#!/bin/bash
  
  # 2016 Modified by Takafumi Moriya at Tokyo Institute of Technology
  # for Japanese speech recognition using CSJ.
  
  # Copyright 2012-2014  Brno University of Technology (Author: Karel Vesely)
  # Apache 2.0
  
  # This example script trains a bottleneck feature extractor with 
  # 'Universal Context' topology as invented by Frantisek Grezl,
  # the network is on top of FBANK+f0 features.
  
  . ./cmd.sh
  . ./path.sh
  
  # Config:
  stage=0 # resume training with --stage=N
  use_dev=false
  # End of config.
  . utils/parse_options.sh || exit 1;
  #
  
  [ ! -e data-fbank/train ] && if [ $stage -le 1 ]; then
      # prepare the FBANK+f0 features
      # all evaluation sets
      for eval_num in eval1 eval2 eval3 ;do
  	dir=data-fbank/$eval_num; srcdir=data/$eval_num
  	(mkdir -p $dir; cp $srcdir/* $dir; )
  	steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 10 $dir $dir/log $dir/data || exit 1;
  	steps/compute_cmvn_stats.sh $dir $dir/log $dir/data || exit 1;
      done
      # training set
      dir=data-fbank/train; srcdir=data/train
      (mkdir -p $dir; cp $srcdir/* $dir; )
      steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 10 $dir $dir/log $dir/data || exit 1;
      steps/compute_cmvn_stats.sh $dir $dir/log $dir/data || exit 1;
  fi
  
  if [ $stage -le 2 ]; then
    # Prepare same subsets as in the main MFCC-GMM recipe, these will be used 
    # during during building GMM system from flat-start, later in the Tandem recipe.
    data=data-fbank
  
    if $use_dev ;then
      dev_set=train_dev
    # Use the first 4k sentences as dev set.  Note: when we trained the LM, we used
      utils/subset_data_dir.sh --first $data/train 4000 $data/$dev_set # 6hr 31min
      n=$[`cat data/train/segments | wc -l` - 4000]
      utils/subset_data_dir.sh --last $data/train $n $data/train_nodev
    else
      cp -r $data/train $data/train_nodev
    fi
    
    # Prepare data for training mono
    # Take the first 30k utterances (about 1/8th of the data)
    utils/subset_data_dir.sh --shortest $data/train_nodev 100000 $data/train_100kshort
    utils/subset_data_dir.sh $data/train_100kshort 30000 $data/train_30kshort
  
    # Take the first 100k utterances (just under half the data); we'll use
    # this for later stages of training.
    utils/subset_data_dir.sh --first $data/train_nodev 100000 $data/train_100k
    local/remove_dup_utts.sh 200 $data/train_100k $data/train_100k_nodup
  
    # Full training dataset,
    local/remove_dup_utts.sh 300 $data/train_nodev $data/train_nodup
  
    # split the data : 90% train 10% cross-validation (held-out)
    dir=$data/train_nodup
    utils/subset_data_dir_tr_cv.sh $dir ${dir}_tr90 ${dir}_cv10 || exit 1
  fi
  
  #########################################################################################
  # Let's build universal-context bottleneck network
  # - Universal context MLP is a hierarchy of two bottleneck neural networks
  # - The first network can see a limited range of frames (11 frames)
  # - The second network sees concatenation of bottlneck outputs of the first 
  #   network, with temporal shifts -10 -5 0 5 10, (in total a range of 31 frames 
  #   in the original feature space)
  # - This structure has been reported to produce superior performance
  #   compared to a network with single bottleneck
  #
  if [ $stage -le 3 ]; then
    # 1st network, overall context +/-5 frames
    # - the topology is 90_1500_1500_80_1500_NSTATES, linear bottleneck
    dir=exp/nnet5b_uc-part1
    ali=exp/tri4_ali_nodup
    $cuda_cmd $dir/log/train_nnet.log \
      steps/nnet/train.sh --hid-layers 2 --hid-dim 1500 --bn-dim 80 --apply-cmvn true \
        --copy-feats false \
        --feat-type traps --splice 5 --traps-dct-basis 6 --learn-rate 0.008 \
      data-fbank/train_nodup_tr90 data-fbank/train_nodup_cv10 data/lang ${ali} ${ali} $dir || exit 1;
  fi
  if [ $stage -le 4 ]; then
    # Compose feature_transform for the next stage, 
    # - remaining part of the first network is fixed
    dir=exp/nnet5b_uc-part1
    feature_transform=$dir/final.feature_transform.part1
    nnet-concat $dir/final.feature_transform \
      "nnet-copy --remove-last-layers=4 --binary=false $dir/final.nnet - |" \
      "utils/nnet/gen_splice.py --fea-dim=80 --splice=2 --splice-step=5 |" \
      $feature_transform || exit 1
    
