Merge pull request #389 from alibaba-damo-academy/main

yhliang

2023-04-20 5de9e75d587b752d8d1063cc7903c4571df99189

Merge pull request #389 from alibaba-damo-academy/main

update dev_lyh

 .github/workflows/main.yml

@@ -18,10 +18,6 @@
        with:
          docs-folder: "docs/"
          pre-build-command: "pip install sphinx-markdown-tables nbsphinx jinja2 recommonmark sphinx_rtd_theme"
      - uses: ammaraskar/sphinx-action@master
        with:
          docs-folder: "docs_cn/"
          pre-build-command: "pip install sphinx-markdown-tables nbsphinx jinja2 recommonmark sphinx_rtd_theme"

      - name: deploy copy
        if: github.ref == 'refs/heads/main' || github.ref == 'refs/heads/dev_wjm' || github.ref == 'refs/heads/dev_lyh'
@@ -31,9 +27,6 @@
          mkdir public/en
          touch public/en/.nojekyll
          cp -r docs/_build/html/* public/en/
          mkdir public/cn
          touch public/cn/.nojekyll
          cp -r docs_cn/_build/html/* public/cn/
          mkdir public/m2met2
          touch public/m2met2/.nojekyll
          cp -r docs_m2met2/_build/html/* public/m2met2/

 .gitignore

@@ -13,3 +13,7 @@
.eggs
MaaS-lib
.gitignore
.egg*
dist
build
funasr.egg-info

 README.md

@@ -12,13 +12,13 @@
[**News**](https://github.com/alibaba-damo-academy/FunASR#whats-new) 
| [**Highlights**](#highlights)
| [**Installation**](#installation)
| [**Docs_EN**](https://alibaba-damo-academy.github.io/FunASR/en/index.html)
| [**Docs**](https://alibaba-damo-academy.github.io/FunASR/en/index.html)
| [**Tutorial**](https://github.com/alibaba-damo-academy/FunASR/wiki#funasr%E7%94%A8%E6%88%B7%E6%89%8B%E5%86%8C)
| [**Papers**](https://github.com/alibaba-damo-academy/FunASR#citations)
| [**Runtime**](https://github.com/alibaba-damo-academy/FunASR/tree/main/funasr/runtime)
| [**Model Zoo**](https://github.com/alibaba-damo-academy/FunASR/blob/main/docs/modelscope_models.md)
| [**Contact**](#contact)

|
[**M2MET2.0 Guidence_CN**](https://alibaba-damo-academy.github.io/FunASR/m2met2_cn/index.html)
| [**M2MET2.0 Guidence_EN**](https://alibaba-damo-academy.github.io/FunASR/m2met2/index.html)


 docs/FQA.md

New file
@@ -0,0 +1,22 @@
# FQA

## How to use VAD model by modelscope pipeline
Ref to [docs](https://github.com/alibaba-damo-academy/FunASR/discussions/236)

## How to use Punctuation model by modelscope pipeline
Ref to [docs](https://github.com/alibaba-damo-academy/FunASR/discussions/238)

## How to use Parafomrer model for streaming by modelscope pipeline
Ref to [docs](https://github.com/alibaba-damo-academy/FunASR/discussions/241)

## How to use vad, asr and punc model by modelscope pipeline
Ref to [docs](https://github.com/alibaba-damo-academy/FunASR/discussions/278)

## How to combine vad, asr, punc and nnlm models inside modelscope pipeline
Ref to [docs](https://github.com/alibaba-damo-academy/FunASR/discussions/134)

## How to combine timestamp prediction model by modelscope pipeline
Ref to [docs](https://github.com/alibaba-damo-academy/FunASR/discussions/246)

## How to switch decoding mode between online and offline for UniASR model
Ref to [docs](https://github.com/alibaba-damo-academy/FunASR/discussions/151)

 docs/benchmark/benchmark_libtorch.md

New file
@@ -0,0 +1 @@
../../funasr/runtime/python/benchmark_libtorch.md

 docs/benchmark/benchmark_onnx.md

New file
@@ -0,0 +1 @@
../../funasr/runtime/python/benchmark_onnx.md

 docs/images/dingding.jpg

 docs/index.rst

@@ -47,6 +47,7 @@
   ./modescope_pipeline/punc_pipeline.md
   ./modescope_pipeline/tp_pipeline.md
   ./modescope_pipeline/sv_pipeline.md
   ./modescope_pipeline/sd_pipeline.md
   ./modescope_pipeline/lm_pipeline.md

.. toctree::
@@ -63,11 +64,22 @@

.. toctree::
   :maxdepth: 1
   :caption: Benchmark and Leadboard

   ./benchmark/benchmark_onnx.md
   ./benchmark/benchmark_libtorch.md

.. toctree::
   :maxdepth: 1
   :caption: Papers

   ./papers.md

.. toctree::
   :maxdepth: 1
   :caption: FQA

   ./FQA.md


Indices and tables

 docs/modelscope_models.md

@@ -80,7 +80,7 @@
| [Xvector](https://www.modelscope.cn/models/damo/speech_xvector_sv-zh-cn-cnceleb-16k-spk3465-pytorch/summary) | CNCeleb (1,200 hours)  |   17.5M    |    3465    |    Xvector, speaker verification, Chinese   |
| [Xvector](https://www.modelscope.cn/models/damo/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch/summary) | CallHome (60 hours) |    61M     |    6135    |   Xvector, speaker verification, English    |

### Speaker diarization Models
### Speaker Diarization Models

|                                                    Model Name                                                    |    Training Data    | Parameters | Notes |
|:----------------------------------------------------------------------------------------------------------------:|:-------------------:|:----------:|:------|

 docs/modescope_pipeline/asr_pipeline.md

@@ -1,20 +1,196 @@
# 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 model of Paraformer and Paraformer-online as example to demonstrate the usage.

