#!/bin/bash
  
  # Copyright 2012-2014  Brno University of Technology (Author: Karel Vesely)
  #                2014  Guoguo Chen
  # Apache 2.0
  
  # This example script trains a DNN on top of fMLLR features. 
  # The training is done in 3 stages,
  #
  # 1) RBM pre-training:
  #    in this unsupervised stage we train stack of RBMs, 
  #    a good starting point for frame cross-entropy trainig.
  # 2) frame cross-entropy training:
  #    the objective is to classify frames to correct pdfs.
  # 3) sequence-training optimizing sMBR: 
  #    the objective is to emphasize state-sequences with better 
  #    frame accuracy w.r.t. reference alignment.
  
  . ./cmd.sh ## You'll want to change cmd.sh to something that will work on your system.
             ## This relates to the queue.
  
  . ./path.sh ## Source the tools/utils (import the queue.pl)
  
  # Config:
  gmmdir=exp/tri5a
  data_fmllr=data-fmllr-tri5a
  stage=0 # resume training with --stage=N
  # End of config.
  . utils/parse_options.sh || exit 1;
  #
  
  if [ $stage -le 0 ]; then
    # Store fMLLR features, so we can train on them easily,
    # dev
    dir=$data_fmllr/dev
    steps/nnet/make_fmllr_feats.sh --nj 10 --cmd "$train_cmd" \
       --transform-dir $gmmdir/decode \
       $dir data/dev $gmmdir $dir/log $dir/data || exit 1
    # train
    dir=$data_fmllr/train
    steps/nnet/make_fmllr_feats.sh --nj 10 --cmd "$train_cmd" \
       --transform-dir ${gmmdir}_ali \
       $dir data/train $gmmdir $dir/log $dir/data || exit 1
    # split the data : 90% train 10% cross-validation (held-out)
    utils/subset_data_dir_tr_cv.sh $dir ${dir}_tr90 ${dir}_cv10 || exit 1
  fi
  
  if [ $stage -le 1 ]; then
    # Pre-train DBN, i.e. a stack of RBMs
    dir=exp/dnn5b_pretrain-dbn
    (tail --pid=$$ -F $dir/log/pretrain_dbn.log 2>/dev/null)& # forward log
    $cuda_cmd $dir/log/pretrain_dbn.log \
      steps/nnet/pretrain_dbn.sh --rbm-iter 1 --nn-depth 4 --hid-dim 2000 \
      $data_fmllr/train $dir || exit 1;
  fi
  
  if [ $stage -le 2 ]; then
    # Train the DNN optimizing per-frame cross-entropy.
    dir=exp/dnn5b_pretrain-dbn_dnn
    ali=${gmmdir}_ali
    feature_transform=exp/dnn5b_pretrain-dbn/final.feature_transform
    dbn=exp/dnn5b_pretrain-dbn/4.dbn
    (tail --pid=$$ -F $dir/log/train_nnet.log 2>/dev/null)& # forward log
    # Train
    $cuda_cmd $dir/log/train_nnet.log \
      steps/nnet/train.sh --feature-transform $feature_transform --dbn $dbn --hid-layers 0 --learn-rate 0.008 \
      $data_fmllr/train_tr90 $data_fmllr/train_cv10 data/lang $ali $ali $dir || exit 1;
    # Decode with the trigram language model.
    steps/nnet/decode.sh --nj 10 --cmd "$decode_cmd" \
      --config conf/decode_dnn.config --acwt 0.1 \
      $gmmdir/graph $data_fmllr/dev \
      $dir/decode || exit 1;
  fi
  
  # Sequence training using sMBR criterion, we do Stochastic-GD 
  # with per-utterance updates. We use usually good acwt 0.1
  # Lattices are re-generated after 1st epoch, to get faster convergence.
  dir=exp/dnn5b_pretrain-dbn_dnn_smbr
  srcdir=exp/dnn5b_pretrain-dbn_dnn
  acwt=0.1
  
  if [ $stage -le 3 ]; then
    # First we generate lattices and alignments:
    steps/nnet/align.sh --nj 10 --cmd "$train_cmd" \
      $data_fmllr/train data/lang $srcdir ${srcdir}_ali || exit 1;
    steps/nnet/make_denlats.sh --nj 10 --sub-split 20 --cmd "$decode_cmd" --config conf/decode_dnn.config \
      --acwt $acwt $data_fmllr/train data/lang $srcdir ${srcdir}_denlats || exit 1;
  fi
  
  if [ $stage -le 4 ]; then
    # Re-train the DNN by 1 iteration of sMBR 
    steps/nnet/train_mpe.sh --cmd "$cuda_cmd" --num-iters 1 --acwt $acwt --do-smbr true \
      $data_fmllr/train data/lang $srcdir ${srcdir}_ali ${srcdir}_denlats $dir || exit 1
    # Decode (reuse HCLG graph)
    for ITER in 1; do
      # Decode with the trigram swbd language model.
      steps/nnet/decode.sh --nj 10 --cmd "$decode_cmd" \
        --config conf/decode_dnn.config \
        --nnet $dir/${ITER}.nnet --acwt $acwt \
        $gmmdir/graph $data_fmllr/dev \
        $dir/decode || exit 1;
    done 
  fi
  
  # Re-generate lattices, run 2 more sMBR iterations
  dir=exp/dnn5b_pretrain-dbn_dnn_smbr_i1lats
  srcdir=exp/dnn5b_pretrain-dbn_dnn_smbr
  acwt=0.0909
  
  if [ $stage -le 5 ]; then
    # First we generate lattices and alignments:
    #steps/nnet/align.sh --nj 10 --cmd "$train_cmd" \
    #  $data_fmllr/train data/lang $srcdir ${srcdir}_ali || exit 1;
    steps/nnet/make_denlats.sh --nj 10 --sub-split 20 --cmd "$decode_cmd" --config conf/decode_dnn.config \
      --acwt $acwt $data_fmllr/train data/lang $srcdir ${srcdir}_denlats || exit 1;
  fi
  
  if [ $stage -le 6 ]; then
    # Re-train the DNN by 2 iteration of sMBR 
    steps/nnet/train_mpe.sh --cmd "$cuda_cmd" --num-iters 2 --acwt $acwt --do-smbr true \
      $data_fmllr/train data/lang $srcdir ${srcdir}_ali ${srcdir}_denlats $dir || exit 1
    # Decode (reuse HCLG graph)
    for ITER in 1 2; do
      # Decode with the trigram language model.
      steps/nnet/decode.sh --nj 10 --cmd "$decode_cmd" \
        --config conf/decode_dnn.config \
        --nnet $dir/${ITER}.nnet --acwt $acwt \
        $gmmdir/graph $data_fmllr/dev \
        $dir/decode || exit 1;
    done 
  fi
  
  # Getting results [see RESULTS file]
  # for x in exp/*/decode*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done