#!/bin/bash
set -e

# run_tdnn_1b.sh's topo is similiar with run_tdnn_1a.sh but we used the xconfigs. Otherwise "frames_per_eg=150,140,100".

#exp/chain_cleaned/tdnn_1b_sp: num-iters=871 nj=3..16 num-params=17.1M dim=40+100->5151 combine=-0.074->-0.074 xent:train/valid[579,870,final]=(-1.02,-0.986,-0.990/-0.985,-0.953,-0.957) logprob:train/valid[579,870,final]=(-0.066,-0.062,-0.063/-0.070,-0.069,-0.069)

# by default, with cleanup:
# local/chain/run_tdnn.sh

# local/chain/compare_wer.sh exp/chain_cleaned/tdnn_1b_sp
# System                      tdnn_1b_sp
# WER on dev(fglarge)              3.87
# WER on dev(tglarge)              3.99
# WER on dev(tgmed)                4.96
# WER on dev(tgsmall)              5.42
# WER on dev_other(fglarge)       10.15
# WER on dev_other(tglarge)       10.77
# WER on dev_other(tgmed)         12.94
# WER on dev_other(tgsmall)       14.39
# WER on test(fglarge)             4.14
# WER on test(tglarge)             4.32
# WER on test(tgmed)               5.28
# WER on test(tgsmall)             5.88
# WER on test_other(fglarge)      10.80
# WER on test_other(tglarge)      11.13
# WER on test_other(tgmed)        13.37
# WER on test_other(tgsmall)      14.92
# Final train prob              -0.0626
# Final valid prob              -0.0687
# Final train prob (xent)       -0.9905
# Final valid prob (xent)       -0.9566

## how you run this (note: this assumes that the run_tdnn.sh soft link points here;
## otherwise call it directly in its location).
# without cleanup:
# local/chain/run_tdnn.sh  --train-set train_960 --gmm tri6b --nnet3-affix "" &

# configs for 'chain'
# this script is adapted from librispeech's 1c script.

# First the options that are passed through to run_ivector_common.sh
# (some of which are also used in this script directly).
stage=0
decode_nj=50
train_set=train_960_cleaned
gmm=tri6b_cleaned # the gmm for the target data
nnet3_affix=_cleaned  # cleanup affix for nnet3 and chain dirs, e.g. _cleaned

# The rest are configs specific to this script.  Most of the parameters
# are just hardcoded at this level, in the commands below.
affix=1b
tree_affix=
train_stage=-10
get_egs_stage=-10
decode_iter=

# TDNN options
# this script uses the new tdnn config generator so it needs a final 0 to reflect that the final layer input has no splicing
# training options
frames_per_eg=150,140,100
relu_dim=725
remove_egs=true
common_egs_dir=
xent_regularize=0.1
self_repair_scale=0.00001


# End configuration section.
echo "$0 $@"  # Print the command line for logging

. ./cmd.sh
. ./path.sh
. ./utils/parse_options.sh

if ! cuda-compiled; then
  cat <<EOF && exit 1
This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
If you want to use GPUs (and have them), go to src/, and configure and make on a machine
where "nvcc" is installed.
EOF
fi

# The iVector-extraction and feature-dumping parts are the same as the standard
# nnet3 setup, and you can skip them by setting "--stage 11" if you have already
# run those things.
local/nnet3/run_ivector_common.sh --stage $stage \
                                  --train-set $train_set \
                                  --gmm $gmm \
                                  --nnet3-affix "$nnet3_affix" || exit 1;

gmm_dir=exp/$gmm
ali_dir=exp/${gmm}_ali_${train_set}_sp
tree_dir=exp/chain${nnet3_affix}/tree_sp${tree_affix:+_$tree_affix}
lang=data/lang_chain
lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_sp_lats
dir=exp/chain${nnet3_affix}/tdnn${affix:+_$affix}_sp
train_data_dir=data/${train_set}_sp_hires
lores_train_data_dir=data/${train_set}_sp
train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires

for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
    $lores_train_data_dir/feats.scp $ali_dir/ali.1.gz; do
  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
done

# Please take this as a reference on how to specify all the options of
# local/chain/run_chain_common.sh
local/chain/run_chain_common.sh --stage $stage \
                                --gmm-dir $gmm_dir \
                                --ali-dir $ali_dir \
                                --lores-train-data-dir ${lores_train_data_dir} \
                                --lang $lang \
                                --lat-dir $lat_dir \
                                --tree-dir $tree_dir || exit 1;


if [ $stage -le 14 ]; then
  mkdir -p $dir

  echo "$0: creating neural net configs";
  # create the config files for nnet initialization

  num_targets=$(tree-info $tree_dir/tree |grep num-pdfs|awk '{print $2}')
  learning_rate_factor=$(echo "print (0.5/$xent_regularize)" | python)

  mkdir -p $dir/configs
  cat <<EOF > $dir/configs/network.xconfig
  input dim=100 name=ivector
  input dim=40 name=input