    # 2nd network, overall context +/-15 frames
    # - the topology will be 400_1500_1500_30_1500_NSTATES, again, the bottleneck is linear
    dir=exp/nnet5b_uc-part2
    ali=exp/tri4_ali_nodup
    $cuda_cmd $dir/log/train_nnet.log \
      steps/nnet/train.sh --hid-layers 2 --hid-dim 1500 --bn-dim 30 --apply-cmvn true \
      --feature-transform $feature_transform --learn-rate 0.008 \
      data-fbank/train_nodup_tr90 data-fbank/train_nodup_cv10 data/lang ${ali} ${ali} $dir || exit 1;
  fi
  #
  #########################################################################################
  
  if [ $stage -le 5 ]; then
    # Store the BN-features
    data=data-bn/nnet5b_uc-part2 
    srcdata=data-fbank
    nnet=exp/nnet5b_uc-part2
    
    # all evaluation sets
    for eval_num in eval1 eval2 eval3 $dev_set ;do
        steps/nnet/make_bn_feats.sh --cmd "$train_cmd" --nj 10 $data/$eval_num $srcdata/$eval_num \
  	  $nnet $data/$eval_num/log $data/$eval_num/data || exit 1
    done
    # trainig data (full set)
    steps/nnet/make_bn_feats.sh --cmd "$train_cmd" --nj 10 $data/train $srcdata/train \
      $nnet $data/train/log $data/train/data || exit 1
  
    # Compute CMVN of the BN-features
    dir=data-bn/nnet5b_uc-part2/train
    steps/compute_cmvn_stats.sh $dir $dir/log $dir/data || exit 1;
  
    for eval_num in eval1 eval2 eval3 $dev_set ;do
        dir=data-bn/nnet5b_uc-part2/$eval_num
        steps/compute_cmvn_stats.sh $dir $dir/log $dir/data || exit 1;
    done
  fi
  
  if [ $stage -le 6 ]; then
    # Prepare BN-feature subsets same as with MFCCs in run.sh 
    data=data-bn/nnet5b_uc-part2
    
    if $use_dev ;then
        dev_set=train_dev
        # Use the first 4k sentences as dev set.
        utils/subset_data_dir.sh --first $data/train 4000 $data/$dev_set # 6hr 31min
        n=$[`cat data/train/segments | wc -l` - 4000]
        utils/subset_data_dir.sh --last $data/train $n $data/train_nodev
    else
        cp -r $data/train $data/train_nodev
    fi
  
    # Prepare data for training mono
    # Take the first 30k utterances (about 1/8th of the data)
    utils/subset_data_dir.sh --shortest $data/train_nodev 100000 $data/train_100kshort
    utils/subset_data_dir.sh $data/train_100kshort 30000 $data/train_30kshort
  
    # Take the first 100k utterances (just under half the data); we'll use
    # this for later stages of training.
    utils/subset_data_dir.sh --first $data/train_nodev 100000 $data/train_100k
    local/remove_dup_utts.sh 200 $data/train_100k $data/train_100k_nodup
  
    # Full dataset
    local/remove_dup_utts.sh 300 $data/train_nodev $data/train_nodup
  fi
  
  
  # Start building the tandem GMM system
  # - train from mono to tri4, run bmmi training
  bndata=data-bn/nnet5b_uc-part2
  
  if [ $stage -le 7 ]; then
    steps/tandem/train_mono.sh --nj 10 --cmd "$train_cmd" \
      data/train_30kshort $bndata/train_30kshort data/lang exp/tandem2uc-mono0a || exit 1;
  
    steps/tandem/align_si.sh --nj 10 --cmd "$train_cmd" \
       data/train_100k_nodup $bndata/train_100k_nodup data/lang exp/tandem2uc-mono0a exp/tandem2uc-mono0a_ali || exit 1;
  
    steps/tandem/train_deltas.sh --cmd "$train_cmd" \
        3200 30000 data/train_100k_nodup $bndata/train_100k_nodup data/lang exp/tandem2uc-mono0a_ali exp/tandem2uc-tri1 || exit 1;
     
