#!/bin/bash
  
  # Copyright 2012-2013  Brno University of Technology (Author: Karel Vesely)
  # Apache 2.0
  
  # In this recipe we build DNN in four stages:
  # 1) Data preparations : the fMLLR features are stored to disk
  # 2) RBM pre-training : in this unsupervised stage we train stack of RBMs, a good starting point for Cross-entropy trainig
  # 3) Frame-level cross-entropy training : in this stage the objective is to classify frames correctly.
  # 4) Sequence-criterion training : in this stage the objective is to classify the whole sequence correctly,
  #     the idea is similar to the 'Discriminative training' in context of GMM-HMMs.
  
  
  . ./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)
  
  
  
  #false && \
  {
  gmmdir=exp/tri3b
  
  ### Generate the alignments of dev93 
  ### (held-out set for Cross-entropy training)
  ###
  steps/align_fmllr.sh --nj 10 --cmd "$train_cmd" \
    data/dev lang $gmmdir exp/tri3b_ali_dev || exit 1
  
  ###
  ### Store the fMLLR features, so we can train on them easily
  ###
  
  # train si284
  # generate the features
  dir=data-fmllr-tri3b/train
  steps/make_fmllr_feats.sh --nj 20 --cmd "$train_cmd" \
     --transform-dir exp/tri3b_ali \
     $dir data/train $gmmdir $dir/_log $dir/_data || exit 1
  
  
  # dev93 (unsupervised fMLLR)
  # held-out set of Cross-entropy training
  dir=data-fmllr-tri3b/dev
  steps/make_fmllr_feats.sh --nj 10 --cmd "$train_cmd" \
     --transform-dir exp/tri3b/decode_dev \
     $dir data/dev $gmmdir $dir/_log $dir/_data || exit 1
  }
  
  
  
  ###
  ### Now we can pre-train stack of RBMs
  ###
  #false && \
  { # Pre-train the DBN
  dir=exp/tri3b_pretrain-dbn
  (tail --pid=$$ -F $dir/_pretrain_dbn.log 2>/dev/null)&
  $cuda_cmd $dir/_pretrain_dbn.log \
    steps/pretrain_dbn.sh --rbm-iter 3 data-fmllr-tri3b/train $dir
  }
  
  
  
  ###
  ### Now we train the DNN optimizing cross-entropy.
  ### This will take quite some time.
  ###
  
  #false && \
  { # Train the MLP
  dir=exp/tri3b_pretrain-dbn_dnn
  ali=exp/tri3b_ali
  feature_transform=exp/tri3b_pretrain-dbn/final.feature_transform
  dbn=exp/tri3b_pretrain-dbn/6.dbn
  (tail --pid=$$ -F $dir/_train_nnet.log 2>/dev/null)& 
  $cuda_cmd $dir/_train_nnet.log \
    steps/train_nnet.sh --feature-transform $feature_transform --dbn $dbn --hid-layers 0 --learn-rate 0.008 \
    data-fmllr-tri3b/train data-fmllr-tri3b/dev lang ${ali} ${ali}_dev $dir || exit 1;
  # decode with 'big-dictionary' (reuse HCLG graph)
  steps/decode_nnet.sh --nj 10 --cmd "$decode_cmd" --acwt 0.10 --config conf/decode_dnn.config \
    exp/tri3b/graph data-fmllr-tri3b/dev $dir/decode_dev || exit 1;
  }
  
  
  
  ###
  ### Finally we train using sMBR criterion.
  ### We do Stochastic-GD with per-utterance updates. 
  ###
  ### To get faster convergence, we will re-generate 
  ### the lattices after 1st epoch of sMBR.
  ###
  
  dir=exp/tri4b_pretrain-dbn_dnn_smbr
  srcdir=exp/tri4b_pretrain-dbn_dnn
  acwt=0.10
  
