#!/bin/bash
  
  # This is neural net training on top of adapted 40-dimensional features.
  #
  
  train_stage=-10
  use_gpu=true
  
  train_set="train-clean-100"
  test_sets="dev-clean dev-other"
  
  . ./cmd.sh
  . ./path.sh
  . utils/parse_options.sh
  
  
  if $use_gpu; then
    if ! cuda-compiled; then
      cat <<EOF && exit 1
  This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
  If you want to use GPUs (and have them), go to src/, and configure and make on a machine
  where "nvcc" is installed.
  EOF
    fi
    parallel_opts="--gpu 1"
    num_threads=1
    minibatch_size=512
    dir=exp/nnet5c_gpu_${train_set}
  else
    num_threads=16
    parallel_opts="--num-threads $num_threads"
    dir=exp/nnet5c_${train_set}
    minibatch_size=128
  fi
  
  if [ ! -f $dir/final.mdl ]; then
    if [ "$USER" == dpovey ]; then
       # spread the egs over various machines.  will help reduce overload of any
       # one machine.
       utils/create_split_dir.pl /export/b0{1,2,3,4}/dpovey/kaldi-pure/egs/wsj/s5/$dir/egs $dir/egs/storage
    fi
    steps/nnet2/train_tanh_fast.sh --stage $train_stage \
      --num-threads "$num_threads" \
      --parallel-opts "$parallel_opts" \
      --minibatch-size "$minibatch_size" \
      --num-jobs-nnet 8 \
      --samples-per-iter 400000 \
      --mix-up 8000 \
      --initial-learning-rate 0.01 --final-learning-rate 0.001 \
      --num-hidden-layers 4 --hidden-layer-dim 1024 \
      --cmd "$decode_cmd" \
       data/$train_set data/lang exp/tri4b_ali_${train_set} $dir || exit 1
  fi
  
  for test in $test_sets; do
    steps/nnet2/decode.sh --nj 20 --cmd "$decode_cmd" \
      --transform-dir exp/tri4b/decode_tgpr_$test \
      exp/tri4b/graph_tgpr data/$test $dir/decode_tgpr_$test || exit 1;
  done
  
  wait
  
  exit 0