-reference
#### Define pipeline
- `task`: `Tasks.auto_speech_recognition`
-- `model`: model name in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/modelscope_models.html#pretrained-models-on-modelscope), or model path in local disk
+- `model`: model name in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/model_zoo/modelscope_models.html#pretrained-models-on-modelscope), or model path in local disk
- `ngpu`: `1` (Default), decoding on GPU. If ngpu=0, decoding on CPU
-- `ncpu`: `1` (Default), sets the number of threads used for intraop parallelism on CPU
+- `ncpu`: `1` (Default), sets the number of threads used for intraop parallelism on CPU
- `output_dir`: `None` (Default), the output path of results if set
- `batch_size`: `1` (Default), batch size when decoding
#### Infer pipeline
-- `audio_in`: the input to decode, which could be:
+- `audio_in`: the input to decode, which could be:
- wav_path, `e.g.`: asr_example.wav,
- - pcm_path, `e.g.`: asr_example.pcm,
+ - pcm_path, `e.g.`: asr_example.pcm,
- audio bytes stream, `e.g.`: bytes data from a microphone
- audio sample point锛宍e.g.`: `audio, rate = soundfile.read("asr_example_zh.wav")`, the dtype is numpy.ndarray or torch.Tensor
- - wav.scp, kaldi style wav list (`wav_id \t wav_path`), `e.g.`:
+ - wav.scp, kaldi style wav list (`wav_id \t wav_path`), `e.g.`:
```text
asr_example1 ./audios/asr_example1.wav
asr_example2 ./audios/asr_example2.wav
@@ -103,7 +195,7 @@
FunASR also offer recipes [egs_modelscope/asr/TEMPLATE/infer.sh](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/TEMPLATE/infer.sh) to decode with multi-thread CPUs, or multi GPUs.
#### Settings of `infer.sh`
-- `model`: model name in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/modelscope_models.html#pretrained-models-on-modelscope), or model path in local disk
+- `model`: model name in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/model_zoo/modelscope_models.html#pretrained-models-on-modelscope), or model path in local disk
- `data_dir`: the dataset dir needs to include `wav.scp`. If `${data_dir}/text` is also exists, CER will be computed
- `output_dir`: output dir of the recognition results
- `batch_size`: `64` (Default), batch size of inference on gpu
@@ -148,15 +240,19 @@
[finetune.py](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/TEMPLATE/finetune.py)
```python
import os
+
from modelscope.metainfo import Trainers
from modelscope.trainers import build_trainer
-from modelscope.msdatasets.audio.asr_dataset import ASRDataset
+
+from funasr.datasets.ms_dataset import MsDataset
+from funasr.utils.modelscope_param import modelscope_args
+
def modelscope_finetune(params):
if not os.path.exists(params.output_dir):
os.makedirs(params.output_dir, exist_ok=True)
# dataset split ["train", "validation"]
- ds_dict = ASRDataset.load(params.data_path, namespace='speech_asr')
+ ds_dict = MsDataset.load(params.data_path)
kwargs = dict(
model=params.model,
data_dir=ds_dict,
@@ -164,21 +260,32 @@
work_dir=params.output_dir,
batch_bins=params.batch_bins,
max_epoch=params.max_epoch,
- lr=params.lr)
+ lr=params.lr,
+ mate_params=params.param_dict)
trainer = build_trainer(Trainers.speech_asr_trainer, default_args=kwargs)
trainer.train()
if __name__ == '__main__':
- from funasr.utils.modelscope_param import modelscope_args
- params = modelscope_args(model="damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")
- params.output_dir = "./checkpoint" # 妯″瀷淇濆瓨璺緞
- params.data_path = "speech_asr_aishell1_trainsets" # 鏁版嵁璺緞锛屽彲浠ヤ负modelscope涓凡涓婁紶鏁版嵁锛屼篃鍙互鏄湰鍦版暟鎹�
- params.dataset_type = "small" # 灏忔暟鎹噺璁剧疆small锛岃嫢鏁版嵁閲忓ぇ浜�1000灏忔椂锛岃浣跨敤large
- params.batch_bins = 2000 # batch size锛屽鏋渄ataset_type="small"锛宐atch_bins鍗曚綅涓篺bank鐗瑰緛甯ф暟锛屽鏋渄ataset_type="large"锛宐atch_bins鍗曚綅涓烘绉掞紝
- params.max_epoch = 50 # 鏈�澶ц缁冭疆鏁�
- params.lr = 0.00005 # 璁剧疆瀛︿範鐜�
-
+ params = modelscope_args(model="damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch", data_path="./data")
+ params.output_dir = "./checkpoint" # m妯″瀷淇濆瓨璺緞
+ params.data_path = "speech_asr_aishell1_trainsets" # 鏁版嵁璺緞
+ params.dataset_type = "small" # 灏忔暟鎹噺璁剧疆small锛岃嫢鏁版嵁閲忓ぇ浜�1000灏忔椂锛岃浣跨敤large
+ params.batch_bins = 2000 # batch size锛屽鏋渄ataset_type="small"锛宐atch_bins鍗曚綅涓篺bank鐗瑰緛甯ф暟锛屽鏋渄ataset_type="large"锛宐atch_bins鍗曚綅涓烘绉掞紝
+ params.max_epoch = 20 # 鏈�澶ц缁冭疆鏁�
+ params.lr = 0.00005 # 璁剧疆瀛︿範鐜�
+ init_param = [] # 鍒濆妯″瀷璺緞锛岄粯璁ゅ姞杞絤odelscope妯″瀷鍒濆鍖栵紝渚嬪: ["checkpoint/20epoch.pb"]
+ freeze_param = [] # 妯″瀷鍙傛暟freeze, 渚嬪: ["encoder"]
+ ignore_init_mismatch = True # 鏄惁蹇界暐妯″瀷鍙傛暟鍒濆鍖栦笉鍖归厤
+ use_lora = False # 鏄惁浣跨敤lora杩涜妯″瀷寰皟
+ params.param_dict = {"init_param":init_param, "freeze_param": freeze_param, "ignore_init_mismatch": ignore_init_mismatch}
+ if use_lora:
+ enable_lora = True
+ lora_bias = "all"
+ lora_params = {"lora_list":['q','v'], "lora_rank":8, "lora_alpha":16, "lora_dropout":0.1}
+ lora_config = {"enable_lora": enable_lora, "lora_bias": lora_bias, "lora_params": lora_params}
+ params.param_dict.update(lora_config)
+
modelscope_finetune(params)
```
@@ -195,6 +302,10 @@
- `batch_bins`: batch size. For dataset_type is `small`, `batch_bins` indicates the feature frames. For dataset_type is `large`, `batch_bins` indicates the duration in ms
- `max_epoch`: number of training epoch
- `lr`: learning rate
+ - `init_param`: `[]`(Default), init model path, load modelscope model initialization by default. For example: ["checkpoint/20epoch.pb"]
+ - `freeze_param`: `[]`(Default), Freeze model parameters. For example锛歔"encoder"]
+ - `ignore_init_mismatch`: `True`(Default), Ignore size mismatch when loading pre-trained model
+ - `use_lora`: `False`(Default), Fine-tuning model use lora, more detail please refer to [LORA](https://arxiv.org/pdf/2106.09685.pdf)
- Training data formats锛�
```sh
--
Gitblit v1.9.1