  # First we need to generate lattices and alignments:
  #false && \
  {
  steps/align_nnet.sh --nj 100 --cmd "$train_cmd" \
    data-fmllr-tri4b/train_si284 lang $srcdir ${srcdir}_ali_si284 || exit 1;
  steps/make_denlats_nnet.sh --nj 100 --cmd "$decode_cmd" \
    --config conf/decode_dnn.config --acwt $acwt \
    data-fmllr-tri4b/train_si284 lang $srcdir ${srcdir}_denlats_si284  || exit 1;
  }
  # Now we re-train the hybrid by single iteration of sMBR 
  #false && \
  {
  steps/train_nnet_mpe.sh --cmd "$cuda_cmd" --num-iters 1 --acwt $acwt --do-smbr true \
    data-fmllr-tri4b/train_si284 lang $srcdir \
    ${srcdir}_ali_si284 ${srcdir}_denlats_si284 $dir || exit 1
  }
  # Decode
  #false && \
  {
  for ITER in 1; do
    # decode dev93 with big dict graph_bd_tgpr
    steps/decode_nnet.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_dnn.config \
      --nnet $dir/${ITER}.nnet --acwt $acwt \
      exp/tri4b/graph_bd_tgpr data-fmllr-tri4b/test_dev93 $dir/decode_dev93_bd_tgpr_it${ITER} || exit 1
    # decode eval92 with big dict graph_bd_tgpr
    steps/decode_nnet.sh --nj 8 --cmd "$decode_cmd" --config conf/decode_dnn.config \
      --nnet $dir/${ITER}.nnet --acwt $acwt \
      exp/tri4b/graph_bd_tgpr data-fmllr-tri4b/test_eval92 $dir/decode_eval92_bd_tgpr_it${ITER} || exit 1
  done 
  }
  
  
  ###
  ### Re-generate lattices and run several more iterations of sMBR
  ###
  
  dir=exp/tri4b_pretrain-dbn_dnn_smbr_iter1-lats
  srcdir=exp/tri4b_pretrain-dbn_dnn_smbr
  acwt=0.10
  
  # First we need to generate lattices and alignments:
  #false && \
  {
  steps/align_nnet.sh --nj 100 --cmd "$train_cmd" \
    data-fmllr-tri4b/train_si284 lang $srcdir ${srcdir}_ali_si284 || exit 1;
  steps/make_denlats_nnet.sh --nj 100 --cmd "$decode_cmd" \
    --config conf/decode_dnn.config --acwt $acwt \
    data-fmllr-tri4b/train_si284 lang $srcdir ${srcdir}_denlats_si284  || exit 1;
  }
  # Now we re-train the hybrid by several iterations of sMBR 
  #false && \
  {
  steps/train_nnet_mpe.sh --cmd "$cuda_cmd" --num-iters 4 --acwt $acwt --do-smbr true \
    data-fmllr-tri4b/train_si284 lang $srcdir \
    ${srcdir}_ali_si284 ${srcdir}_denlats_si284 $dir 
  }
  # Decode
  #false && \
  {
  for ITER in 1 2 3 4; do
    # decode dev93 with big dict graph_bd_tgpr
    steps/decode_nnet.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_dnn.config \
      --nnet $dir/${ITER}.nnet --acwt $acwt \
      exp/tri4b/graph_bd_tgpr data-fmllr-tri4b/test_dev93 $dir/decode_dev93_bd_tgpr_it${ITER} || exit 1
    # decode eval92 with big dict graph_bd_tgpr
    steps/decode_nnet.sh --nj 8 --cmd "$decode_cmd" --config conf/decode_dnn.config \
      --nnet $dir/${ITER}.nnet --acwt $acwt \
      exp/tri4b/graph_bd_tgpr data-fmllr-tri4b/test_eval92 $dir/decode_eval92_bd_tgpr_it${ITER} || exit 1
  done 
  }
  
  
  # Getting results [see RESULTS file]
  # for x in exp/*/decode*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done