## Inference

### Quick start
#### [Paraformer 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 with you data
inference_pipeline = pipeline(
    task=Tasks.auto_speech_recognition,
    model='damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
)

#### Inference with multi-threads on CPU
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-online model](https://www.modelscope.cn/models/damo/speech_paraformer_asr_nat-zh-cn-16k-common-vocab8404-online/summary)
```python
inference_pipeline = pipeline(
    task=Tasks.auto_speech_recognition,
    model='damo/speech_paraformer_asr_nat-zh-cn-16k-common-vocab8404-online',
    )
import soundfile
speech, sample_rate = soundfile.read("example/asr_example.wav")

#### Inference with multi GPU
param_dict = {"cache": dict(), "is_final": False}
chunk_stride = 7680# 480ms
# first chunk, 480ms
speech_chunk = speech[0:chunk_stride] 
rec_result = inference_pipeline(audio_in=speech_chunk, param_dict=param_dict)
print(rec_result)
# next chunk, 480ms
speech_chunk = speech[chunk_stride:chunk_stride+chunk_stride]
rec_result = inference_pipeline(audio_in=speech_chunk, param_dict=param_dict)
print(rec_result)
```
Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/FunASR/discussions/241)

#### [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`、`normal`、`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)
```python
decoding_model = "fast" # "fast"、"normal"、"offline"
inference_pipeline = pipeline(
    task=Tasks.auto_speech_recognition,
    model='damo/speech_UniASR_asr_2pass-minnan-16k-common-vocab3825',
    param_dict={"decoding_model": decoding_model})

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)
```
The decoding mode of `fast` and `normal`
Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/FunASR/discussions/151)
#### [RNN-T-online model]()
Undo

#### 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
- `ngpu`: 1 (Defalut), decoding on GPU. If ngpu=0, decoding on CPU
- `ncpu`: 1 (Defalut), sets the number of threads used for intraop parallelism on CPU 
- `output_dir`: None (Defalut), the output path of results if set
- `batch_size`: 1 (Defalut), batch size when decoding
##### infer pipeline
- `audio_in`: the input to decode, which could be: 
  - wav_path, `e.g.`: asr_example.wav,
  - pcm_path, `e.g.`: asr_example.pcm, 
  - audio bytes stream, `e.g.`: bytes data from a microphone
  - audio sample point，`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.`: 
  ```cat wav.scp
  asr_example1  ./audios/asr_example1.wav
  asr_example2  ./audios/asr_example2.wav
  ```
  In this case of `wav.scp` input, `output_dir` must be set to save the output results
- `audio_fs`: audio sampling rate, only set when audio_in is pcm audio
- `output_dir`: None (Defalut), the output path of results if set

### Inference with multi-thread CPUs or multi GPUs
FunASR also offer recipes [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.

- Setting parameters in `infer.sh`
    - <strong>model:</strong> # model name on ModelScope
    - <strong>data_dir:</strong> # the dataset dir needs to include `${data_dir}/wav.scp`. If `${data_dir}/text` is also exists, CER will be computed
    - <strong>output_dir:</strong> # result dir
    - <strong>batch_size:</strong> # batchsize of inference
    - <strong>gpu_inference:</strong> # whether to perform gpu decoding, set false for cpu decoding
    - <strong>gpuid_list:</strong> # set gpus, e.g., gpuid_list="0,1"
    - <strong>njob:</strong> # the number of jobs for CPU decoding, if `gpu_inference`=false, use CPU decoding, please set `njob`

- Decode with multi GPUs:
```shell
    bash infer.sh \
    --model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" \
    --data_dir "./data/test" \
    --output_dir "./results" \
    --batch_size 64 \
    --gpu_inference true \
    --gpuid_list "0,1"
```
- Decode with multi-thread CPUs:
```shell
    bash infer.sh \
    --model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" \
    --data_dir "./data/test" \
    --output_dir "./results" \
    --gpu_inference false \
    --njob 64
```

- Results

The decoding results can be found in `$output_dir/1best_recog/text.cer`, which includes recognition results of each sample and the CER metric of the whole test set.

If you decode the SpeechIO test sets, you can use textnorm with `stage`=3, and `DETAILS.txt`, `RESULTS.txt` record the results and CER after text normalization.


## Finetune with pipeline

### Quick start
[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

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')
    kwargs = dict(
        model=params.model,
        data_dir=ds_dict,
        dataset_type=params.dataset_type,
        work_dir=params.output_dir,
        batch_bins=params.batch_bins,
        max_epoch=params.max_epoch,
        lr=params.lr)
    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"                      # 模型保存路径
    params.data_path = "speech_asr_aishell1_trainsets"      # 数据路径，可以为modelscope中已上传数据，也可以是本地数据
    params.dataset_type = "small"                           # 小数据量设置small，若数据量大于1000小时，请使用large
    params.batch_bins = 2000                                # batch size，如果dataset_type="small"，batch_bins单位为fbank特征帧数，如果dataset_type="large"，batch_bins单位为毫秒，
    params.max_epoch = 50                                   # 最大训练轮数
    params.lr = 0.00005                                     # 设置学习率
    
    modelscope_finetune(params)
```

```shell
python finetune.py &> log.txt &
```

### Finetune with your data

## Inference with your finetuned model
- Modify finetune training related parameters in [finetune.py](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/TEMPLATE/finetune.py)
    - <strong>output_dir:</strong> # result dir
    - <strong>data_dir:</strong> # the dataset dir needs to include files: `train/wav.scp`, `train/text`; `validation/wav.scp`, `validation/text`
    - <strong>dataset_type:</strong> # for dataset larger than 1000 hours, set as `large`, otherwise set as `small`
    - <strong>batch_bins:</strong> # 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
    - <strong>max_epoch:</strong> # number of training epoch
    - <strong>lr:</strong> # learning rate

- Then you can run the pipeline to finetune with:
```shell
python finetune.py
```
If you want finetune with multi-GPUs, you could:
```shell
CUDA_VISIBLE_DEVICES=1,2 python -m torch.distributed.launch --nproc_per_node 2 finetune.py > log.txt 2>&1
```
## Inference with your finetuned model
- Modify inference related parameters in [infer_after_finetune.py](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/TEMPLATE/infer_after_finetune.py)
    - <strong>modelscope_model_name: </strong> # model name on ModelScope
    - <strong>output_dir:</strong> # result dir
    - <strong>data_dir:</strong> # the dataset dir needs to include `test/wav.scp`. If `test/text` is also exists, CER will be computed
    - <strong>decoding_model_name:</strong> # set the checkpoint name for decoding, e.g., `valid.cer_ctc.ave.pb`
    - <strong>batch_size:</strong> # batchsize of inference  

- Then you can run the pipeline to finetune with:
```python
    python infer_after_finetune.py
```

 docs/modescope_pipeline/lm_pipeline.md

@@ -1,10 +1,10 @@
# Speech Recognition
# Language Models

## Inference with pipeline
### Quick start
#### Inference with you data
#### Inference with multi-threads on CPU
#### Inference with multi GPU
### Inference with you data
### Inference with multi-threads on CPU
### Inference with multi GPU

## Finetune with pipeline
### Quick start

 docs/modescope_pipeline/punc_pipeline.md

@@ -4,11 +4,11 @@

### Quick start

#### Inference with you data
### Inference with you data

#### Inference with multi-threads on CPU
### Inference with multi-threads on CPU

#### Inference with multi GPU
### Inference with multi GPU

## Finetune with pipeline


 docs/modescope_pipeline/quick_start.md

@@ -59,8 +59,7 @@

inference_pipeline = pipeline(
    task=Tasks.speech_timestamp,
    model='damo/speech_timestamp_prediction-v1-16k-offline',
    output_dir='./tmp')
    model='damo/speech_timestamp_prediction-v1-16k-offline',)

rec_result = inference_pipeline(
    audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_timestamps.wav',
@@ -86,6 +85,71 @@
# speaker verification
rec_result = inference_sv_pipline(audio_in=('https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_enroll.wav','https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_same.wav'))
print(rec_result["scores"][0])
```