    utils/mkgraph.sh data/lang_csj_tg exp/tandem2uc-tri1 exp/tandem2uc-tri1/graph_csj_tg
  
    for eval_num in eval1 eval2 eval3 $dev_set ;do
        steps/tandem/decode.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_tandem.config \
  	  exp/tandem2uc-tri1/graph_csj_tg data/$eval_num $bndata/$eval_num exp/tandem2uc-tri1/decode_${eval_num}_csj
    done
  fi
  
  if [ $stage -le 8 ]; then
    steps/tandem/align_si.sh --nj 10 --cmd "$train_cmd" \
       data/train_100k_nodup $bndata/train_100k_nodup data/lang exp/tandem2uc-tri1 exp/tandem2uc-tri1_ali || exit 1;
  
    steps/tandem/train_deltas.sh --cmd "$train_cmd" \
       4000 70000 data/train_100k_nodup $bndata/train_100k_nodup data/lang exp/tandem2uc-tri1_ali exp/tandem2uc-tri2 || exit 1;
  
    utils/mkgraph.sh data/lang_csj_tg exp/tandem2uc-tri2 exp/tandem2uc-tri2/graph_csj_tg || exit 1;
    for eval_num in eval1 eval2 eval3 $dev_set ;do
        steps/tandem/decode.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_tandem.config \
  	  exp/tandem2uc-tri2/graph_csj_tg data/$eval_num $bndata/$eval_num exp/tandem2uc-tri2/decode_${eval_num}_csj || exit 1;
    done
  fi
  
  if [ $stage -le 9 ]; then
    steps/tandem/align_si.sh --nj 10 --cmd "$train_cmd" \
      data/train_nodup $bndata/train_nodup data/lang exp/tandem2uc-tri2 exp/tandem2uc-tri2_ali || exit 1;
  
    # Train tri3, which is LDA+MLLT, on train_nodup data.
    steps/tandem/train_lda_mllt.sh --cmd "$train_cmd" \
       --splice-opts "--left-context=3 --right-context=3" \
       6000 140000 data/train_nodup $bndata/train_nodup data/lang exp/tandem2uc-tri2_ali exp/tandem2uc-tri3 || exit 1;
  
    utils/mkgraph.sh data/lang_csj_tg exp/tandem2uc-tri3 exp/tandem2uc-tri3/graph_csj_tg || exit 1;
    for eval_num in eval1 eval2 eval3 $dev_set ;do
        steps/tandem/decode.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_tandem.config \
  	  exp/tandem2uc-tri3/graph_csj_tg data/$eval_num $bndata/$eval_num exp/tandem2uc-tri3/decode_${eval_num}_csj || exit 1;
    done
  fi
  
  if [ $stage -le 10 ]; then
    # From now, we start building a more serious system (with SAT), 
    # and we'll do the alignment with fMLLR.
    steps/tandem/align_fmllr.sh --nj 10 --cmd "$train_cmd" \
      data/train_nodup $bndata/train_nodup data/lang exp/tandem2uc-tri3 exp/tandem2uc-tri3_ali_nodup || exit 1;
  
    steps/tandem/train_sat.sh  --cmd "$train_cmd" \
      11500 200000 data/train_nodup $bndata/train_nodup data/lang exp/tandem2uc-tri3_ali_nodup exp/tandem2uc-tri4 || exit 1;
  
    utils/mkgraph.sh data/lang_csj_tg exp/tandem2uc-tri4 exp/tandem2uc-tri4/graph_csj_tg || exit 1
    for eval_num in eval1 eval2 eval3 $dev_set ;do
        steps/tandem/decode_fmllr.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_tandem.config \
  	  exp/tandem2uc-tri4/graph_csj_tg data/$eval_num $bndata/$eval_num exp/tandem2uc-tri4/decode_${eval_num}_csj || exit 1
    done
  fi
  
