#!/bin/bash
  
  # _1b is as _1a, but with pitch feats, i-vector and dropout schedule added, referenced from wsj
  
  # basic info:
  # steps/info/chain_dir_info.pl exp/chain/tdnn_1f_nopitch_ivec_sp/exp/chain/tdnn_1f_nopitch_ivec_sp/: num-iters=578 nj=2..8 num-params=19.3M dim=43+100->4520 combine=-0.082->-0.081 (over 6) xent:train/valid[384,577,final]=(-0.863,-0.752,-0.740/-0.901,-0.791,-0.784) logprob:train/valid[384,577,final]=(-0.083,-0.076,-0.075/-0.084,-0.077,-0.076)
  
  # results:
  # local/chain/compare_wer.sh exp/chain/tdnn_1f_nopitch_ivec_sp/
  # Model                tdnn_1f_nopitch_ivec_sp
  # Num. of params             19.3M
  # WER(%)                     8.81
  # Final train prob        -0.0749
  # Final valid prob        -0.0756
  # Final train prob (xent)   -0.7401
  # Final valid prob (xent)   -0.7837
  
  set -e
  
  # configs for 'chain'
  affix=all
  stage=0
  train_stage=-10
  get_egs_stage=-10
  dir=exp/chain/tdnn_1b  # Note: _sp will get added to this
  decode_iter=
  
  # training options
  num_epochs=4
  initial_effective_lrate=0.001
  final_effective_lrate=0.0001
  max_param_change=2.0
  final_layer_normalize_target=0.5
  num_jobs_initial=2
  num_jobs_final=4
  nj=15
  minibatch_size=128
  dropout_schedule='0,0@0.20,0.3@0.50,0'
  frames_per_eg=150,110,90
  remove_egs=true
  common_egs_dir=
  xent_regularize=0.1
  
  # End configuration section.
  echo "$0 $@"  # Print the command line for logging
  
  . ./cmd.sh
  . ./path.sh
  . ./utils/parse_options.sh
  
  if ! cuda-compiled; then
    cat <<EOF && exit 1
  This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
  If you want to use GPUs (and have them), go to src/, and configure and make on a machine
  where "nvcc" is installed.
  EOF
  fi
  
  dir=${dir}${affix:+_$affix}_sp
  train_set=train
  test_sets="dev test"
  ali_dir=exp/tri3_ali
  treedir=exp/chain/tri4_cd_tree_sp
  lang=data/lang_chain
  
  if [ $stage -le 5 ]; then
    mfccdir=mfcc_hires
    for datadir in ${train_set} ${test_sets}; do
    	utils/copy_data_dir.sh data/${datadir} data/${datadir}_hires
      utils/data/perturb_data_dir_volume.sh data/${datadir}_hires || exit 1;
  	steps/make_mfcc_pitch.sh --mfcc-config conf/mfcc_hires.conf --pitch-config conf/pitch.conf \
        --nj $nj data/${datadir}_hires exp/make_mfcc/ ${mfccdir}
      steps/compute_cmvn_stats.sh data/${datadir}_hires exp/make_mfcc ${mfccdir}
      utils/data/limit_feature_dim.sh 0:39 data/${datadir}_hires data/${datadir}_hires_nopitch
      steps/compute_cmvn_stats.sh data/${datadir}_hires_nopitch exp/make_mfcc ${mfccdir}
    done
  fi
  
  # extract ivector from unified data using the trained
  if [ $stage -le 6 ]; then
    echo "$0: computing a subset of data to train the diagonal UBM."
    # We'll use about a quarter of the data.
    mkdir -p exp/chain/diag_ubm_${affix}
    temp_data_root=exp/chain/diag_ubm_${affix}
  
    num_utts_total=$(wc -l < data/${train_set}_hires_nopitch/utt2spk)
    num_utts=$[$num_utts_total/4]
    utils/data/subset_data_dir.sh data/${train_set}_hires_nopitch \
      $num_utts ${temp_data_root}/${train_set}_subset
  
