run.sh

#!/bin/bash

set -e

#===== begin config =======

stage=0
fix_scp=false
show_vpc_scores=true
skip_stage=

anon_exp_parameter="x_vector_vpc__crossgender=false__f0transformation=false__diffpseudospeaker"
verbose=""

########
# Once trained the rotation is applied on the anon-trial set to revert the VoicePrivacy anonymization,
# Top1 accuracy is calculated between de-anon-trial and trial-anon to asses the reversibility of procrustes (in the x-vector domain)
# EER / linkability is calculated between de-anon-trial and original enroll
########

# anon xvector extracted with anon model
retrained_anon_xtractor=false

# procrustes trained on ("f", "m" or "")
filter_gender=""

# train rotation with wasserstein procrustes instead of procrustes
wp=false

# Frontend params for wasserstein procrustes
frontend_train=""
# frontend_train="--pca --pca_n_dim 70"
wass_procrustes_param="--niter 512 --bsz 8 --lr 10"  # Hyperparameter found with grid search

# oracle:
oracle_f=false
oracle_m=false


#=====  end config  =======
. utils/parse_options.sh || exit 1;
. ./env.sh


mkdir -p exp

frontend_test="$frontend_train --pca_load_path exp/enroll_train_wp"

anon_xtractor="_retrained_xtractor"
if ! $retrained_anon_xtractor; then
  anon_xtractor=""
fi

wp_flag=""
if $wp; then
  wp_flag="--wp"
fi

if [ $stage -le -1 ]; then
  printf "${GREEN}Stage -1: testing x-vector loader${NC}\n"

  for anon_exp_parameter in \
  "x_vector_vpc__crossgender=false__f0transformation=false__diffpseudospeaker" \
  "x_vector_vpc__crossgender=false__f0transformation=false__diffpseudospeaker_retrained_xtractor" \
  "x_vector_vpc__crossgender=false__f0transformation=true__diffpseudospeaker" \
  "x_vector_vpc__crossgender=false__f0transformation=true__diffpseudospeaker_retrained_xtractor" \
  "x_vector_vpc__crossgender=true__f0transformation=false__diffpseudospeaker" \
  "x_vector_vpc__crossgender=true__f0transformation=false__diffpseudospeaker_retrained_xtractor" \
  "x_vector_vpc__crossgender=true__f0transformation=true__diffpseudospeaker" \
  "x_vector_vpc__crossgender=true__f0transformation=true__diffpseudospeaker_retrained_xtractor" \
  ;do

    printf "$anon_exp_parameter\n"

    if $fix_scp; then
      cd "data/$anon_exp_parameter"
      rg "exp/models/asv_eval[^/]*/xvect_01709_1" --files-with-matches | \
        xargs sed -i "s|exp/models/asv_eval[^/]*/xvect_01709_1|data/$anon_exp_parameter|g" || true
      cd -
    fi

    for suffix in "" "_anon"; do
      for original_dset in xvect_libri_test_trials_f xvect_libri_test_trials_m xvect_libri_test_enrolls; do
        original_dset=${original_dset}${suffix}
        printf "  $original_dset\n"
        python ./prep_dset.py \
          ./data/$anon_exp_parameter/$original_dset/xvector.scp \
          ./data/$anon_exp_parameter/$original_dset/xvector.scp  \
          "/tmp/test/Emb_U" "/tmp/test/User_U" \
          "/tmp/test/Emb_L" "/tmp/test/User_L" \
          --test
      done
    done
  done
  printf "${GREEN}Stage -1: All Scp files could be read${NC}\n"
  exit 0
fi

slug=original
if [ $stage -le 0 ]; then
  # quite useful step to test if the compute_spk_cosine script works
  printf "${GREEN}Reproduce VoicePrivacy EER results with cosine scoring${NC}\n"
  index=0
  for exp in "$anon_exp_parameter" \
              "$anon_exp_parameter${anon_xtractor}"; do
  if [[ $index == 1 ]]; then
    slug=anon
    if ! $retrained_anon_xtractor; then
      break
    fi

    printf "${GREEN}Anonymized x-vector -> (extracted by a x-vector trained on anonymized speech)${NC}\n"

  else
    for dset in "f" "m";do
      printf "**ASV ($slug): ${RED}test_trials_${dset} ${GREEN}original${NC} <=> ${RED}test_enrolls - ${GREEN}original${RED}${NC}**\n"
      python compute_spk_cosine.py \
        ./data/${exp}/xvect_libri_test_trials_${dset}/meta/trials \
        ./data/${exp}/xvect_libri_test_trials_${dset}/ \
        ./data/${exp}/xvect_libri_test_enrolls/ \
       ./exp/cosine_scores.txt

