Yannick Estève / ONTRAC-Kaldi

Blame view

egs/wsj/s5/local/nnet/run_dnn.sh 4.53 KB
edit raw normal view history
8dcb6dfcb   Yannick Estève   first commit 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
 
  #!/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.
  
  . ./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/tri4b
  data_fmllr=data-fmllr-tri4b
  stage=0 # resume training with --stage=N
  # End of config.
  . utils/parse_options.sh || exit 1;
  #
  
  if [ $stage -le 0 ]; then
    # Store fMLLR features, so we can train on them easily,
    # test_dev93
    dir=$data_fmllr/test_dev93
    steps/nnet/make_fmllr_feats.sh --nj 10 --cmd "$train_cmd" \
       --transform-dir $gmmdir/decode_bd_tgpr_dev93 \
       $dir data/test_dev93 $gmmdir $dir/log $dir/data || exit 1
    # test_eval92
    dir=$data_fmllr/test_eval92
    steps/nnet/make_fmllr_feats.sh --nj 8 --cmd "$train_cmd" \
       --transform-dir $gmmdir/decode_bd_tgpr_eval92 \
       $dir data/test_eval92 $gmmdir $dir/log $dir/data || exit 1
    # train
    dir=$data_fmllr/train_si284
    steps/nnet/make_fmllr_feats.sh --nj 10 --cmd "$train_cmd" \
       --transform-dir ${gmmdir}_ali_si284 \
       $dir data/train_si284 $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 3 $data_fmllr/train_si284 $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_si284
    feature_transform=exp/dnn5b_pretrain-dbn/final.feature_transform
    dbn=exp/dnn5b_pretrain-dbn/6.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_si284_tr90 $data_fmllr/train_si284_cv10 data/lang $ali $ali $dir || exit 1;
    # Decode (reuse HCLG graph)
    steps/nnet/decode.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_dnn.config --acwt 0.1 \
      $gmmdir/graph_bd_tgpr $data_fmllr/test_dev93 $dir/decode_bd_tgpr_dev93 || exit 1;
    steps/nnet/decode.sh --nj 8 --cmd "$decode_cmd" --config conf/decode_dnn.config --acwt 0.1 \
      $gmmdir/graph_bd_tgpr $data_fmllr/test_eval92 $dir/decode_bd_tgpr_eval92 || 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 [ $stage -le 3 ]; then
    # First we generate lattices and alignments:
    steps/nnet/align.sh --nj 100 --cmd "$train_cmd" \
      $data_fmllr/train_si284 data/lang $srcdir ${srcdir}_ali || exit 1;
    steps/nnet/make_denlats.sh --nj 100 --cmd "$decode_cmd" --config conf/decode_dnn.config --acwt $acwt \
      $data_fmllr/train_si284 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 5 --acwt $acwt --do-smbr true \
      $data_fmllr/train_si284 data/lang $srcdir ${srcdir}_ali ${srcdir}_denlats $dir || exit 1
    # Decode (reuse HCLG graph)
    for ITER in 5 4 3 1; do
      steps/nnet/decode.sh --nj 10 --cmd "$decode_cmd" --config conf/decode_dnn.config \
        --nnet $dir/${ITER}.nnet --acwt $acwt \
        $gmmdir/graph_bd_tgpr $data_fmllr/test_dev93 $dir/decode_bd_tgpr_dev93_it${ITER} || exit 1;
      steps/nnet/decode.sh --nj 8 --cmd "$decode_cmd" --config conf/decode_dnn.config \
        --nnet $dir/${ITER}.nnet --acwt $acwt \
        $gmmdir/graph_bd_tgpr $data_fmllr/test_eval92 $dir/decode_bd_tgpr_eval92_it${ITER} || 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