Merge pull request #807 from alibaba-damo-academy/dev_wjm

jmwang66

2023-08-07 cf8e000a84e888495dcf30c4dbfecea1ee7ab4e2

 egs/callhome/diarization/sond/finetune.sh

@@ -8,25 +8,46 @@
# [2] Speaker Overlap-aware Neural Diarization for Multi-party Meeting Analysis, EMNLP 2022
# We recommend you run this script stage by stage.

# This recipe includes:
# 1. downloading a pretrained model on the simulated data from switchboard and NIST,
# 2. finetuning the pretrained model on Callhome1.
# Finally, you will get a slightly better DER result 9.95% on Callhome2 than that in the paper 10.14%.

# environment configuration
kaldi_root=

if [ -z "${kaldi_root}" ]; then
  echo "We need kaldi to prepare dataset, extract fbank features, please install kaldi first and set kaldi_root."
  echo "Kaldi installation guide can be found at https://kaldi-asr.org/"
  exit;
fi

if [ ! -e local ]; then
  ln -s ${kaldi_root}/egs/callhome_diarization/v2/local ./local
fi

if [ ! -e utils ]; then
  ln -s ../../../aishell/transformer/utils ./utils
  ln -s ${kaldi_root}/egs/callhome_diarization/v2/utils ./utils
fi

# callhome data root like path/to/NIST/LDC2001S97
callhome_root=
if [ -z "${kaldi_root}" ]; then
  echo "We need callhome corpus to prepare data."
  exit;
fi

# machines configuration
gpu_devices="0,1,2,3"
gpu_devices="0,1,2,3"  # for V100-16G, need 4 gpus.
gpu_num=4
count=1

# general configuration
stage=1
stop_stage=1
stage=0
stop_stage=10
# number of jobs for data process
nj=16
sr=8000

# dataset related
data_root=

# experiment configuration
lang=en
@@ -73,15 +94,201 @@
    _ngpu=0
fi

# Download required resources
# Prepare datasets
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
  echo "Stage 0: Download required resources."
  wget told_finetune_resources.zip
  echo "Stage 0: Prepare callhome data."
  local/make_callhome.sh ${callhome_root} ${datadir}/

  # split ref.rttm
  for dset in callhome1 callhome2; do
    rm -rf ${datadir}/${dset}/ref.rttm
    for name in `awk '{print $1}' ${datadir}/${dset}/wav.scp`; do
      grep ${name} ${datadir}/callhome/fullref.rttm >> ${datadir}/${dset}/ref.rttm;
    done

    # filter out records which don't have rttm labels.
    awk '{print $2}' ${datadir}/${dset}/ref.rttm | sort | uniq > ${datadir}/${dset}/uttid
    mv ${datadir}/${dset}/wav.scp ${datadir}/${dset}/wav.scp.bak
    awk '{if (NR==FNR){a[$1]=1}else{if (a[$1]==1){print $0}}}' ${datadir}/${dset}/uttid ${datadir}/${dset}/wav.scp.bak > ${datadir}/${dset}/wav.scp
    mkdir ${datadir}/${dset}/raw
    mv ${datadir}/${dset}/{reco2num_spk,segments,spk2utt,utt2spk,uttid,wav.scp.bak} ${datadir}/${dset}/raw/
    awk '{print $1,$1}' ${datadir}/${dset}/wav.scp > ${datadir}/${dset}/utt2spk
  done
fi

# Finetune model on callhome1
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
  echo "Stage 1: Finetune pretrained model on callhome1."
  echo "Stage 1: Dump sph file to wav"
  export PATH=${kaldi_root}/tools/sph2pipe/:${PATH}
  if [ ! -f ${kaldi_root}/tools/sph2pipe/sph2pipe ]; then
    echo "Can not find sph2pipe in ${kaldi_root}/tools/sph2pipe/,"
    echo "please install sph2pipe and put it in the right place."
    exit;
  fi

  for dset in callhome1 callhome2; do
    echo "Stage 1: start to dump ${dset}."
    mv ${datadir}/${dset}/wav.scp ${datadir}/${dset}/sph.scp