  # please note that it is important to have input layer with the name=input
  # as the layer immediately preceding the fixed-affine-layer to enable
  # the use of short notation for the descriptor
  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat

  # the first splicing is moved before the lda layer, so no splicing here
  relu-batchnorm-layer name=tdnn1 dim=$relu_dim
  relu-batchnorm-layer name=tdnn2 dim=$relu_dim input=Append(-1,0,1,2)
  relu-batchnorm-layer name=tdnn3 dim=$relu_dim input=Append(-3,0,3)
  relu-batchnorm-layer name=tdnn4 dim=$relu_dim input=Append(-3,0,3)
  relu-batchnorm-layer name=tdnn5 dim=$relu_dim input=Append(-3,0,3)
  relu-batchnorm-layer name=tdnn6 dim=$relu_dim input=Append(-6,-3,0)

  ## adding the layers for chain branch
  relu-batchnorm-layer name=prefinal-chain dim=$relu_dim target-rms=0.5
  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5

  # adding the layers for xent branch
  # This block prints the configs for a separate output that will be
  # trained with a cross-entropy objective in the 'chain' models... this
  # has the effect of regularizing the hidden parts of the model.  we use
  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
  # 0.5 / args.xent_regularize is suitable as it means the xent
  # final-layer learns at a rate independent of the regularization
  # constant; and the 0.5 was tuned so as to make the relative progress
  # similar in the xent and regular final layers.
  relu-batchnorm-layer name=prefinal-xent input=tdnn6 dim=$relu_dim target-rms=0.5
  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
EOF
  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/


fi



if [ $stage -le 15 ]; then
  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
    utils/create_split_dir.pl \
     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/librispeech-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
  fi

  steps/nnet3/chain/train.py --stage $train_stage \
    --cmd "$decode_cmd" \
    --feat.online-ivector-dir $train_ivector_dir \
    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
    --chain.xent-regularize $xent_regularize \
    --chain.leaky-hmm-coefficient 0.1 \
    --chain.l2-regularize 0.00005 \
    --chain.apply-deriv-weights false \
    --chain.lm-opts="--num-extra-lm-states=2000" \
    --egs.stage $get_egs_stage \
    --egs.opts "--frames-overlap-per-eg 0" \
    --egs.chunk-width $frames_per_eg \
    --egs.dir "$common_egs_dir" \
    --trainer.num-chunk-per-minibatch 128 \
    --trainer.frames-per-iter 1500000 \
    --trainer.num-epochs 4 \
    --trainer.optimization.num-jobs-initial 3 \
    --trainer.optimization.num-jobs-final 16 \
    --trainer.optimization.initial-effective-lrate 0.001 \
    --trainer.optimization.final-effective-lrate 0.0001 \
    --trainer.max-param-change 2 \
    --cleanup.remove-egs $remove_egs \
    --feat-dir $train_data_dir \
    --tree-dir $tree_dir \
    --lat-dir $lat_dir \
    --dir $dir  || exit 1;
fi


graph_dir=$dir/graph_tgsmall
if [ $stage -le 16 ]; then
  # Note: it might appear that this $lang directory is mismatched, and it is as
  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
  # the lang directory.
  utils/mkgraph.sh --self-loop-scale 1.0 --remove-oov data/lang_test_tgsmall $dir $graph_dir
  # remove <UNK> from the graph, and convert back to const-FST.
  fstrmsymbols --apply-to-output=true --remove-arcs=true "echo 3|" $graph_dir/HCLG.fst - | \
    fstconvert --fst_type=const > $graph_dir/temp.fst
  mv $graph_dir/temp.fst $graph_dir/HCLG.fst
fi


if [ $stage -le 17 ]; then
  iter_opts=
  if [ ! -z $decode_iter ]; then
    iter_opts=" --iter $decode_iter "
  fi
  rm $dir/.error 2>/dev/null || true
  for decode_set in test_clean test_other dev_clean dev_other; do
      (
      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
          --nj $decode_nj --cmd "$decode_cmd" $iter_opts \
          --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_tgsmall || exit 1
      steps/lmrescore.sh --cmd "$decode_cmd" --self-loop-scale 1.0 data/lang_test_{tgsmall,tgmed} \
          data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_{tgsmall,tgmed} || exit 1
      steps/lmrescore_const_arpa.sh \
          --cmd "$decode_cmd" data/lang_test_{tgsmall,tglarge} \
          data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_{tgsmall,tglarge} || exit 1
      steps/lmrescore_const_arpa.sh \
          --cmd "$decode_cmd" data/lang_test_{tgsmall,fglarge} \
          data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_{tgsmall,fglarge} || exit 1
      ) || touch $dir/.error &
  done
  wait
  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
fi

exit 0;