### Speaker diarization
#### SOND
```python
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

inference_diar_pipline = pipeline(
    mode="sond_demo",
    num_workers=0,
    task=Tasks.speaker_diarization,
    diar_model_config="sond.yaml",
    model='damo/speech_diarization_sond-en-us-callhome-8k-n16k4-pytorch',
    sv_model="damo/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch",
    sv_model_revision="master",
)

audio_list=[
    "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_data/record.wav",
    "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_data/spk_A.wav",
    "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_data/spk_B.wav",
    "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_data/spk_B1.wav"
]

results = inference_diar_pipline(audio_in=audio_list)
print(results)
```

### FAQ
#### How to switch device from GPU to CPU with pipeline

The pipeline defaults to decoding with GPU (`ngpu=1`) when GPU is available. If you want to switch to CPU, you could set `ngpu=0`
```python
inference_pipeline = pipeline(
    task=Tasks.auto_speech_recognition,
    model='damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
    ngpu=0,
)
```

#### How to infer from local model path
Download model to local dir, by modelscope-sdk

```python
from modelscope.hub.snapshot_download import snapshot_download

local_dir_root = "./models_from_modelscope"
model_dir = snapshot_download('damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch', cache_dir=local_dir_root)
```

Or download model to local dir, by git lfs
```shell
git lfs install
# git clone https://www.modelscope.cn/<namespace>/<model-name>.git
git clone https://www.modelscope.cn/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch.git
```

Infer with local model path
```python
local_dir_root = "./models_from_modelscope/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
inference_pipeline = pipeline(
    task=Tasks.auto_speech_recognition,
    model=local_dir_root,
)
```

## Finetune with pipeline
@@ -132,6 +196,10 @@
```shell
python finetune.py &> log.txt &
```

### FAQ
### Multi GPUs training and distributed training

If you want finetune with multi-GPUs, you could:
```shell
CUDA_VISIBLE_DEVICES=1,2 python -m torch.distributed.launch --nproc_per_node 2 finetune.py > log.txt 2>&1

 docs/modescope_pipeline/sd_pipeline.md

New file
@@ -0,0 +1,20 @@
# Speaker Diarization

## Inference with pipeline

### Quick start

### Inference with you data

### Inference with multi-threads on CPU

### Inference with multi GPU

## Finetune with pipeline

### Quick start

### Finetune with your data

## Inference with your finetuned model


 docs/modescope_pipeline/sv_pipeline.md

@@ -4,11 +4,11 @@

### Quick start

#### Inference with you data
### Inference with you data

#### Inference with multi-threads on CPU
### Inference with multi-threads on CPU

#### Inference with multi GPU
### Inference with multi GPU

## Finetune with pipeline


 docs/modescope_pipeline/tp_pipeline.md

@@ -4,11 +4,11 @@

### Quick start

#### Inference with you data
### Inference with you data

#### Inference with multi-threads on CPU
### Inference with multi-threads on CPU

#### Inference with multi GPU
### Inference with multi GPU

## Finetune with pipeline


 docs/modescope_pipeline/vad_pipeline.md

@@ -1,14 +1,107 @@
# Voice Activity Detection

## Inference with pipeline
> **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 model of FSMN-VAD as example to demonstrate the usage.

## Inference

### Quick start
#### [FSMN-VAD model](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary)
```python
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

#### Inference with you data
inference_pipeline = pipeline(
    task=Tasks.voice_activity_detection,
    model='damo/speech_fsmn_vad_zh-cn-16k-common-pytorch',
)

#### Inference with multi-threads on CPU
segments_result = inference_pipeline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/vad_example.wav')
print(segments_result)
```
#### [FSMN-VAD-online model](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary)
```python
inference_pipeline = pipeline(
    task=Tasks.auto_speech_recognition,
    model='damo/speech_fsmn_vad_zh-cn-16k-common-pytorch',
    )
import soundfile
speech, sample_rate = soundfile.read("example/asr_example.wav")