    mkdir -p ${dumpdir}/${dset}/wavs
    python -Wignore script/dump_pipe_wav.py ${datadir}/${dset}/sph.scp ${dumpdir}/${dset}/wavs \
      --sr ${sr} --nj ${nj} --no_pbar
    find `pwd`/${dumpdir}/${dset}/wavs -iname "*.wav" | sort | awk -F'[/.]' '{print $(NF-1),$0}' > ${datadir}/${dset}/wav.scp
  done
fi

if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
  echo "Stage 2: Extract non-overlap segments from callhome dataset"
  for dset in callhome1 callhome2; do
    echo "Stage 2: Extracting non-overlap segments for "${dset}
    mkdir -p ${dumpdir}/${dset}/nonoverlap_0s
    python -Wignore script/extract_nonoverlap_segments.py \
      ${datadir}/${dset}/wav.scp ${datadir}/${dset}/ref.rttm ${dumpdir}/${dset}/nonoverlap_0s \
      --min_dur 0.1 --max_spk_num 8 --sr ${sr} --no_pbar --nj ${nj}

    mkdir -p ${datadir}/${dset}/nonoverlap_0s
    find ${dumpdir}/${dset}/nonoverlap_0s/ -iname "*.wav" | sort | awk -F'[/.]' '{print $(NF-1),$0}' > ${datadir}/${dset}/nonoverlap_0s/wav.scp
    awk -F'[/.]' '{print $(NF-1),$(NF-2)}' ${datadir}/${dset}/nonoverlap_0s/wav.scp > ${datadir}/${dset}/nonoverlap_0s/utt2spk
    echo "Done."
  done
fi

if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
  echo "Stage 3: Generate fbank features"
  home_path=$(pwd)
  cd ${kaldi_root}/egs/callhome_diarization/v2 || exit

  export train_cmd="run.pl"
  export cmd="run.pl"
  . ./path.sh
  cd $home_path || exit

  ln -s ${kaldi_root}/egs/callhome_diarization/v2/steps ./
  for dset in callhome1 callhome2; do
    utils/fix_data_dir.sh ${datadir}/${dset}
    steps/make_fbank.sh --write-utt2num-frames true --fbank-config conf/fbank.conf --nj ${nj} --cmd "$train_cmd" \
        ${datadir}/${dset} ${expdir}/make_fbank/${dset} ${dumpdir}/${dset}/fbank
  done
  rm -f steps

fi

if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
  echo "Stage 4: Extract speaker embeddings."
  sv_exp_dir=exp/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch

  if [ ! -e ${sv_exp_dir} ]; then
    echo "start to download sv models"
    git lfs install
    git clone https://www.modelscope.cn/damo/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch.git
    mv speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch ${expdir}/
    echo "Done."
  fi

  for dset in callhome1/nonoverlap_0s callhome2/nonoverlap_0s; do
    echo "Start to extract speaker embeddings for ${dset}"
    key_file=${datadir}/${dset}/wav.scp
    num_scp_file="$(<${key_file} wc -l)"
    _nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
    _logdir=${dumpdir}/${dset}/xvecs
    mkdir -p ${_logdir}
    split_scps=
    for n in $(seq "${_nj}"); do
        split_scps+=" ${_logdir}/keys.${n}.scp"
    done
    # shellcheck disable=SC2086
    utils/split_scp.pl "${key_file}" ${split_scps}

    ${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/sv_inference.JOB.log \
      python -m funasr.bin.sv_inference_launch \
        --batch_size 1 \
        --njob ${njob} \
        --ngpu "${_ngpu}" \
        --gpuid_list ${gpuid_list} \
        --data_path_and_name_and_type "${key_file},speech,sound" \
        --key_file "${_logdir}"/keys.JOB.scp \
        --sv_train_config ${sv_exp_dir}/sv.yaml \
        --sv_model_file ${sv_exp_dir}/sv.pth \
        --output_dir "${_logdir}"/output.JOB
    cat ${_logdir}/output.*/xvector.scp | sort > ${datadir}/${dset}/utt2xvec

    python script/calc_num_frames.py ${key_file} ${datadir}/${dset}/utt2num_frames
    echo "Done."
  done

fi

if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
  echo "Stage 5: Generate label files."

