提交

kongdeqiang

5 天以前 28ccfbfc51068a663a80764e14074df5edf2b5ba

 examples/aishell/conformer/run.sh

@@ -5,7 +5,7 @@

# general configuration
feats_dir="../DATA" #feature output dictionary
exp_dir="."
exp_dir=`pwd`
lang=zh
token_type=char
stage=0
@@ -14,10 +14,10 @@
# feature configuration
nj=32

inference_device="cuda" #"cpu"
inference_checkpoint="model.pt"
inference_device="cuda" #"cpu", "cuda:0", "cuda:1"
inference_checkpoint="model.pt.avg10"
inference_scp="wav.scp"
inference_batch_size=32
inference_batch_size=1

# data
raw_data=../raw_data
@@ -26,6 +26,8 @@
# exp tag
tag="exp1"
workspace=`pwd`

master_port=12345

. utils/parse_options.sh || exit 1;

@@ -77,7 +79,7 @@
    --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

fi

token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
@@ -105,18 +107,16 @@
  echo "stage 4: ASR Training"

  mkdir -p ${exp_dir}/exp/${model_dir}
  log_file="${exp_dir}/exp/${model_dir}/train.log.txt"
  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}"

  gpu_num=$(echo CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
#  torchrun \
#  --nnodes 1 \
#  --nproc_per_node ${gpu_num}
  cmd="python"
  if [ ${gpu_num} -gt 1  ];then
    cmd="torchrun --nnodes 1 --nproc_per_node ${gpu_num}"
  fi
  ${cmd} \
  export CUDA_VISIBLE_DEVICES=$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}/conf" \
  --config-name "${config}" \
@@ -133,7 +133,7 @@
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
  echo "stage 5: Inference"

  if ${inference_device} == "cuda"; then
  if [ ${inference_device} == "cuda" ]; then
      nj=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
  else
      inference_batch_size=1
@@ -145,8 +145,9 @@

  for dset in ${test_sets}; do

    inference_dir="${exp_dir}/exp/${model_dir}/${inference_checkpoint}/${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/${dset}"
@@ -158,7 +159,7 @@
    done
    utils/split_scp.pl "${key_file}" ${split_scps}

    gpuid_list_array=(${gpuid_list//,/ })
    gpuid_list_array=(${CUDA_VISIBLE_DEVICES//,/ })
    for JOB in $(seq ${nj}); do
        {
          id=$((JOB-1))
@@ -174,7 +175,9 @@
          ++input="${_logdir}/keys.${JOB}.scp" \
          ++output_dir="${inference_dir}/${JOB}" \
          ++device="${inference_device}" \
          ++batch_size="${inference_batch_size}"
          ++ncpu=1 \
          ++disable_log=true \
          ++batch_size="${inference_batch_size}" &> ${_logdir}/log.${JOB}.txt
        }&

    done
@@ -190,8 +193,8 @@
    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