From 6400d078948b9284ccfee9802076172ca68a5585 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=8C=97=E5=BF=B5?= Date: Wed, 22 Mar 2023 10:52:03 +0800 Subject: [PATCH] update paraformer_large inference recipe --- .../README.md | 11 +- .../infer.py | 108 +++--------------- .../infer.sh | 93 +++++++++++++++ 3 files changed, 116 insertions(+), 96 deletions(-) create mode 100644 egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.sh diff --git a/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/README.md b/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/README.md index a0443614f..79cc3c3bf 100644 --- a/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/README.md +++ b/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/README.md @@ -21,23 +21,26 @@ Or you can use the finetuned model for inference directly. -- Setting parameters in `infer.py` +- Setting parameters in `infer.sh` - model: # model name on ModelScope - data_dir: # the dataset dir needs to include `test/wav.scp`. If `test/text` is also exists, CER will be computed - output_dir: # result dir - - ngpu: # the number of GPUs for decoding, if `ngpu` > 0, use GPU decoding - - njob: # the number of jobs for CPU decoding, if `ngpu` = 0, use CPU decoding, please set `njob` - batch_size: # batchsize of inference + - gpu_inference: # whether to perform gpu decoding, set false for cpu decoding + - gpuid_list: # set gpus, e.g., gpuid_list="0,1" + - njob: # 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 - python infer.py + sh infer.sh ``` - 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. + ### Inference using local finetuned model - Modify inference related parameters in `infer_after_finetune.py` diff --git a/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.py b/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.py index 795a1e7c5..19731912e 100644 --- a/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.py +++ b/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.py @@ -1,101 +1,25 @@ import os import shutil -from multiprocessing import Pool - +import argparse from modelscope.pipelines import pipeline from modelscope.utils.constant import Tasks -from funasr.utils.compute_wer import compute_wer - - -def modelscope_infer_core(output_dir, split_dir, njob, idx, batch_size, ngpu, model): - output_dir_job = os.path.join(output_dir, "output.{}".format(idx)) - if ngpu > 0: - use_gpu = 1 - gpu_id = int(idx) - 1 - else: - use_gpu = 0 - gpu_id = -1 - if "CUDA_VISIBLE_DEVICES" in os.environ.keys(): - gpu_list = os.environ['CUDA_VISIBLE_DEVICES'].split(",") - os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu_list[gpu_id]) - else: - os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu_id) - inference_pipline = pipeline( +def modelscope_infer(args): + os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpuid) + inference_pipeline = pipeline( task=Tasks.auto_speech_recognition, - model=model, - output_dir=output_dir_job, - batch_size=batch_size, - ngpu=use_gpu, + model=args.model, + output_dir=args.output_dir, + batch_size=args.batch_size, ) - audio_in = os.path.join(split_dir, "wav.{}.scp".format(idx)) - inference_pipline(audio_in=audio_in) - - -def modelscope_infer(params): - # prepare for multi-GPU decoding - ngpu = params["ngpu"] - njob = params["njob"] - batch_size = params["batch_size"] - output_dir = params["output_dir"] - model = params["model"] - if os.path.exists(output_dir): - shutil.rmtree(output_dir) - os.mkdir(output_dir) - split_dir = os.path.join(output_dir, "split") - os.mkdir(split_dir) - if ngpu > 0: - nj = ngpu - elif ngpu == 0: - nj = njob - wav_scp_file = os.path.join(params["data_dir"], "wav.scp") - with open(wav_scp_file) as f: - lines = f.readlines() - num_lines = len(lines) - num_job_lines = num_lines // nj - start = 0 - for i in range(nj): - end = start + num_job_lines - file = os.path.join(split_dir, "wav.{}.scp".format(str(i + 1))) - with open(file, "w") as f: - if i == nj - 1: - f.writelines(lines[start:]) - else: - f.writelines(lines[start:end]) - start = end - - p = Pool(nj) - for i in range(nj): - p.apply_async(modelscope_infer_core, - args=(output_dir, split_dir, njob, str(i + 1), batch_size, ngpu, model)) - p.close() - p.join() - - # combine decoding results - best_recog_path = os.path.join(output_dir, "1best_recog") - os.mkdir(best_recog_path) - files = ["text", "token", "score"] - for file in files: - with open(os.path.join(best_recog_path, file), "w") as f: - for i in range(nj): - job_file = os.path.join(output_dir, "output.{}/1best_recog".format(str(i + 1)), file) - with open(job_file) as f_job: - lines = f_job.readlines() - f.writelines(lines) - - # If text exists, compute CER - text_in = os.path.join(params["data_dir"], "text") - if os.path.exists(text_in): - text_proc_file = os.path.join(best_recog_path, "token") - compute_wer(text_in, text_proc_file, os.path.join(best_recog_path, "text.cer")) - + inference_pipeline(audio_in=args.audio_in) if __name__ == "__main__": - params = {} - params["model"] = "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" - params["data_dir"] = "./data/test" - params["output_dir"] = "./results" - params["ngpu"] = 1 # if ngpu > 0, will use gpu decoding - params["njob"] = 1 # if ngpu = 0, will use cpu decoding - params["batch_size"] = 64 - modelscope_infer(params) \ No newline at end of file + parser = argparse.ArgumentParser() + parser.add_argument('--model', type=str, default="speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch") + parser.add_argument('--audio_in', type=str, default="./data/test") + 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) \ No newline at end of file diff --git a/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.sh b/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.sh new file mode 100644 index 000000000..770cf97c7 --- /dev/null +++ b/egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/infer.sh @@ -0,0 +1,93 @@ +#!/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 +gpuid_list="0,1" +njob=4 +gpu_inference=true + +if ${gpu_inference}; then + nj=$(echo $gpuid_list | awk -F "," '{print NF}') +else + nj=$njob + 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 ..." + python utils/proce_text.py ${output_dir}/1best_recog/text ${output_dir}/1best_recog/text.proc + python utils/proce_text.py ${data_dir}/text ${data_dir}/text.proc + python utils/compute_wer.py ${data_dir}/text.proc ${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 \ + ${data_dir}/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 ${data_dir}/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 +