    echo "$0: computing a PCA transform from the hires data."
    steps/online/nnet2/get_pca_transform.sh --cmd "$train_cmd" \
      --splice-opts "--left-context=3 --right-context=3" \
      --max-utts 10000 --subsample 2 \
      --dim $(feat-to-dim scp:${temp_data_root}/${train_set}_subset/feats.scp -) \
      ${temp_data_root}/${train_set}_subset \
      exp/chain/pca_transform_${affix}
  
    echo "$0: training the diagonal UBM."
    # Use 512 Gaussians in the UBM.
    steps/online/nnet2/train_diag_ubm.sh --cmd "$train_cmd" --nj $nj \
      --num-frames 700000 \
      --num-threads 8 \
      ${temp_data_root}/${train_set}_subset 512 \
      exp/chain/pca_transform_${affix} exp/chain/diag_ubm_${affix}
  
    echo "$0: training the iVector extractor"
    steps/online/nnet2/train_ivector_extractor.sh --cmd "$train_cmd" --nj $nj \
      data/${train_set}_hires_nopitch exp/chain/diag_ubm_${affix} \
      exp/chain/extractor_${affix} || exit 1;
  
    for datadir in ${train_set} ${test_sets}; do
      steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 data/${datadir}_hires_nopitch data/${datadir}_hires_nopitch_max2
      steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
        data/${datadir}_hires_nopitch_max2 exp/chain/extractor_${affix} exp/chain/ivectors_${datadir}_${affix} || exit 1;
    done
  fi
  
  if [ $stage -le 7 ]; then
    # Get the alignments as lattices (gives the LF-MMI training more freedom).
    # use the same num-jobs as the alignments
    nj=$(cat $ali_dir/num_jobs) || exit 1;
    steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
      data/lang exp/tri3 exp/tri4_sp_lats
    rm exp/tri4_sp_lats/fsts.*.gz # save space
  fi
  
  if [ $stage -le 8 ]; then
    # Create a version of the lang/ directory that has one state per phone in the
    # topo file. [note, it really has two states.. the first one is only repeated
    # once, the second one has zero or more repeats.]
    rm -rf $lang
    cp -r data/lang $lang
    silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
    nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
    # Use our special topology... note that later on may have to tune this
    # topology.
    steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
  fi
  
  if [ $stage -le 9 ]; then
    # Build a tree using our new topology. This is the critically different
    # step compared with other recipes.
    steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
      --context-opts "--context-width=2 --central-position=1" \
      --cmd "$train_cmd" 5000 data/$train_set $lang $ali_dir $treedir
  fi
  
  if [ $stage -le 10 ]; then
    echo "$0: creating neural net configs using the xconfig parser";
    feat_dim=$(feat-to-dim scp:data/${train_set}_hires/feats.scp -)
    num_targets=$(tree-info $treedir/tree | grep num-pdfs | awk '{print $2}')
    learning_rate_factor=$(echo "print (0.5/$xent_regularize)" | python)
    opts="l2-regularize=0.002"
    linear_opts="orthonormal-constraint=1.0"
    output_opts="l2-regularize=0.0005 bottleneck-dim=256"
  
    mkdir -p $dir/configs
    cat <<EOF > $dir/configs/network.xconfig
    input dim=100 name=ivector
    input dim=$feat_dim name=input
  
    # please note that it is important to have input layer with the name=input
    # as the layer immediately preceding the fixed-affine-layer to enable
    # the use of short notation for the descriptor
    fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
  