  # bMMI starting from system in tandem2uc-tri4, use full dataset.
  if [ $stage -le 11 ]; then
    steps/tandem/align_fmllr.sh --nj 10 --cmd "$train_cmd" \
      data/train_nodup $bndata/train_nodup data/lang exp/tandem2uc-tri4 exp/tandem2uc-tri4_ali || exit 1;
    steps/tandem/make_denlats.sh --nj 10 --cmd "$decode_cmd" --transform-dir exp/tandem2uc-tri4_ali \
      --sub-split 100 data/train_nodup $bndata/train_nodup data/lang exp/tandem2uc-tri4 exp/tandem2uc-tri4_denlats || exit 1;
  fi
  if [ $stage -le 12 ]; then
    steps/tandem/train_mmi.sh --cmd "$decode_cmd" --boost 0.1 --acwt 0.039 \
      data/train_nodup $bndata/train_nodup data/lang exp/tandem2uc-tri4_{ali,denlats} exp/tandem2uc-tri4_mmi_b0.1 || exit 1;
  
    for eval_num in eval1 eval2 eval3 $dev_set ;do
        steps/tandem/decode.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_tandem.config \
  	  --transform-dir exp/tandem2uc-tri4/decode_${eval_num}_csj \
  	  exp/tandem2uc-tri4/graph_csj_tg data/$eval_num $bndata/$eval_num exp/tandem2uc-tri4_mmi_b0.1/decode_${eval_num}_csj || exit 1;
    done
  fi
  
  echo success
  exit 0
  
  # Getting results [see RESULTS file]
  # for x in exp/*/decode*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done
  # We use config parameters of swbd resipe.
  :<<EOF
  === evaluation set 1 ===
  %WER 16.46 [ 4285 / 26028, 556 ins, 910 del, 2819 sub ] exp/tandem2uc-tri1/decode_eval1_csj/wer_20_0.0
  %WER 15.32 [ 3987 / 26028, 523 ins, 825 del, 2639 sub ] exp/tandem2uc-tri2/decode_eval1_csj/wer_20_0.0
  %WER 14.28 [ 3718 / 26028, 475 ins, 744 del, 2499 sub ] exp/tandem2uc-tri3/decode_eval1_csj/wer_20_0.0
  %WER 13.51 [ 3517 / 26028, 450 ins, 738 del, 2329 sub ] exp/tandem2uc-tri4/decode_eval1_csj/wer_20_0.5
  %WER 14.93 [ 3885 / 26028, 584 ins, 711 del, 2590 sub ] exp/tandem2uc-tri4/decode_eval1_csj.si/wer_20_0.0
  %WER 12.42 [ 3232 / 26028, 399 ins, 671 del, 2162 sub ] exp/tandem2uc-tri4_mmi_b0.1/decode_eval1_csj/wer_20_1.0
  === evaluation set 2 ===
  %WER 12.54 [ 3343 / 26661, 474 ins, 525 del, 2344 sub ] exp/tandem2uc-tri1/decode_eval2_csj/wer_20_0.0
  %WER 12.19 [ 3250 / 26661, 371 ins, 596 del, 2283 sub ] exp/tandem2uc-tri2/decode_eval2_csj/wer_20_1.0
  %WER 11.19 [ 2984 / 26661, 354 ins, 511 del, 2119 sub ] exp/tandem2uc-tri3/decode_eval2_csj/wer_20_0.5
  %WER 9.96 [ 2655 / 26661, 349 ins, 427 del, 1879 sub ] exp/tandem2uc-tri4/decode_eval2_csj/wer_20_0.5
  %WER 11.96 [ 3188 / 26661, 504 ins, 427 del, 2257 sub ] exp/tandem2uc-tri4/decode_eval2_csj.si/wer_20_0.0
  %WER 9.30 [ 2480 / 26661, 312 ins, 387 del, 1781 sub ] exp/tandem2uc-tri4_mmi_b0.1/decode_eval2_csj/wer_20_1.0
  === evaluation set 3 ===
  %WER 18.19 [ 3127 / 17189, 555 ins, 510 del, 2062 sub ] exp/tandem2uc-tri1/decode_eval3_csj/wer_20_0.5
  %WER 17.80 [ 3060 / 17189, 522 ins, 535 del, 2003 sub ] exp/tandem2uc-tri2/decode_eval3_csj/wer_20_1.0
  %WER 15.88 [ 2729 / 17189, 520 ins, 423 del, 1786 sub ] exp/tandem2uc-tri3/decode_eval3_csj/wer_20_0.5
  %WER 14.88 [ 2557 / 17189, 556 ins, 359 del, 1642 sub ] exp/tandem2uc-tri4/decode_eval3_csj/wer_20_0.5
  %WER 17.03 [ 2927 / 17189, 592 ins, 417 del, 1918 sub ] exp/tandem2uc-tri4/decode_eval3_csj.si/wer_20_1.0
  %WER 13.44 [ 2311 / 17189, 430 ins, 340 del, 1541 sub ] exp/tandem2uc-tri4_mmi_b0.1/decode_eval3_csj/wer_20_1.0
  EOF