      PYTHONPATH=$(realpath ./utils/anonymization_metrics) \
        python ./utils/compute_metrics.py \
        -k ./data/${exp}/xvect_libri_test_trials_${dset}/meta/trials \
        -s ./exp/cosine_scores.txt
    done
    for dset in "f" "m";do
      printf "**ASV ($slug): ${RED}test_trials_${dset} ${GREEN}anonymized${NC} <=> ${RED}test_enrolls - ${GREEN}original${RED}${NC}**\n"
      python compute_spk_cosine.py \
        ./data/${exp}/xvect_libri_test_trials_${dset}/meta/trials \
        ./data/${exp}/xvect_libri_test_trials_${dset}_anon/ \
        ./data/${exp}/xvect_libri_test_enrolls/ \
       ./exp/cosine_scores.txt

      PYTHONPATH=$(realpath ./utils/anonymization_metrics) \
        python ./utils/compute_metrics.py \
        -k ./data/${exp}/xvect_libri_test_trials_${dset}/meta/trials \
        -s ./exp/cosine_scores.txt
    done
  index=1
  fi

    for dset in "f" "m";do
      printf "**ASV ($slug): ${RED}test_trials_${dset} ${GREEN}anonymized${NC} <=> ${RED}test_enrolls - ${GREEN}anonymized${RED}${NC}**\n"
      python compute_spk_cosine.py \
        ./data/${exp}/xvect_libri_test_trials_${dset}/meta/trials \
        ./data/${exp}/xvect_libri_test_trials_${dset}_anon/ \
        ./data/${exp}/xvect_libri_test_enrolls_anon/ \
       ./exp/cosine_scores.txt

      PYTHONPATH=$(realpath ./utils/anonymization_metrics) \
        python ./utils/compute_metrics.py \
        -k ./data/${exp}/xvect_libri_test_trials_${dset}/meta/trials \
        -s ./exp/cosine_scores.txt
    done
    printf "\n"
  done

  if $show_vpc_scores; then
    anon_dset=xvect_libri_test_enrolls_anon
    original_dset=xvect_libri_test_enrolls
    printf "${RED}Spk verif scores:${NC}\n"
    cat ./data/$anon_exp_parameter/results/results.txt | grep ".*$(echo $anon_dset | sed -e 's/xvect_//').*" -A 3
    printf "${RED}with retrained x-vector:${NC}\n"
    cat ./data/${anon_exp_parameter}${anon_xtractor}/results/results.txt | grep ".*$(echo $anon_dset | sed -e 's/xvect_//').*" -A 3
    printf "${RED}---${NC}\n"
  fi

fi

if [ $stage -le 1 ] && ! echo $skip_stage | grep -w -q 1; then
  anon_dset=xvect_libri_test_enrolls_anon
  original_dset=xvect_libri_test_enrolls

  # Calculate theorical best likability after training procrustes on test datatest
  if $oracle_f; then
    filter_gender="f"
    anon_dset=xvect_libri_test_trials_f_anon
    original_dset=xvect_libri_test_trials_f
  fi

  # Calculate theorical best likability after training procrustes on test datatest
  if $oracle_m; then
    filter_gender="m"
    anon_dset=xvect_libri_test_trials_m_anon
    original_dset=xvect_libri_test_trials_m
  fi

  printf "${GREEN}   == Data used to train rotation ==\n     - $original_dset \n     - $anon_dset${NC}\n"

  expe_dir=exp/enroll_train_wp
  mkdir -p $expe_dir

  python ./prep_dset.py \
     ./data/$anon_exp_parameter/$original_dset/xvector.scp \
     ./data/${anon_exp_parameter}${anon_xtractor}/$anon_dset/xvector.scp \
     "$expe_dir/Emb_U" "$expe_dir/User_U" \
     "$expe_dir/Emb_L" "$expe_dir/User_L" \
     --noplot \
     --filter_gender $filter_gender \
     # --spk_utt_all_combinations