    # the first splicing is moved before the lda layer, so no splicing here
    relu-batchnorm-dropout-layer name=tdnn1 $opts dim=1280
    linear-component name=tdnn2l dim=256 $linear_opts input=Append(-1,0)
    relu-batchnorm-dropout-layer name=tdnn2 $opts input=Append(0,1) dim=1280
    linear-component name=tdnn3l dim=256 $linear_opts
    relu-batchnorm-dropout-layer name=tdnn3 $opts dim=1280
    linear-component name=tdnn4l dim=256 $linear_opts input=Append(-1,0)
    relu-batchnorm-dropout-layer name=tdnn4 $opts input=Append(0,1) dim=1280
    linear-component name=tdnn5l dim=256 $linear_opts
    relu-batchnorm-dropout-layer name=tdnn5 $opts dim=1280 input=Append(tdnn5l, tdnn3l)
    linear-component name=tdnn6l dim=256 $linear_opts input=Append(-3,0)
    relu-batchnorm-dropout-layer name=tdnn6 $opts input=Append(0,3) dim=1280
    linear-component name=tdnn7l dim=256 $linear_opts input=Append(-3,0)
    relu-batchnorm-dropout-layer name=tdnn7 $opts input=Append(0,3,tdnn6l,tdnn4l,tdnn2l) dim=1280
    linear-component name=tdnn8l dim=256 $linear_opts input=Append(-3,0)
    relu-batchnorm-dropout-layer name=tdnn8 $opts input=Append(0,3) dim=1280
    linear-component name=tdnn9l dim=256 $linear_opts input=Append(-3,0)
    relu-batchnorm-dropout-layer name=tdnn9 $opts input=Append(0,3,tdnn8l,tdnn6l,tdnn4l) dim=1280
    linear-component name=tdnn10l dim=256 $linear_opts input=Append(-3,0)
    relu-batchnorm-dropout-layer name=tdnn10 $opts input=Append(0,3) dim=1280
    linear-component name=tdnn11l dim=256 $linear_opts input=Append(-3,0)
    relu-batchnorm-dropout-layer name=tdnn11 $opts input=Append(0,3,tdnn10l,tdnn8l,tdnn6l) dim=1280
    linear-component name=prefinal-l dim=256 $linear_opts
  
    relu-batchnorm-layer name=prefinal-chain input=prefinal-l $opts dim=1280
    output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
  
    relu-batchnorm-layer name=prefinal-xent input=prefinal-l $opts dim=1280
    output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
  
  EOF
    steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
  fi
  
  if [ $stage -le 11 ]; then
    #if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
    #  utils/create_split_dir.pl \
    #   /export/b0{5,6,7,8}/$USER/kaldi-data/egs/aishell-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
    #fi
  
    steps/nnet3/chain/train.py --stage $train_stage \
      --cmd "$decode_cmd" \
      --feat.online-ivector-dir exp/chain/ivectors_${train_set}_${affix} \
      --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
      --chain.xent-regularize $xent_regularize \
      --chain.leaky-hmm-coefficient 0.1 \
      --chain.l2-regularize 0.00005 \
      --chain.apply-deriv-weights false \
      --chain.lm-opts="--num-extra-lm-states=2000" \
      --egs.dir "$common_egs_dir" \
      --egs.stage $get_egs_stage \
      --egs.opts "--frames-overlap-per-eg 0" \
      --egs.chunk-width $frames_per_eg \
      --trainer.dropout-schedule $dropout_schedule \
      --trainer.num-chunk-per-minibatch $minibatch_size \
      --trainer.frames-per-iter 1500000 \
      --trainer.num-epochs $num_epochs \
      --trainer.optimization.num-jobs-initial $num_jobs_initial \
      --trainer.optimization.num-jobs-final $num_jobs_final \
      --trainer.optimization.initial-effective-lrate $initial_effective_lrate \
      --trainer.optimization.final-effective-lrate $final_effective_lrate \
      --trainer.max-param-change $max_param_change \
      --cleanup.remove-egs $remove_egs \
      --feat-dir data/${train_set}_hires \
      --tree-dir $treedir \
      --lat-dir exp/tri4_sp_lats \
      --dir $dir  || exit 1;
  fi
  
  if [ $stage -le 12 ]; then
    # Note: it might appear that this $lang directory is mismatched, and it is as
    # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
    # the lang directory.
    utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test $dir $dir/graph
  fi
  
  graph_dir=$dir/graph
  if [ $stage -le 13 ]; then
    for test_set in $test_sets; do
      nj=$(wc -l data/${test_set}_hires/spk2utt | awk '{print $1}')
      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
        --nj $nj --cmd "$decode_cmd" \
        --online-ivector-dir exp/chain/ivectors_${test_set}_${affix} \
        $graph_dir data/${test_set}_hires $dir/decode_${test_set} || exit 1;
    done
  fi
  
  echo "local/chain/run_tdnn.sh succeeded"
  exit 0;