#!/bin/bash
  
  # Copyright 2012-2014  Brno University of Technology (Author: Karel Vesely)
  # 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.
  
  # Note: With DNNs in RM, the optimal LMWT is 2-6. Don't be tempted to try acwt's like 0.2, 
  # the value 0.1 is better both for decoding and sMBR.
  
  . ./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)
  
  set -eu
  
  # Config:
  gmm=exp/tri3b
  data_fmllr=data-fmllr-tri3b
  stage=0 # resume training with --stage=N
  # End of config.
  . utils/parse_options.sh
  #
  
  [ ! -e $data_fmllr/test ] && if [ $stage -le 0 ]; then
    # Store fMLLR features, so we can train on them easily,
    # test
    dir=$data_fmllr/test
    steps/nnet/make_fmllr_feats.sh --nj 10 --cmd "$train_cmd" \
       --transform-dir $gmm/decode \
       $dir data/test $gmm $dir/log $dir/data
    # train
    dir=$data_fmllr/train
    steps/nnet/make_fmllr_feats.sh --nj 10 --cmd "$train_cmd" \
       --transform-dir ${gmm}_ali \
       $dir data/train $gmm $dir/log $dir/data
    # split the data : 90% train 10% cross-validation (held-out)
    utils/subset_data_dir_tr_cv.sh $dir ${dir}_tr90 ${dir}_cv10
  fi
  
  if [ $stage -le 1 ]; then
    # Pre-train DBN, i.e. a stack of RBMs (small database, smaller DNN)
    dir=exp/dnn4b_pretrain-dbn
    $cuda_cmd $dir/log/pretrain_dbn.log \
      steps/nnet/pretrain_dbn.sh --hid-dim 1024 --rbm-iter 20 $data_fmllr/train $dir
  fi
  
  if [ $stage -le 2 ]; then
    # Train the DNN optimizing per-frame cross-entropy.
    dir=exp/dnn4b_pretrain-dbn_dnn
    ali=${gmm}_ali
    feature_transform=exp/dnn4b_pretrain-dbn/final.feature_transform
    dbn=exp/dnn4b_pretrain-dbn/6.dbn
    # 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 $dir
    # Decode (reuse HCLG graph)
    steps/nnet/decode.sh --nj 20 --cmd "$decode_cmd" --config conf/decode_dnn.config --acwt 0.1 \
      $gmm/graph $data_fmllr/test $dir/decode
    steps/nnet/decode.sh --nj 20 --cmd "$decode_cmd" --config conf/decode_dnn.config --acwt 0.1 \
      $gmm/graph_ug $data_fmllr/test $dir/decode_ug
  fi
  
  
  # Sequence training using sMBR criterion, we do Stochastic-GD with per-utterance updates.
  # Note: With DNNs in RM, the optimal LMWT is 2-6. Don't be tempted to try acwt's like 0.2, 
  # the value 0.1 is better both for decoding and sMBR.
  dir=exp/dnn4b_pretrain-dbn_dnn_smbr
  srcdir=exp/dnn4b_pretrain-dbn_dnn
  acwt=0.1
  
  if [ $stage -le 3 ]; then
    # First we generate lattices and alignments:
    steps/nnet/align.sh --nj 20 --cmd "$train_cmd" \
      $data_fmllr/train data/lang $srcdir ${srcdir}_ali
    steps/nnet/make_denlats.sh --nj 20 --cmd "$decode_cmd" --config conf/decode_dnn.config --acwt $acwt \
      $data_fmllr/train data/lang $srcdir ${srcdir}_denlats
  fi
  
  if [ $stage -le 4 ]; then
    # Re-train the DNN by 6 iterations of sMBR 
    steps/nnet/train_mpe.sh --cmd "$cuda_cmd" --num-iters 6 --acwt $acwt --do-smbr true \
      $data_fmllr/train data/lang $srcdir ${srcdir}_ali ${srcdir}_denlats $dir
    # Decode
    for ITER in 6 3 1; do
      steps/nnet/decode.sh --nj 20 --cmd "$decode_cmd" --config conf/decode_dnn.config \
        --nnet $dir/${ITER}.nnet --acwt $acwt \
        $gmm/graph $data_fmllr/test $dir/decode_it${ITER}
    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
  
  # to see how model conversion to nnet2 works, run run_dnn_convert_nnet2.sh at this point.