#### Inference with multi GPU
param_dict = {"in_cache": dict(), "is_final": False}
chunk_stride = 1600# 100ms
# first chunk, 100ms
speech_chunk = speech[0:chunk_stride] 
rec_result = inference_pipeline(audio_in=speech_chunk, param_dict=param_dict)
print(rec_result)
# next chunk, 480ms
speech_chunk = speech[chunk_stride:chunk_stride+chunk_stride]
rec_result = inference_pipeline(audio_in=speech_chunk, param_dict=param_dict)
print(rec_result)
```
Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/FunASR/discussions/236)


#### 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
- `ngpu`: 1 (Defalut), decoding on GPU. If ngpu=0, decoding on CPU
- `ncpu`: 1 (Defalut), sets the number of threads used for intraop parallelism on CPU 
- `output_dir`: None (Defalut), the output path of results if set
- `batch_size`: 1 (Defalut), batch size when decoding
##### infer pipeline
- `audio_in`: the input to decode, which could be: 
  - wav_path, `e.g.`: asr_example.wav,
  - pcm_path, `e.g.`: asr_example.pcm, 
  - audio bytes stream, `e.g.`: bytes data from a microphone
  - audio sample point，`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.`: 
  ```cat wav.scp
  asr_example1  ./audios/asr_example1.wav
  asr_example2  ./audios/asr_example2.wav
  ```
  In this case of `wav.scp` input, `output_dir` must be set to save the output results
- `audio_fs`: audio sampling rate, only set when audio_in is pcm audio
- `output_dir`: None (Defalut), the output path of results if set

### Inference with multi-thread CPUs or multi GPUs
FunASR also offer recipes [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.

- Setting parameters in `infer.sh`
    - <strong>model:</strong> # model name on ModelScope
    - <strong>data_dir:</strong> # the dataset dir needs to include `${data_dir}/wav.scp`. If `${data_dir}/text` is also exists, CER will be computed
    - <strong>output_dir:</strong> # result dir
    - <strong>batch_size:</strong> # batchsize of inference
    - <strong>gpu_inference:</strong> # whether to perform gpu decoding, set false for cpu decoding
    - <strong>gpuid_list:</strong> # set gpus, e.g., gpuid_list="0,1"
    - <strong>njob:</strong> # the number of jobs for CPU decoding, if `gpu_inference`=false, use CPU decoding, please set `njob`

- Decode with multi GPUs:
```shell
    bash infer.sh \
    --model "damo/speech_fsmn_vad_zh-cn-16k-common-pytorch" \
    --data_dir "./data/test" \
    --output_dir "./results" \
    --gpu_inference true \
    --gpuid_list "0,1"
```
- Decode with multi-thread CPUs:
```shell
    bash infer.sh \
    --model "damo/speech_fsmn_vad_zh-cn-16k-common-pytorch" \
    --data_dir "./data/test" \
    --output_dir "./results" \
    --gpu_inference false \
    --njob 64
```

- Results

The decoding results can be found in `$output_dir/1best_recog/text.cer`, which includes recognition results of each sample and the CER metric of the whole test set.

If you decode the SpeechIO test sets, you can use textnorm with `stage`=3, and `DETAILS.txt`, `RESULTS.txt` record the results and CER after text normalization.


## Finetune with pipeline


 docs/recipe/sd_recipe.md

New file
@@ -0,0 +1,129 @@
# Speaker Diarization
Here we take "Training a paraformer model from scratch using the AISHELL-1 dataset" as an example to introduce how to use FunASR. According to this example, users can similarly employ other datasets (such as AISHELL-2 dataset, etc.) to train other models (such as conformer, transformer, etc.).

## Overall Introduction
We provide a recipe `egs/aishell/paraformer/run.sh` for training a paraformer model on AISHELL-1 dataset. This recipe consists of five stages, supporting training on multiple GPUs and decoding by CPU or GPU. Before introducing each stage in detail, we first explain several parameters which should be set by users.
- `CUDA_VISIBLE_DEVICES`: visible gpu list
- `gpu_num`: the number of GPUs used for training
- `gpu_inference`: whether to use GPUs for decoding
- `njob`: for CPU decoding, indicating the total number of CPU jobs; for GPU decoding, indicating the number of jobs on each GPU
- `data_aishell`: the raw path of AISHELL-1 dataset
- `feats_dir`: the path for saving processed data
- `nj`: the number of jobs for data preparation
- `speed_perturb`: the range of speech perturbed
- `exp_dir`: the path for saving experimental results
- `tag`: the suffix of experimental result directory

## Stage 0: Data preparation
This stage processes raw AISHELL-1 dataset `$data_aishell` and generates the corresponding `wav.scp` and `text` in `$feats_dir/data/xxx`. `xxx` means `train/dev/test`. Here we assume users have already downloaded AISHELL-1 dataset. If not, users can download data [here](https://www.openslr.org/33/) and set the path for `$data_aishell`. The examples of `wav.scp` and `text` are as follows:
* `wav.scp`
```
BAC009S0002W0122 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0122.wav
BAC009S0002W0123 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0123.wav
BAC009S0002W0124 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0124.wav
...
```
* `text`
```
BAC009S0002W0122 而 对 楼 市 成 交 抑 制 作 用 最 大 的 限 购
BAC009S0002W0123 也 成 为 地 方 政 府 的 眼 中 钉
BAC009S0002W0124 自 六 月 底 呼 和 浩 特 市 率 先 宣 布 取 消 限 购 后
...
```
These two files both have two columns, while the first column is wav ids and the second column is the corresponding wav paths/label tokens.

## Stage 1: Feature Generation
This stage extracts FBank features from `wav.scp` and apply speed perturbation as data augmentation according to `speed_perturb`. Users can set `nj` to control the number of jobs for feature generation. The generated features are saved in `$feats_dir/dump/xxx/ark` and the corresponding `feats.scp` files are saved as `$feats_dir/dump/xxx/feats.scp`. An example of `feats.scp` can be seen as follows:
* `feats.scp`
```
...
BAC009S0002W0122_sp0.9 /nfs/funasr_data/aishell-1/dump/fbank/train/ark/feats.16.ark:592751055
...
```
Note that samples in this file have already been shuffled randomly. This file contains two columns. The first column is wav ids while the second column is kaldi-ark feature paths. Besides, `speech_shape` and `text_shape` are also generated in this stage, denoting the speech feature shape and text length of each sample. The examples are shown as follows:
* `speech_shape`
```
...
BAC009S0002W0122_sp0.9 665,80
...
```
* `text_shape`
```
...
BAC009S0002W0122_sp0.9 15
...
```
These two files have two columns. The first column is wav ids and the second column is the corresponding speech feature shape and text length.

