set -e
# The run_tdnn_1a.sh is a first attempt at an TDNN system, based on configs.
# It created befroe we introduced xconfigs.
# See run_tdnn_1b.sh for comparative results.


# chain_cleaned/tdnn_1a_sp: num_params=16.8M (12.7M after excluding the xent branch), average training time=71.8s per job(on Tesla K80), real-time factor=0.558894


# for x in exp/chain_cleaned/tdnn_1a_sp/decode_*; do grep WER $x/wer_* | utils/best_wer.sh ; done
#System                      tdnn_1a_sp
#WER on dev(fglarge)           3.87
#WER on dev(tglarge)           3.97
#WER on dev(tgmed)             4.95
#WER on dev(tgsmall)           5.57
#WER on dev_other(fglarge)    10.22
#WER on dev_other(tglarge)    10.79
#WER on dev_other(tgmed)      13.01
#WER on dev_other(tgsmall)    14.36
#WER on test(fglarge)          4.17
#WER on test(tglarge)          4.36
#WER on test(tgmed)            5.33
#WER on test(tgsmall)          5.93
#WER on test_other(fglarge)   10.62
#WER on test_other(tglarge)   10.96
#WER on test_other(tgmed)     13.24
#WER on test_other(tgsmall)   14.53

## how you run this (note: this assumes that the run_tdnn.sh soft link points here;
## otherwise call it directly in its location).
# by default, with cleanup:
# local/chain/run_tdnn.sh

# without cleanup:
# local/chain/run_tdnn_1a.sh  --train-set train_960 --gmm tri6b --nnet3-affix "" &

# configs for 'chain'
# this script is adapted from swbd's 7b script, but the relu-dim is larger.

# 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=1a
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
relu_dim=725
remove_egs=false
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 "$0: creating neural net configs";
  # create the config files for nnet initialization
  repair_opts=${self_repair_scale:+" --self-repair-scale-nonlinearity $self_repair_scale "}

  steps/nnet3/tdnn/make_configs.py $repair_opts \
    --feat-dir $train_data_dir \
    --ivector-dir $train_ivector_dir \
    --tree-dir $tree_dir \
    --relu-dim $relu_dim \
    --splice-indexes "-1,0,1 -1,0,1,2 -3,0,3 -3,0,3 -3,0,3 -6,-3,0 0" \
    --use-presoftmax-prior-scale false \
    --xent-regularize $xent_regularize \
    --xent-separate-forward-affine true \
    --include-log-softmax false \
    --final-layer-normalize-target 0.5 \
    $dir/configs || exit 1;
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

  touch $dir/egs/.nodelete # keep egs around when that run dies.

  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
  # romove <UNK> from the graph
  fstrmsymbols --apply-to-output=true --remove-arcs=true "echo 3|" $graph_dir/HCLG.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;