#!/bin/bash
  
  # This is neural net training on top of adapted 40-dimensional features.
  # This is an alternative to the run_5c_gpu.sh that will train faster if you
  # have 8 gpus because it uses more jobs, but the results are slightly worse.
  # [note: possibly we could raise the learning rate and match the run_5c_gpu.sh
  # results.]
  
  
  train_stage=-100
  temp_dir=  # e.g. --temp-dir /export/m1-02/dpovey/kaldi-dan2/egs/wsj/s5/
  parallel_opts="--gpu 1"  # This is suitable for the CLSP network, you'll likely have to change it.
  dir=exp/nnet5c2_gpu
  
  # Note: since we multiplied the num-jobs by 1/4, we halved the
  # learning rate, relative to run_5c.sh
  
  . ././cmd.sh
  . ./path.sh
  ! cuda-compiled && 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
  . utils/parse_options.sh
  
  ( 
  
    if [ ! -z "$temp_dir" ] && [ ! -e $dir/egs ]; then
      mkdir -p $dir
      mkdir -p $temp_dir/$dir/egs
      ln -s $temp_dir/$dir/egs $dir/
    fi
  
    steps/nnet2/train_tanh.sh \
     --num-jobs-nnet 8 --num-threads 1 --parallel-opts "$parallel_opts" \
     --mix-up 8000 \
     --initial-learning-rate 0.0075 --final-learning-rate 0.00075 \
     --num-hidden-layers 4 --hidden-layer-dim 1024 \
     --cmd "$decode_cmd" \
      data/train_si284 data/lang exp/tri4b_ali_si284 $dir || exit 1
    
    steps/nnet2/decode.sh --cmd "$decode_cmd" --nj 10 \
      --transform-dir exp/tri4b/decode_bd_tgpr_dev93 \
       exp/tri4b/graph_bd_tgpr data/test_dev93 $dir/decode_bd_tgpr_dev93
  
    steps/nnet2/decode.sh --cmd "$decode_cmd" --nj 8 \
      --transform-dir exp/tri4b/decode_bd_tgpr_eval92 \
       exp/tri4b/graph_bd_tgpr data/test_eval92 $dir/decode_bd_tgpr_eval92
  )