  printf "${GREEN}== Training rotation ==${NC}\n"

  expe_dir=exp/enroll_train_wp

  python ./get_align_procrustes.py \
    --emb_src $expe_dir/Emb_U.npy \
    --label_src $expe_dir/User_U.npy \
    --emb_tgt $expe_dir/Emb_L.npy \
    --label_tgt $expe_dir/User_L.npy \
    --rotation exp/WP_R.npy \
    $frontend_train $wass_procrustes_param $wp_flag $verbose

  printf "${GREEN}Done${NC}\n"
fi

if [ $stage -le 2 ] && ! echo $skip_stage | grep -w -q 2; then
  printf "${GREEN}== TEST rotation irreversibility ==${NC}\n"

  expe_dir=exp/trials_test
  mkdir -p $expe_dir

  for dset in trials_f trials_m; do

    gender_test=${dset: -1}
    if [[ "${gender_test}" != "$filter_gender" ]] && [[ "${filter_gender}" != "" ]]; then
      continue
    fi

    original_dset=xvect_libri_test_${dset}
    anon_dset=xvect_libri_test_${dset}_anon


    printf "**Accuracy ${RED}$dset ${GREEN}anonymized => procrustes${NC} => ${RED}$dset - ${GREEN}original${RED}${NC}**\n"
    python ./prep_dset.py \
       ./data/$anon_exp_parameter/$original_dset/xvector.scp \
       ./data/${anon_exp_parameter}${anon_xtractor}/$anon_dset/xvector.scp \
       "$expe_dir/Emb_U" "$expe_dir/User_U" \
       "$expe_dir/Emb_L" "$expe_dir/User_L" \
       --noplot # --test-only-train-spk # --test-remove-train-spk

    python ./get_align_procrustes.py \
      --emb_src $expe_dir/Emb_U.npy \
      --label_src $expe_dir/User_U.npy \
      --emb_tgt $expe_dir/Emb_L.npy \
      --label_tgt $expe_dir/User_L.npy \
      --rotation exp/WP_R.npy \
      $frontend_test \
      --test

  done
fi


if [ $stage -le 3 ] && ! echo $skip_stage | grep -w -q 3; then
  printf "==${GREEN} TEST likability between Anonymized and Orignal speech ==\\n\
  Anonymized x-vector -> (extracted by a x-vector trained on anonymized speech)\\n\
  Original x-vector -> (extracted by a x-vector trained on anonymized speech) ${NC}\n"

  for dset in "f" "m";do
    if [[ "${dset}" != "$filter_gender" ]] && [[ "${filter_gender}" != "" ]]; then
      continue
    fi

    exp="$anon_exp_parameter${anon_xtractor}"
    python ./apply_procrustes.py \
      --emb_in ./data/${exp}/xvect_libri_test_trials_${dset}_anon/xvector.scp \
      --emb_out ./data/${exp}/xvect_libri_test_trials_${dset}_anon/ \
      --emb_src $expe_dir/nan --test \
      --rotation ./exp/WP_R.npy \
      $frontend_test
  done

  for dset in "f" "m";do
    if [[ "${dset}" != "$filter_gender" ]] && [[ "${filter_gender}" != "" ]]; then
      continue
    fi

    exp_o="$anon_exp_parameter"
    exp_a="$anon_exp_parameter${anon_xtractor}"

    printf "**ASV: ${RED}test_trials_${dset} ${GREEN}anonymized => procrustes${NC} <=> ${RED}test_enrolls - ${GREEN}original${RED}${NC}**\n"
    python compute_spk_cosine.py \
      ./data/${exp_a}/xvect_libri_test_trials_${dset}/meta/trials \
      ./data/${exp_a}/xvect_libri_test_trials_${dset}_anon/ \
      ./data/${exp_o}/xvect_libri_test_enrolls/ \
     ./exp/cosine_scores.txt \
     --trial-scp transformed_xvector.scp \
      $frontend_test --test

    PYTHONPATH=$(realpath ./utils/anonymization_metrics) \
      python ./utils/compute_metrics.py \
      -k ./data/${exp_a}/xvect_libri_test_trials_${dset}/meta/trials \
      -s ./exp/cosine_scores.txt
  done
fi