docs

docs/modescope_pipeline/asr_pipeline.md

@@ -82,7 +82,7 @@ Undo
 - `output_dir`: None (Defalut), the output path of results if set
 
 ### Inference with multi-thread CPUs or multi GPUs
-FunASR also offer recipes [run.sh](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.sh) to decode 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
@@ -123,7 +123,7 @@ If you decode the SpeechIO test sets, you can use textnorm with `stage`=3, and `
 ## Finetune with pipeline
 
 ### Quick start
-[finetune.py](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/finetune.py)
+[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
@@ -166,7 +166,7 @@ python finetune.py &> log.txt &
 
 ### Finetune with your data
 
-- Modify finetune training related parameters in [finetune.py](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/finetune.py)
+- 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`
@@ -183,7 +183,7 @@ If you want finetune with multi-GPUs, you could:
 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`
+- 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

docs/modescope_pipeline/vad_pipeline.md

@@ -66,7 +66,7 @@ Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/
 - `output_dir`: None (Defalut), the output path of results if set
 
 ### Inference with multi-thread CPUs or multi GPUs
-FunASR also offer recipes [run.sh](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.sh) to decode 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

egs_modelscope/asr/TEMPLATE/finetune.py

@@ -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

@@ -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

@@ -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

@@ -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

@@ -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 modelscope.utils.constant import Tasks
 
 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