  for dset in callhome1 callhome2; do
    echo "Stage 5: Generate labels for ${dset}."
    python -Wignore script/calc_real_meeting_frame_labels.py \
          ${datadir}/${dset} ${dumpdir}/${dset}/labels \
          --n_spk 8 --frame_shift 0.01 --nj 16 --sr 8000
    find `pwd`/${dumpdir}/${dset}/labels/ -iname "*.lbl.mat" | awk -F'[/.]' '{print $(NF-2),$0}' | sort > ${datadir}/${dset}/labels.scp
  done

fi

if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
  echo "Stage 6: Make training and evaluation files."

  # dump callhome1 data in training mode.
  data_dir=${datadir}/callhome1/files_for_dump
  mkdir ${data_dir}
  # filter out zero duration segments
  LC_ALL=C awk '{if ($5 > 0){print $0}}' ${datadir}/callhome1/ref.rttm > ${data_dir}/ref.rttm
  cp ${datadir}/callhome1/{feats.scp,labels.scp} ${data_dir}/
  cp ${datadir}/callhome1/nonoverlap_0s/{utt2spk,utt2xvec,utt2num_frames} ${data_dir}/

  echo "Stage 6: start to dump for callhome1."
  python -Wignore script/dump_meeting_chunks.py --dir ${data_dir} \
    --out ${dumpdir}/callhome1/dumped_files/data --n_spk 16 --no_pbar --sr 8000 --mode train \
    --chunk_size 1600 --chunk_shift 400 --add_mid_to_speaker true

  mkdir -p ${datadir}/callhome1/dumped_files
  cat ${dumpdir}/callhome1/dumped_files/data_parts*_feat.scp | sort > ${datadir}/callhome1/dumped_files/feats.scp
  cat ${dumpdir}/callhome1/dumped_files/data_parts*_xvec.scp | sort > ${datadir}/callhome1/dumped_files/profile.scp
  cat ${dumpdir}/callhome1/dumped_files/data_parts*_label.scp | sort > ${datadir}/callhome1/dumped_files/label.scp
  mkdir -p ${expdir}/callhome1_states
  awk '{print $1,"1600"}' ${datadir}/callhome1/dumped_files/feats.scp | shuf > ${expdir}/callhome1_states/speech_shape
  python -Wignore script/convert_rttm_to_seg_file.py --rttm_scp ${data_dir}/ref.rttm --seg_file ${data_dir}/org_vad.txt

  # dump callhome2 data in test mode.
  data_dir=${datadir}/callhome2/files_for_dump
  mkdir ${data_dir}
  # filter out zero duration segments
  LC_ALL=C awk '{if ($5 > 0){print $0}}' ${datadir}/callhome2/ref.rttm > ${data_dir}/ref.rttm
  cp ${datadir}/callhome2/{feats.scp,labels.scp} ${data_dir}/
  cp ${datadir}/callhome2/nonoverlap_0s/{utt2spk,utt2xvec,utt2num_frames} ${data_dir}/

  echo "Stage 6: start to dump for callhome2."
  python -Wignore script/dump_meeting_chunks.py --dir ${data_dir} \
    --out ${dumpdir}/callhome2/dumped_files/data --n_spk 16 --no_pbar --sr 8000 --mode test \
    --chunk_size 1600 --chunk_shift 400 --add_mid_to_speaker true

  mkdir -p ${datadir}/callhome2/dumped_files
  cat ${dumpdir}/callhome2/dumped_files/data_parts*_feat.scp | sort > ${datadir}/callhome2/dumped_files/feats.scp
  cat ${dumpdir}/callhome2/dumped_files/data_parts*_xvec.scp | sort > ${datadir}/callhome2/dumped_files/profile.scp
  cat ${dumpdir}/callhome2/dumped_files/data_parts*_label.scp | sort > ${datadir}/callhome2/dumped_files/label.scp
  mkdir -p ${expdir}/callhome2_states
  awk '{print $1,"1600"}' ${datadir}/callhome2/dumped_files/feats.scp | shuf > ${expdir}/callhome2_states/speech_shape
  python -Wignore script/convert_rttm_to_seg_file.py --rttm_scp ${data_dir}/ref.rttm --seg_file ${data_dir}/org_vad.txt

fi

# Finetune model on callhome1, this will take about 1.5 hours.
if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
  echo "Stage 7: Finetune pretrained model on callhome1."