## Stage 2: Dictionary Preparation
This stage processes the dictionary, which is used as a mapping between label characters and integer indices during ASR training. The processed dictionary file is saved as `$feats_dir/data/$lang_toekn_list/$token_type/tokens.txt`. An example of `tokens.txt` is as follows:
* `tokens.txt`
```
<blank>
<s>
</s>
一
丁
...
龚
龟
<unk>
```
* `<blank>`: indicates the blank token for CTC
* `<s>`: indicates the start-of-sentence token
* `</s>`: indicates the end-of-sentence token
* `<unk>`: indicates the out-of-vocabulary token

## Stage 3: Training
This stage achieves the training of the specified model. To start training, users should manually set `exp_dir`, `CUDA_VISIBLE_DEVICES` and `gpu_num`, which have already been explained above. By default, the best `$keep_nbest_models` checkpoints on validation dataset will be averaged to generate a better model and adopted for decoding.

* DDP Training

We support the DistributedDataParallel (DDP) training and the detail can be found [here](https://pytorch.org/tutorials/intermediate/ddp_tutorial.html). To enable DDP training, please set `gpu_num` greater than 1. For example, if you set `CUDA_VISIBLE_DEVICES=0,1,5,6,7` and `gpu_num=3`, then the gpus with ids 0, 1 and 5 will be used for training.

* DataLoader

We support an optional iterable-style DataLoader based on [Pytorch Iterable-style DataPipes](https://pytorch.org/data/beta/torchdata.datapipes.iter.html) for large dataset and users can set `dataset_type=large` to enable it. 

* Configuration

The parameters of the training, including model, optimization, dataset, etc., can be set by a YAML file in `conf` directory. Also, users can directly set the parameters in `run.sh` recipe. Please avoid to set the same parameters in both the YAML file and the recipe.

* Training Steps

We support two parameters to specify the training steps, namely `max_epoch` and `max_update`. `max_epoch` indicates the total training epochs while `max_update` indicates the total training steps. If these two parameters are specified at the same time, once the training reaches any one of these two parameters, the training will be stopped.

* Tensorboard

Users can use tensorboard to observe the loss, learning rate, etc. Please run the following command:
```
tensorboard --logdir ${exp_dir}/exp/${model_dir}/tensorboard/train
```

## Stage 4: Decoding
This stage generates the recognition results and calculates the `CER` to verify the performance of the trained model. 

* Mode Selection

As we support paraformer, uniasr, conformer and other models in FunASR, a `mode` parameter should be specified as `asr/paraformer/uniasr` according to the trained model.

* Configuration

We support CTC decoding, attention decoding and hybrid CTC-attention decoding in FunASR, which can be specified by `ctc_weight` in a YAML file in `conf` directory. Specifically, `ctc_weight=1.0` indicates CTC decoding, `ctc_weight=0.0` indicates attention decoding, `0.0<ctc_weight<1.0` indicates hybrid CTC-attention decoding.

* CPU/GPU Decoding

We support CPU and GPU decoding in FunASR. For CPU decoding, you should set `gpu_inference=False` and set `njob` to specify the total number of CPU decoding jobs. For GPU decoding, you should set `gpu_inference=True`. You should also set `gpuid_list` to indicate which GPUs are used for decoding and `njobs` to indicate the number of decoding jobs on each GPU.

* Performance

We adopt `CER` to verify the performance. The results are in `$exp_dir/exp/$model_dir/$decoding_yaml_name/$average_model_name/$dset`, namely `text.cer` and `text.cer.txt`. `text.cer` saves the comparison between the recognized text and the reference text while `text.cer.txt` saves the final `CER` result. The following is an example of `text.cer`:
* `text.cer`
```
...
BAC009S0764W0213(nwords=11,cor=11,ins=0,del=0,sub=0) corr=100.00%,cer=0.00%
ref:    构 建 良 好 的 旅 游 市 场 环 境
res:    构 建 良 好 的 旅 游 市 场 环 境
...
```


 docs_cn/Makefile

File was deleted

 docs_cn/build_task.md

File was deleted

 docs_cn/conf.py

File was deleted

 docs_cn/get_started.md

File was deleted

 docs_cn/images/DeepScience.png

Binary files differ

 docs_cn/images/dingding.jpg

Binary files differ

 docs_cn/images/funasr_logo.jpg

Binary files differ

 docs_cn/images/wechat.png

Binary files differ

 docs_cn/index.rst

File was deleted

 docs_cn/installation.md

File was deleted

 docs_cn/make.bat

File was deleted

 docs_cn/modelscope_models.md

File was deleted

 docs_cn/modelscope_usages.md

File was deleted

 docs_cn/papers.md

File was deleted

 egs_modelscope/asr/TEMPLATE/finetune.py

New file
@@ -0,0 +1,36 @@
import os

from modelscope.metainfo import Trainers
from modelscope.trainers import build_trainer

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 = MsDataset.load(params.data_path)
    kwargs = dict(
        model=params.model,
        data_dir=ds_dict,
        dataset_type=params.dataset_type,
        work_dir=params.output_dir,
        batch_bins=params.batch_bins,
        max_epoch=params.max_epoch,
        lr=params.lr)
    trainer = build_trainer(Trainers.speech_asr_trainer, default_args=kwargs)
    trainer.train()


if __name__ == '__main__':
    params = modelscope_args(model="damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch", data_path="./data")
    params.output_dir = "./checkpoint"              # m模型保存路径
    params.data_path = "./example_data/"            # 数据路径
    params.dataset_type = "small"                   # 小数据量设置small，若数据量大于1000小时，请使用large
    params.batch_bins = 2000                       # batch size，如果dataset_type="small"，batch_bins单位为fbank特征帧数，如果dataset_type="large"，batch_bins单位为毫秒，
    params.max_epoch = 50                           # 最大训练轮数
    params.lr = 0.00005                             # 设置学习率
    
    modelscope_finetune(params)

 egs_modelscope/asr/TEMPLATE/infer.py

New file
@@ -0,0 +1,25 @@
import os
import shutil
import argparse
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

def modelscope_infer(args):
    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpuid)
    inference_pipeline = pipeline(
        task=Tasks.auto_speech_recognition,
        model=args.model,
        output_dir=args.output_dir,
        batch_size=args.batch_size,
    )
    inference_pipeline(audio_in=args.audio_in)

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--model', type=str, default="damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")
    parser.add_argument('--audio_in', type=str, default="./data/test/wav.scp")
    parser.add_argument('--output_dir', type=str, default="./results/")
    parser.add_argument('--batch_size', type=int, default=64)
    parser.add_argument('--gpuid', type=str, default="0")
    args = parser.parse_args()
