提交

kongdeqiang

2026-03-13 28ccfbfc51068a663a80764e14074df5edf2b5ba

 examples/aishell/paraformer/run.sh

@@ -1,24 +1,23 @@
#!/usr/bin/env bash

workspace=`pwd`

# machines configuration
CUDA_VISIBLE_DEVICES="0,1"
gpu_num=2
gpu_inference=true  # Whether to perform gpu decoding, set false for cpu decoding
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
njob=1

# general configuration
feats_dir="../DATA" #feature output dictionary
exp_dir="."
exp_dir=`pwd`
lang=zh
token_type=char
stage=0
stop_stage=5

# feature configuration
nj=64
nj=32

inference_device="cuda" #"cpu"
inference_checkpoint="model.pt.avg10"
inference_scp="wav.scp"
inference_batch_size=32

# data
raw_data=../raw_data
@@ -26,6 +25,9 @@

# exp tag
tag="exp1"
workspace=`pwd`

master_port=12345

. utils/parse_options.sh || exit 1;

@@ -39,17 +41,13 @@
valid_set=dev
test_sets="dev test"

config=train_asr_paraformer_conformer_12e_6d_2048_256.yaml
config=paraformer_conformer_12e_6d_2048_256.yaml
model_dir="baseline_$(basename "${config}" .yaml)_${lang}_${token_type}_${tag}"

inference_device="cuda" #"cpu"
inference_checkpoint="model.pt"
inference_scp="wav.scp"



if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
    echo "stage -1: Data Download"
    mkdir -p ${raw_data}
    local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
    local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
fi
@@ -76,13 +74,11 @@

if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
    echo "stage 1: Feature and CMVN Generation"
#    utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$config" --scale 1.0
    python ../../../funasr/bin/compute_audio_cmvn.py \
    --config-path "${workspace}" \
    --config-path "${workspace}/conf" \
    --config-name "${config}" \
    ++train_data_set_list="${feats_dir}/data/${train_set}/audio_datasets.jsonl" \
    ++cmvn_file="${feats_dir}/data/${train_set}/cmvn.json" \
    ++dataset_conf.num_workers=$nj
    ++cmvn_file="${feats_dir}/data/${train_set}/cmvn.json"
fi

token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
@@ -109,15 +105,22 @@
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
  echo "stage 4: ASR Training"

  log_file="${exp_dir}/exp/${model_dir}/train.log.txt"
  mkdir -p ${exp_dir}/exp/${model_dir}
  current_time=$(date "+%Y-%m-%d_%H-%M")
  log_file="${exp_dir}/exp/${model_dir}/train.log.txt.${current_time}"
  echo "log_file: ${log_file}"

  export CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES
  gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
  torchrun \
  --nnodes 1 \
  --nproc_per_node ${gpu_num} \
  --master_port ${master_port} \
  ../../../funasr/bin/train.py \
  --config-path "${workspace}" \
  --config-path "${workspace}/conf" \
  --config-name "${config}" \
  ++train_data_set_list="${feats_dir}/data/${train_set}/audio_datasets.jsonl" \
  ++valid_data_set_list="${feats_dir}/data/${valid_set}/audio_datasets.jsonl" \
  ++tokenizer_conf.token_list="${token_list}" \
  ++frontend_conf.cmvn_file="${feats_dir}/data/${train_set}/am.mvn" \
  ++output_dir="${exp_dir}/exp/${model_dir}" &> ${log_file}
@@ -129,21 +132,21 @@
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
  echo "stage 5: Inference"

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

  for dset in ${test_sets}; do

    inference_dir="${asr_exp}/${inference_checkpoint}/${dset}"
    inference_dir="${exp_dir}/exp/${model_dir}/inference-${inference_checkpoint}/${dset}"
    _logdir="${inference_dir}/logdir"
    echo "inference_dir: ${inference_dir}"

    mkdir -p "${_logdir}"
    data_dir="${feats_dir}/data/${dset}"
@@ -155,8 +158,13 @@
    done
    utils/split_scp.pl "${key_file}" ${split_scps}

    gpuid_list_array=(${CUDA_VISIBLE_DEVICES//,/ })
    for JOB in $(seq ${nj}); do
        {
          id=$((JOB-1))
          gpuid=${gpuid_list_array[$id]}

          export CUDA_VISIBLE_DEVICES=${gpuid}
          python ../../../funasr/bin/inference.py \
          --config-path="${exp_dir}/exp/${model_dir}" \
          --config-name="config.yaml" \
@@ -165,7 +173,10 @@
          ++frontend_conf.cmvn_file="${feats_dir}/data/${train_set}/am.mvn" \
          ++input="${_logdir}/keys.${JOB}.scp" \
          ++output_dir="${inference_dir}/${JOB}" \
          ++device="${inference_device}"
          ++device="${inference_device}" \
          ++ncpu=1 \
          ++disable_log=true \
          ++batch_size="${inference_batch_size}" &> ${_logdir}/log.${JOB}.txt
        }&

    done
@@ -181,10 +192,10 @@
    done

    echo "Computing WER ..."
    cp ${inference_dir}/1best_recog/text ${inference_dir}/1best_recog/text.proc
    cp ${data_dir}/text ${inference_dir}/1best_recog/text.ref
    python utils/postprocess_text_zh.py ${inference_dir}/1best_recog/text ${inference_dir}/1best_recog/text.proc
    python utils/postprocess_text_zh.py  ${data_dir}/text ${inference_dir}/1best_recog/text.ref
    python utils/compute_wer.py ${inference_dir}/1best_recog/text.ref ${inference_dir}/1best_recog/text.proc ${inference_dir}/1best_recog/text.cer
    tail -n 3 ${inference_dir}/1best_recog/text.cer
  done

fi
fi