  if [ ! -e ${expdir}/speech_diarization_sond-en-us-swbd_sre-8k-n16k4-pytorch ]; then
    echo "start to download pretrained models"
    git lfs install
    git clone https://www.modelscope.cn/damo/speech_diarization_sond-en-us-swbd_sre-8k-n16k4-pytorch.git
    mv speech_diarization_sond-en-us-swbd_sre-8k-n16k4-pytorch ${expdir}/
    echo "Done."
  fi

  world_size=$gpu_num  # run on one machine
  mkdir -p ${expdir}/${model_dir}
  mkdir -p ${expdir}/${model_dir}/log
@@ -122,7 +329,7 @@
              --valid_data_path_and_name_and_type ${datadir}/${valid_set}/dumped_files/profile.scp,profile,kaldi_ark \
              --valid_data_path_and_name_and_type ${datadir}/${valid_set}/dumped_files/label.scp,binary_labels,kaldi_ark \
              --valid_shape_file ${expdir}/${valid_set}_states/speech_shape \
              --init_param exp/pretrained_models/phase2.pth \
              --init_param ${expdir}/speech_diarization_sond-en-us-swbd_sre-8k-n16k4-pytorch/sond.pth \
              --unused_parameters true \
              ${init_opt} \
              ${freeze_opt} \
@@ -145,8 +352,8 @@


# evaluate for finetuned model
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
    echo "stage 2: evaluation for finetuned model ${inference_model}."
if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
    echo "stage 8: evaluation for finetuned model ${inference_model}."
    for dset in ${test_sets}; do
        echo "Processing for $dset"
        exp_model_dir=${expdir}/${model_dir}
@@ -191,16 +398,16 @@

# Scoring for finetuned model, you may get a DER like:
# oracle_vad  |  system_vad
#   7.28      |     8.06
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
  echo "stage 3: Scoring finetuned models"
#   7.32      |     8.14
if [ ${stage} -le 9 ] && [ ${stop_stage} -ge 9 ]; then
  echo "stage 9: Scoring finetuned models"
  if [ ! -e dscore ]; then
    git clone https://github.com/nryant/dscore.git
    pip install intervaltree
    # add intervaltree to setup.py
  fi
  for dset in ${test_sets}; do
    echo "stage 3: Scoring for ${dset}"
    echo "stage 9: Scoring for ${dset}"
    diar_exp=${expdir}/${model_dir}
    _data="${datadir}/${dset}"
    _inference_tag="$(basename "${inference_config}" .yaml)${inference_tag}"
@@ -230,12 +437,12 @@
# Then find the wav files to construct wav.scp and put it at data/callhome2/wav.scp.
# After iteratively perform SOAP, you will get DER results like:
# iters : oracle_vad  |  system_vad
# iter_0:   9.68      |     10.51
# iter_1:   9.26      |     10.14  (reported in the paper)
# iter_2:   9.18      |     10.08
# iter_3:   9.24      |     10.15
# iter_4:   9.27      |     10.17
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# iter_0:   9.58      |     10.46
# iter_1:   9.22      |     10.15
# iter_2:   9.21      |     10.14
# iter_3:   9.30      |     10.24
# iter_4:   9.29      |     10.23
if [ ${stage} -le 10 ] && [ ${stop_stage} -ge 10 ]; then
  if [ ! -e ${expdir}/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch ]; then
    git lfs install
    git clone https://www.modelscope.cn/damo/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch.git
@@ -243,7 +450,7 @@
  fi

  for dset in ${test_sets}; do
    echo "stage 4: Evaluating finetuned system on ${dset} set with medfilter_size=83 clustering=EEND-OLA"
    echo "stage 10: Evaluating finetuned system on ${dset} set with medfilter_size=83 clustering=EEND-OLA"
    sv_exp_dir=${expdir}/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch
    diar_exp=${expdir}/${model_dir}
    _data="${datadir}/${dset}/dumped_files"