    modelscope_infer(args)

 egs_modelscope/asr/TEMPLATE/infer.sh

New file
@@ -0,0 +1,96 @@
#!/usr/bin/env bash

set -e
set -u
set -o pipefail

stage=1
stop_stage=2
model="damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
data_dir="./data/test"
output_dir="./results"
batch_size=64
gpu_inference=true    # whether to perform gpu decoding
gpuid_list="0,1"    # set gpus, e.g., gpuid_list="0,1"
njob=4    # the number of jobs for CPU decoding, if gpu_inference=false, use CPU decoding, please set njob

. utils/parse_options.sh || exit 1;

if ${gpu_inference} == "true"; then
    nj=$(echo $gpuid_list | awk -F "," '{print NF}')
else
    nj=$njob
    batch_size=1
    gpuid_list=""
    for JOB in $(seq ${nj}); do
        gpuid_list=$gpuid_list"-1,"
    done
fi

mkdir -p $output_dir/split
split_scps=""
for JOB in $(seq ${nj}); do
    split_scps="$split_scps $output_dir/split/wav.$JOB.scp"
done
perl utils/split_scp.pl ${data_dir}/wav.scp ${split_scps}

if [ $stage -le 1 ] && [ $stop_stage -ge 1 ];then
    echo "Decoding ..."
    gpuid_list_array=(${gpuid_list//,/ })
    for JOB in $(seq ${nj}); do
        {
        id=$((JOB-1))
        gpuid=${gpuid_list_array[$id]}
        mkdir -p ${output_dir}/output.$JOB
        python infer.py \
            --model ${model} \
            --audio_in ${output_dir}/split/wav.$JOB.scp \
            --output_dir ${output_dir}/output.$JOB \
            --batch_size ${batch_size} \
            --gpuid ${gpuid}
        }&
    done
    wait

    mkdir -p ${output_dir}/1best_recog
    for f in token score text; do
        if [ -f "${output_dir}/output.1/1best_recog/${f}" ]; then
          for i in $(seq "${nj}"); do
              cat "${output_dir}/output.${i}/1best_recog/${f}"
          done | sort -k1 >"${output_dir}/1best_recog/${f}"
        fi
    done
fi

if [ $stage -le 2 ] && [ $stop_stage -ge 2 ];then
    echo "Computing WER ..."
    cp ${output_dir}/1best_recog/text ${output_dir}/1best_recog/text.proc
    cp ${data_dir}/text ${output_dir}/1best_recog/text.ref
    python utils/compute_wer.py ${output_dir}/1best_recog/text.ref ${output_dir}/1best_recog/text.proc ${output_dir}/1best_recog/text.cer
    tail -n 3 ${output_dir}/1best_recog/text.cer
fi

if [ $stage -le 3 ] && [ $stop_stage -ge 3 ];then
    echo "SpeechIO TIOBE textnorm"
    echo "$0 --> Normalizing REF text ..."
    ./utils/textnorm_zh.py \
        --has_key --to_upper \
        ${data_dir}/text \
        ${output_dir}/1best_recog/ref.txt

    echo "$0 --> Normalizing HYP text ..."
    ./utils/textnorm_zh.py \
        --has_key --to_upper \
        ${output_dir}/1best_recog/text.proc \
        ${output_dir}/1best_recog/rec.txt
    grep -v $'\t$' ${output_dir}/1best_recog/rec.txt > ${output_dir}/1best_recog/rec_non_empty.txt

    echo "$0 --> computing WER/CER and alignment ..."
    ./utils/error_rate_zh \
        --tokenizer char \
        --ref ${output_dir}/1best_recog/ref.txt \
        --hyp ${output_dir}/1best_recog/rec_non_empty.txt \
        ${output_dir}/1best_recog/DETAILS.txt | tee ${output_dir}/1best_recog/RESULTS.txt
    rm -rf ${output_dir}/1best_recog/rec.txt ${output_dir}/1best_recog/rec_non_empty.txt
fi


 egs_modelscope/asr/TEMPLATE/infer_after_finetune.py

New file
@@ -0,0 +1,48 @@
import json
import os
import shutil

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from modelscope.hub.snapshot_download import snapshot_download

from funasr.utils.compute_wer import compute_wer

def modelscope_infer_after_finetune(params):
    # prepare for decoding

    try:
        pretrained_model_path = snapshot_download(params["modelscope_model_name"], cache_dir=params["output_dir"])
    except BaseException:
        raise BaseException(f"Please download pretrain model from ModelScope firstly.")
    shutil.copy(os.path.join(params["output_dir"], params["decoding_model_name"]), os.path.join(pretrained_model_path, "model.pb"))
    decoding_path = os.path.join(params["output_dir"], "decode_results")
    if os.path.exists(decoding_path):
        shutil.rmtree(decoding_path)
    os.mkdir(decoding_path)

    # decoding
    inference_pipeline = pipeline(
        task=Tasks.auto_speech_recognition,
        model=pretrained_model_path,
        output_dir=decoding_path,
        batch_size=params["batch_size"]
    )
    audio_in = os.path.join(params["data_dir"], "wav.scp")
    inference_pipeline(audio_in=audio_in)

    # computer CER if GT text is set
    text_in = os.path.join(params["data_dir"], "text")
    if os.path.exists(text_in):
        text_proc_file = os.path.join(decoding_path, "1best_recog/text")
        compute_wer(text_in, text_proc_file, os.path.join(decoding_path, "text.cer"))


if __name__ == '__main__':
    params = {}
    params["modelscope_model_name"] = "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
    params["output_dir"] = "./checkpoint"
    params["data_dir"] = "./data/test"
    params["decoding_model_name"] = "valid.acc.ave_10best.pb"
    params["batch_size"] = 64
    modelscope_infer_after_finetune(params)

 egs_modelscope/asr/TEMPLATE/utils

New file
@@ -0,0 +1 @@
../../../egs/aishell/transformer/utils

 egs_modelscope/asr/mfcca/speech_mfcca_asr-zh-cn-16k-alimeeting-vocab4950/infer.py

@@ -7,7 +7,6 @@

from funasr.utils.compute_wer import compute_wer

import pdb;
def modelscope_infer_core(output_dir, split_dir, njob, idx):
    output_dir_job = os.path.join(output_dir, "output.{}".format(idx))
    gpu_id = (int(idx) - 1) // njob

 egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/README.md

@@ -23,21 +23,37 @@

- Setting parameters in `infer.sh`
    - <strong>model:</strong> # model name on ModelScope
    - <strong>data_dir:</strong> # the dataset dir needs to include `test/wav.scp`. If `test/text` is also exists, CER will be computed
    - <strong>data_dir:</strong> # the dataset dir needs to include `${data_dir}/wav.scp`. If `${data_dir}/text` is also exists, CER will be computed
    - <strong>output_dir:</strong> # result dir
    - <strong>batch_size:</strong> # batchsize of inference
    - <strong>gpu_inference:</strong> # whether to perform gpu decoding, set false for cpu decoding
    - <strong>gpuid_list:</strong> # set gpus, e.g., gpuid_list="0,1"
    - <strong>njob:</strong> # the number of jobs for CPU decoding, if `gpu_inference`=false, use CPU decoding, please set `njob`

- Then you can run the pipeline to infer with:
```python
    sh infer.sh
- Decode with multi GPUs:
```shell
    bash infer.sh \
    --model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" \
    --data_dir "./data/test" \
    --output_dir "./results" \
    --batch_size 64 \
    --gpu_inference true \
    --gpuid_list "0,1"
```

- Decode with multi-thread CPUs:
```shell
    bash infer.sh \
    --model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" \
    --data_dir "./data/test" \
    --output_dir "./results" \
    --gpu_inference false \
    --njob 64
```

- Results

The decoding results can be found in `$output_dir/1best_recog/text.cer`, which includes recognition results of each sample and the CER metric of the whole test set.
The decoding results can be found in `${output_dir}/1best_recog/text.cer`, which includes recognition results of each sample and the CER metric of the whole test set.

If you decode the SpeechIO test sets, you can use textnorm with `stage`=3, and `DETAILS.txt`, `RESULTS.txt` record the results and CER after text normalization.


 egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.sh

@@ -14,8 +14,9 @@
gpuid_list="0,1"    # set gpus, e.g., gpuid_list="0,1"
njob=4    # the number of jobs for CPU decoding, if gpu_inference=false, use CPU decoding, please set njob

. utils/parse_options.sh || exit 1;

if ${gpu_inference}; then
if ${gpu_inference} == "true"; then
    nj=$(echo $gpuid_list | awk -F "," '{print NF}')
else
    nj=$njob

 funasr/bin/asr_inference.py

@@ -346,6 +346,8 @@
    **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if word_lm_train_config is not None:

 funasr/bin/asr_inference_launch.py

@@ -1,9 +1,4 @@
#!/usr/bin/env python3
# Copyright ESPnet (https://github.com/espnet/espnet). All Rights Reserved.
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

import torch
torch.set_num_threads(1)

import argparse
import logging

 funasr/bin/asr_inference_mfcca.py

@@ -472,6 +472,8 @@
    **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if word_lm_train_config is not None:

 funasr/bin/asr_inference_paraformer.py

@@ -612,6 +612,8 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)

    if word_lm_train_config is not None:
        raise NotImplementedError("Word LM is not implemented")
@@ -630,6 +632,8 @@
    else:
        hotword_list_or_file = None

    if kwargs.get("device", None) == "cpu":
        ngpu = 0
    if ngpu >= 1 and torch.cuda.is_available():
        device = "cuda"
    else:

 funasr/bin/asr_inference_paraformer_streaming.py

@@ -536,6 +536,8 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)

    if word_lm_train_config is not None:
        raise NotImplementedError("Word LM is not implemented")

 funasr/bin/asr_inference_paraformer_vad.py

@@ -157,6 +157,8 @@
    **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    
    if word_lm_train_config is not None:
        raise NotImplementedError("Word LM is not implemented")

 funasr/bin/asr_inference_paraformer_vad_punc.py

@@ -484,6 +484,8 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)

    if word_lm_train_config is not None:
        raise NotImplementedError("Word LM is not implemented")

 funasr/bin/asr_inference_uniasr.py

@@ -379,6 +379,8 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if word_lm_train_config is not None:

 funasr/bin/diar_inference_launch.py

@@ -2,8 +2,6 @@
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
#  MIT License  (https://opensource.org/licenses/MIT)

import torch
torch.set_num_threads(1)

import argparse
import logging

 funasr/bin/eend_ola_inference.py

@@ -158,6 +158,8 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if ngpu > 1:

 funasr/bin/lm_inference.py

@@ -89,10 +89,9 @@
    **kwargs,
):
    assert check_argument_types()
    logging.basicConfig(
        level=log_level,
        format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
    )
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)


    if ngpu >= 1 and torch.cuda.is_available():
        device = "cuda"

 funasr/bin/lm_inference_launch.py

@@ -1,9 +1,6 @@
#!/usr/bin/env python3
# Copyright ESPnet (https://github.com/espnet/espnet). All Rights Reserved.
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

import torch
torch.set_num_threads(1)


import argparse
import logging

 funasr/bin/punc_inference_launch.py

@@ -1,9 +1,5 @@
#!/usr/bin/env python3
# Copyright ESPnet (https://github.com/espnet/espnet). All Rights Reserved.
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

import torch
torch.set_num_threads(1)

import argparse
import logging

 funasr/bin/punctuation_infer_vadrealtime.py

@@ -203,10 +203,8 @@
    **kwargs,
):
    assert check_argument_types()
    logging.basicConfig(
        level=log_level,
        format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
    )
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)

    if ngpu >= 1 and torch.cuda.is_available():
        device = "cuda"

 funasr/bin/sond_inference.py

@@ -252,6 +252,8 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if ngpu > 1:

 funasr/bin/sv_inference.py

@@ -179,6 +179,9 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if ngpu > 1:

 funasr/bin/sv_inference_launch.py

@@ -2,8 +2,6 @@
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
#  MIT License  (https://opensource.org/licenses/MIT)

import torch
torch.set_num_threads(1)

import argparse
import logging

 funasr/bin/tp_inference.py

@@ -54,7 +54,7 @@
        assert check_argument_types()
        # 1. Build ASR model
        tp_model, tp_train_args = ASRTask.build_model_from_file(
            timestamp_infer_config, timestamp_model_file, device
            timestamp_infer_config, timestamp_model_file, device=device
        )
        if 'cuda' in device:
            tp_model = tp_model.cuda()  # force model to cuda
@@ -179,6 +179,9 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if ngpu > 1:

 funasr/bin/tp_inference_launch.py

@@ -1,9 +1,5 @@
#!/usr/bin/env python3
# Copyright ESPnet (https://github.com/espnet/espnet). All Rights Reserved.
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

import torch
torch.set_num_threads(1)

import argparse
import logging

 funasr/bin/vad_inference.py

@@ -192,6 +192,9 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if ngpu > 1:

 funasr/bin/vad_inference_launch.py

@@ -1,9 +1,4 @@
#!/usr/bin/env python3
# Copyright ESPnet (https://github.com/espnet/espnet). All Rights Reserved.
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

import torch
torch.set_num_threads(1)

import argparse
import logging

 funasr/bin/vad_inference_online.py

@@ -151,6 +151,9 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)
    
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")
    if ngpu > 1:

 funasr/models/joint_net/__init__.py

New file
@@ -0,0 +1 @@


 funasr/runtime/python/benchmark_libtorch.md

@@ -1,27 +1,32 @@
# Benchmark 
# CPU Benchmark (Libtorch)

## Configuration
### Data set:
Aishell1 [test set](https://www.openslr.org/33/) , the total audio duration is 36108.919 seconds.

### Tools
- Install ModelScope and FunASR

#### Install Requirements
Install ModelScope and FunASR
    ```shell
    pip install "modelscope[audio_asr]" --upgrade -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html
pip install -U modelscope funasr
# For the users in China, you could install with the command:
#pip install -U funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple
```

Install requirements
```shell
    git clone https://github.com/alibaba-damo-academy/FunASR.git && cd FunASR
    pip install --editable ./
    cd funasr/runtime/python/utils
    pip install -r requirements.txt
    ```

- recipe
#### Recipe

    set the model, data path and output_dir

    ```shell
    nohup bash test_rtf.sh &> log.txt &
    ```



## [Paraformer-large](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) 

 funasr/runtime/python/benchmark_onnx.md

@@ -1,20 +1,26 @@
# Benchmark 
# CPU Benchmark (ONNX)

## Configuration
### Data set:
Aishell1 [test set](https://www.openslr.org/33/) , the total audio duration is 36108.919 seconds.

### Tools
- Install ModelScope and FunASR

#### Install Requirements
Install ModelScope and FunASR
    ```shell
    pip install "modelscope[audio_asr]" --upgrade -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html
pip install -U modelscope funasr
# For the users in China, you could install with the command:
#pip install -U funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple
```

Install requirements
```shell
    git clone https://github.com/alibaba-damo-academy/FunASR.git && cd FunASR
    pip install --editable ./
    cd funasr/runtime/python/utils
    pip install -r requirements.txt
    ```

- recipe
#### Recipe

    set the model, data path and output_dir


 funasr/runtime/python/onnxruntime/README.md

@@ -35,22 +35,114 @@
# pip install -e ./ -i https://mirror.sjtu.edu.cn/pypi/web/simple
```

## Run the demo
- Model_dir: the model path, which contains `model.onnx`, `config.yaml`, `am.mvn`.
- Input: wav formt file, support formats: `str, np.ndarray, List[str]`
- Output: `List[str]`: recognition result.
- Example:
## Inference with runtime

### Speech Recognition
#### Paraformer
     ```python
     from funasr_onnx import Paraformer

     model_dir = "/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
 model_dir = "./export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
     model = Paraformer(model_dir, batch_size=1)

     wav_path = ['/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/example/asr_example.wav']
 wav_path = ['./export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/example/asr_example.wav']

     result = model(wav_path)
     print(result)
     ```
- Model_dir: the model path, which contains `model.onnx`, `config.yaml`, `am.mvn`
- Input: wav formt file, support formats: `str, np.ndarray, List[str]`
- Output: `List[str]`: recognition result

#### Paraformer-online

### Voice Activity Detection
#### FSMN-VAD
```python
from funasr_onnx import Fsmn_vad

model_dir = "./export/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch"
wav_path = "./export/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/example/vad_example.wav"
model = Fsmn_vad(model_dir)

result = model(wav_path)
print(result)
```
- Model_dir: the model path, which contains `model.onnx`, `config.yaml`, `am.mvn`
- Input: wav formt file, support formats: `str, np.ndarray, List[str]`
- Output: `List[str]`: recognition result

#### FSMN-VAD-online
```python
from funasr_onnx import Fsmn_vad_online
import soundfile


model_dir = "./export/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch"
wav_path = "./export/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/example/vad_example.wav"
model = Fsmn_vad_online(model_dir)


##online vad
speech, sample_rate = soundfile.read(wav_path)
speech_length = speech.shape[0]
#
sample_offset = 0
step = 1600
param_dict = {'in_cache': []}
for sample_offset in range(0, speech_length, min(step, speech_length - sample_offset)):
    if sample_offset + step >= speech_length - 1:
        step = speech_length - sample_offset
        is_final = True
    else:
        is_final = False
    param_dict['is_final'] = is_final
    segments_result = model(audio_in=speech[sample_offset: sample_offset + step],
                            param_dict=param_dict)
    if segments_result:
        print(segments_result)
```
- Model_dir: the model path, which contains `model.onnx`, `config.yaml`, `am.mvn`
- Input: wav formt file, support formats: `str, np.ndarray, List[str]`
- Output: `List[str]`: recognition result

### Punctuation Restoration
#### CT-Transformer
```python
from funasr_onnx import CT_Transformer

model_dir = "./export/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch"
model = CT_Transformer(model_dir)

text_in="跨境河流是养育沿岸人民的生命之源长期以来为帮助下游地区防灾减灾中方技术人员在上游地区极为恶劣的自然条件下克服巨大困难甚至冒着生命危险向印方提供汛期水文资料处理紧急事件中方重视印方在跨境河流问题上的关切愿意进一步完善双方联合工作机制凡是中方能做的我们都会去做而且会做得更好我请印度朋友们放心中国在上游的任何开发利用都会经过科学规划和论证兼顾上下游的利益"
result = model(text_in)
print(result[0])
```
- Model_dir: the model path, which contains `model.onnx`, `config.yaml`, `am.mvn`
- Input: wav formt file, support formats: `str, np.ndarray, List[str]`
- Output: `List[str]`: recognition result

#### CT-Transformer-online
```python
from funasr_onnx import CT_Transformer_VadRealtime

model_dir = "./export/damo/punc_ct-transformer_zh-cn-common-vad_realtime-vocab272727"
model = CT_Transformer_VadRealtime(model_dir)

text_in  = "跨境河流是养育沿岸|人民的生命之源长期以来为帮助下游地区防灾减灾中方技术人员|在上游地区极为恶劣的自然条件下克服巨大困难甚至冒着生命危险|向印方提供汛期水文资料处理紧急事件中方重视印方在跨境河流>问题上的关切|愿意进一步完善双方联合工作机制|凡是|中方能做的我们|都会去做而且会做得更好我请印度朋友们放心中国在上游的|任何开发利用都会经过科学|规划和论证兼顾上下游的利益"

vads = text_in.split("|")
rec_result_all=""
param_dict = {"cache": []}
for vad in vads:
    result = model(vad, param_dict=param_dict)
    rec_result_all += result[0]

print(rec_result_all)
```
- Model_dir: the model path, which contains `model.onnx`, `config.yaml`, `am.mvn`
- Input: wav formt file, support formats: `str, np.ndarray, List[str]`
- Output: `List[str]`: recognition result

## Performance benchmark


 funasr/train/trainer.py

@@ -187,9 +187,6 @@
                trainer_options.keep_nbest_models = [1]
            keep_nbest_models = trainer_options.keep_nbest_models
     
        #assert batch_interval is set and >0
        assert trainer_options.batch_interval > 0
 
        output_dir = Path(trainer_options.output_dir)
        reporter = Reporter()
        if trainer_options.use_amp: