#!/bin/bash

# Copyright 2012-2014  Johns Hopkins University (Author: Daniel Povey).
#           2013  Xiaohui Zhang
#           2013  Guoguo Chen
#           2014  Vimal Manohar
#           2014  Vijayaditya Peddinti
# Apache 2.0.

# This is a modified version of train_multisplice_accel2.sh in
# steps/nnet2/ for language recognition. The main difference is
# that it uses different get_lda.sh and get_egs2.sh scripts.
#
# The original train_multisplice_accel2.sh was a modified version of
# train_pnorm_multisplice2.sh (still using pnorm).  The "accel" refers to the
# fact that we increase the number of jobs during training (from
# --num-jobs-initial to --num-jobs-final).  We dropped "pnorm" from the name as
# it was getting too long.


# Begin configuration section.
cmd=run.pl
num_epochs=15      # Number of epochs of training;
                   # the number of iterations is worked out from this.
initial_effective_lrate=0.01
final_effective_lrate=0.001
bias_stddev=0.5
pnorm_input_dim=3000
pnorm_output_dim=300
minibatch_size=128 # by default use a smallish minibatch size for neural net
                   # training; this controls instability which would otherwise
                   # be a problem with multi-threaded update.

samples_per_iter=400000 # each iteration of training, see this many samples
                        # per job.  This option is passed to get_egs.sh
num_jobs_initial=1  # Number of neural net jobs to run in parallel at the start of training
num_jobs_final=8   # Number of neural net jobs to run in parallel at the end of training
prior_subset_size=10000 # 10k samples per job, for computing priors.  Should be
                        # more than enough.
num_jobs_compute_prior=10 # these are single-threaded, run on CPU.
get_egs_stage=0
online_ivector_dir=
remove_egs=true  # set to false to disable removing egs.

max_models_combine=20 # The "max_models_combine" is the maximum number of models we give
  # to the final 'combine' stage, but these models will themselves be averages of
  # iteration-number ranges.

shuffle_buffer_size=5000 # This "buffer_size" variable controls randomization of the samples
                # on each iter.  You could set it to 0 or to a large value for complete
                # randomization, but this would both consume memory and cause spikes in
                # disk I/O.  Smaller is easier on disk and memory but less random.  It's
                # not a huge deal though, as samples are anyway randomized right at the start.
                # (the point of this is to get data in different minibatches on different iterations,
                # since in the preconditioning method, 2 samples in the same minibatch can
                # affect each others' gradients.

add_layers_period=2 # by default, add new layers every 2 iterations.
num_hidden_layers=3
stage=-4
exit_stage=-100 # you can set this to terminate the training early.  Exits before running this stage

splice_indexes="layer0/-4:-3:-2:-1:0:1:2:3:4 layer2/-5:-1:3"
# Format : layer<hidden_layer>/<frame_indices>....layer<hidden_layer>/<frame_indices> "
# note: hidden layers which are composed of one or more components,
# so hidden layer indexing is different from component count


io_opts="--max-jobs-run 5" # for jobs with a lot of I/O, limits the number running at one time.   These don't
randprune=4.0 # speeds up LDA.
alpha=4.0 # relates to preconditioning.
update_period=4 # relates to online preconditioning: says how often we update the subspace.
num_samples_history=2000 # relates to online preconditioning
max_change_per_sample=0.075
precondition_rank_in=20  # relates to online preconditioning
precondition_rank_out=80 # relates to online preconditioning

mix_up=0 # Number of components to mix up to (should be > #tree leaves, if
        # specified.)
num_threads=16
parallel_opts="--num-threads 16 --mem 1G"
  # by default we use 16 threads; this lets the queue know.
  # note: parallel_opts doesn't automatically get adjusted if you adjust num-threads.
combine_num_threads=8
combine_parallel_opts="--num-threads 8"  # queue options for the "combine" stage.
cleanup=true
egs_dir=
lda_opts=
lda_dim=
egs_opts=
transform_dir=     # If supplied, overrides alidir
feat_type=  # Can be used to force "raw" features.
align_cmd=              # The cmd that is passed to steps/nnet2/align.sh
align_use_gpu=          # Passed to use_gpu in steps/nnet2/align.sh [yes/no]
realign_times=          # List of times on which we realign.  Each time is
                        # floating point number strictly between 0 and 1, which
                        # will be multiplied by the num-iters to get an iteration
                        # number.
num_jobs_align=30       # Number of jobs for realignment
# End configuration section.
frames_per_eg=8 # to be passed on to get_egs2.sh

trap 'for pid in $(jobs -pr); do kill -KILL $pid; done' INT QUIT TERM

echo "$0 $@"  # Print the command line for logging

if [ -f path.sh ]; then . ./path.sh; fi
. parse_options.sh || exit 1;

if [ $# != 4 ]; then
  echo "Usage: $0 [opts] <data> <lang> <ali-dir> <exp-dir>"
  echo " e.g.: $0 data/train data/lang exp/tri3_ali exp/tri4_nnet"
  echo ""
  echo "Main options (for others, see top of script file)"
  echo "  --config <config-file>                           # config file containing options"
  echo "  --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs."
  echo "  --num-epochs <#epochs|15>                        # Number of epochs of training"
  echo "  --initial-effective-lrate <lrate|0.02> # effective learning rate at start of training."
  echo "  --final-effective-lrate <lrate|0.004>   # effective learning rate at end of training."
  echo "                                                   # data, 0.00025 for large data"
  echo "  --num-hidden-layers <#hidden-layers|2>           # Number of hidden layers, e.g. 2 for 3 hours of data, 4 for 100hrs"
  echo "  --add-layers-period <#iters|2>                   # Number of iterations between adding hidden layers"
  echo "  --mix-up <#pseudo-gaussians|0>                   # Can be used to have multiple targets in final output layer,"
  echo "                                                   # per context-dependent state.  Try a number several times #states."
  echo "  --num-jobs-initial <num-jobs|1>                  # Number of parallel jobs to use for neural net training, at the start."
  echo "  --num-jobs-final <num-jobs|8>                    # Number of parallel jobs to use for neural net training, at the end"
  echo "  --num-threads <num-threads|16>                   # Number of parallel threads per job (will affect results"
  echo "                                                   # as well as speed; may interact with batch size; if you increase"
  echo "                                                   # this, you may want to decrease the batch size."
  echo "  --parallel-opts <opts|\"--num-threads 16 --mem 1G\">      # extra options to pass to e.g. queue.pl for processes that"
  echo "                                                   # use multiple threads... note, you might have to reduce --mem"
  echo "                                                   # versus your defaults, because it gets multiplied by the --num-threads argument."
  echo "  --io-opts <opts|\"--max-jobs-run 10\">                      # Options given to e.g. queue.pl for jobs that do a lot of I/O."
  echo "  --minibatch-size <minibatch-size|128>            # Size of minibatch to process (note: product with --num-threads"
  echo "                                                   # should not get too large, e.g. >2k)."
  echo "  --samples-per-iter <#samples|400000>             # Number of samples of data to process per iteration, per"
  echo "                                                   # process."
  echo "  --splice-indexes <string|layer0/-4:-3:-2:-1:0:1:2:3:4> "
  echo "                                                   # Frame indices used for each splice layer."
  echo "                                                   # Format : layer<hidden_layer_index>/<frame_indices>....layer<hidden_layer>/<frame_indices> "
  echo "                                                   # (note: we splice processed, typically 40-dimensional frames"
  echo "  --lda-dim <dim|''>                               # Dimension to reduce spliced features to with LDA"
  echo "  --realign-epochs <list-of-epochs|''>             # A list of space-separated epoch indices the beginning of which"
  echo "                                                   # realignment is to be done"
  echo "  --align-cmd (utils/run.pl|utils/queue.pl <queue opts>) # passed to align.sh"
  echo "  --align-use-gpu (yes/no)                         # specify is gpu is to be used for realignment"
  echo "  --num-jobs-align <#njobs|30>                     # Number of jobs to perform realignment"
  echo "  --stage <stage|-4>                               # Used to run a partially-completed training process from somewhere in"
  echo "                                                   # the middle."


  exit 1;
fi

data=$1
lang=$2
alidir=$3
dir=$4

if [ ! -z "$realign_times" ]; then
  [ -z "$align_cmd" ] && echo "$0: realign_times specified but align_cmd not specified" && exit 1
  [ -z "$align_use_gpu" ] && echo "$0: realign_times specified but align_use_gpu not specified" && exit 1
fi

# Check some files.
for f in $data/feats.scp $lang/L.fst $alidir/ali.1.gz $alidir/final.mdl $alidir/tree; do
  [ ! -f $f ] && echo "$0: no such file $f" && exit 1;
done


# Set some variables.
num_leaves=`tree-info $alidir/tree 2>/dev/null | grep num-pdfs | awk '{print $2}'` || exit 1
[ -z $num_leaves ] && echo "\$num_leaves is unset" && exit 1
[ "$num_leaves" -eq "0" ] && echo "\$num_leaves is 0" && exit 1

nj=`cat $alidir/num_jobs` || exit 1;  # number of jobs in alignment dir...
# in this dir we'll have just one job.
sdata=$data/split$nj
utils/split_data.sh $data $nj

mkdir -p $dir/log
echo $nj > $dir/num_jobs
cp $alidir/tree $dir

# process the splice_inds string, to get a layer-wise context string
# to be processed by the nnet-components
# this would be mainly used by SpliceComponent|SpliceMaxComponent
python steps/nnet2/make_multisplice_configs.py contexts --splice-indexes "$splice_indexes" $dir || exit -1;
context_string=$(cat $dir/vars) || exit -1
echo $context_string
eval $context_string || exit -1; #
  # initializes variables used by get_lda.sh and get_egs.sh
  # get_lda.sh : first_left_context, first_right_context,
  # get_egs.sh : nnet_left_context & nnet_right_context

extra_opts=()
[ ! -z "$feat_type" ] && extra_opts+=(--feat-type $feat_type)
[ ! -z "$online_ivector_dir" ] && extra_opts+=(--online-ivector-dir $online_ivector_dir)
[ -z "$transform_dir" ] && transform_dir=$alidir
extra_opts+=(--transform-dir $transform_dir)

if [ $stage -le -4 ]; then
  echo "$0: calling get_lda.sh"
  lid/nnet2/get_lda.sh $lda_opts "${extra_opts[@]}" --left-context $first_left_context --right-context $first_right_context --cmd "$cmd" $data $lang $alidir $dir || exit 1;
fi
# these files will have been written by get_lda.sh
feat_dim=$(cat $dir/feat_dim) || exit 1;
ivector_dim=$(cat $dir/ivector_dim) || exit 1;
lda_dim=$(cat $dir/lda_dim) || exit 1;

if [ $stage -le -3 ] && [ -z "$egs_dir" ]; then

  extra_opts+=(--left-context $nnet_left_context )
  extra_opts+=(--right-context $nnet_right_context )
  echo "$0: calling get_egs2.sh"
  lid/nnet2/get_egs2.sh $egs_opts "${extra_opts[@]}" \
      --samples-per-iter $samples_per_iter --stage $get_egs_stage \
      --io-opts "$io_opts" \
      --cmd "$cmd" $egs_opts \
      --frames-per-eg $frames_per_eg \
      $data $alidir $dir/egs || exit 1;
fi

if [ -z $egs_dir ]; then
  egs_dir=$dir/egs
  # confirm that the provided egs_dir has the necessary context
  egs_left_context=$(cat $egs_dir/info/left_context) || exit -1
  egs_right_context=$(cat $egs_dir/info/right_context) || exit -1
  echo $egs_left_context  $nnet_left_context $egs_right_context $nnet_right_context
  ([[ $egs_left_context -lt $nnet_left_context ]] || [[ $egs_right_context -lt $nnet_right_context ]]) &&
    echo "Provided egs_dir $egs_dir does not have sufficient context to train the neural network." && exit -1;
fi

frames_per_eg=$(cat $egs_dir/info/frames_per_eg) || { echo "error: no such file $egs_dir/info/frames_per_eg"; exit 1; }
num_archives=$(cat $egs_dir/info/num_archives) || { echo "error: no such file $egs_dir/info/frames_per_eg"; exit 1; }

# num_archives_expanded considers each separate label-position from
# 0..frames_per_eg-1 to be a separate archive.
num_archives_expanded=$[$num_archives*$frames_per_eg]

[ $num_jobs_initial -gt $num_jobs_final ] && \
  echo "$0: --initial-num-jobs cannot exceed --final-num-jobs" && exit 1;

[ $num_jobs_final -gt $num_archives_expanded ] && \
  echo "$0: --final-num-jobs cannot exceed #archives $num_archives_expanded." && exit 1;

if ! [ $num_hidden_layers -ge 1 ]; then
  echo "Invalid num-hidden-layers $num_hidden_layers"
  exit 1
fi

if [ $stage -le -2 ]; then
  echo "$0: initializing neural net";
  lda_mat=$dir/lda.mat
  tot_input_dim=$[$feat_dim+$ivector_dim]

  online_preconditioning_opts="alpha=$alpha num-samples-history=$num_samples_history update-period=$update_period rank-in=$precondition_rank_in rank-out=$precondition_rank_out max-change-per-sample=$max_change_per_sample"

  initial_lrate=$(perl -e "print ($initial_effective_lrate*$num_jobs_initial);")

  # create the config files for nnet initialization
  python steps/nnet2/make_multisplice_configs.py  \
    --splice-indexes "$splice_indexes"  \
    --total-input-dim $tot_input_dim  \
    --ivector-dim $ivector_dim  \
    --lda-mat "$lda_mat"  \
    --lda-dim $lda_dim  \
    --pnorm-input-dim $pnorm_input_dim  \
    --pnorm-output-dim  $pnorm_output_dim \
    --online-preconditioning-opts "$online_preconditioning_opts"  \
    --initial-learning-rate $initial_lrate \
    --bias-stddev  $bias_stddev  \
    --num-hidden-layers $num_hidden_layers \
    --num-targets  $num_leaves  \
    configs  $dir || exit -1;

  $cmd $dir/log/nnet_init.log \
    nnet-am-init $alidir/tree $lang/topo "nnet-init $dir/nnet.config -|" \
    $dir/0.mdl || exit 1;
fi

if [ $stage -le -1 ]; then
  echo "Training transition probabilities and setting priors"
  $cmd $dir/log/train_trans.log \
    nnet-train-transitions $dir/0.mdl "ark:gunzip -c $alidir/ali.*.gz|" $dir/0.mdl \
    || exit 1;
fi

# set num_iters so that as close as possible, we process the data $num_epochs
# times, i.e. $num_iters*$avg_num_jobs) == $num_epochs*$num_archives_expanded,
# where avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.

num_archives_to_process=$[$num_epochs*$num_archives_expanded]
num_archives_processed=0
num_iters=$[($num_archives_to_process*2)/($num_jobs_initial+$num_jobs_final)]

! [ $num_iters -gt $[$finish_add_layers_iter+2] ] \
  && echo "$0: Insufficient epochs" && exit 1

finish_add_layers_iter=$[$num_hidden_layers * $add_layers_period]


# mix up at the iteration where we've processed about half the data; this keeps
# the overall training procedure fairly invariant to the number of initial and
# final jobs.
# j = initial, k = final, n = num-iters, x = half-of-data epoch,
# p is proportion of data we want to process (e.g. p=0.5 here).
# solve for x if the amount of data processed by epoch x is p
# times the amount by iteration n.
# put this in wolfram alpha:
# solve { x*j + (k-j)*x*x/(2*n) = p * (j*n + (k-j)*n/2), {x} }
# got: x = (j n-sqrt(-n^2 (j^2 (p-1)-k^2 p)))/(j-k) and j!=k and n!=0
# simplified manually to: n * (sqrt(((1-p)j^2 + p k^2)/2) - j)/(j-k)
mix_up_iter=$(perl -e '($j,$k,$n,$p)=@ARGV; print int(0.5 + ($j==$k ? $n*$p : $n*(sqrt((1-$p)*$j*$j+$p*$k*$k)-$j)/($k-$j))); ' $num_jobs_initial $num_jobs_final $num_iters 0.5)
! [ $mix_up_iter -gt $finish_add_layers_iter ] && \
  echo "Mix-up-iter is $mix_up_iter, should be greater than $finish_add_layers_iter -> add more epochs?" \
  && exit 1;

echo "$0: Will train for $num_epochs epochs = $num_iters iterations"
[ $mix_up -gt 0 ] && echo "$0: Will mix up on iteration $mix_up_iter"

if [ $num_threads -eq 1 ]; then
  parallel_suffix="-simple" # this enables us to use GPU code if
                         # we have just one thread.
  parallel_train_opts=
  if ! cuda-compiled; then
    echo "$0: WARNING: you are running with one thread but you have not compiled"
    echo "   for CUDA.  You may be running a setup optimized for GPUs.  If you have"
    echo "   GPUs and have nvcc installed, go to src/ and do ./configure; make"
  fi
else
  parallel_suffix="-parallel"
  parallel_train_opts="--num-threads=$num_threads"
fi


approx_iters_per_epoch_final=$[$num_archives_expanded/$num_jobs_final]
# First work out how many models we want to combine over in the final
# nnet-combine-fast invocation.  This equals
# min(max(max_models_combine, approx_iters_per_epoch_final),
#     2/3 * iters_after_mixup)
num_models_combine=$max_models_combine
if [ $num_models_combine -lt $approx_iters_per_epoch_final ]; then
   num_models_combine=$approx_iters_per_epoch_final
fi
iters_after_mixup_23=$[(($num_iters-$mix_up_iter-1)*2)/3]
if [ $num_models_combine -gt $iters_after_mixup_23 ]; then
  num_models_combine=$iters_after_mixup_23
fi
first_model_combine=$[$num_iters-$num_models_combine+1]

x=0


for realign_time in $realign_times; do
  # Work out the iterations on which we will re-align, if the --realign-times
  # option was used.  This is slightly approximate.
  ! perl -e "exit($realign_time > 0.0 && $realign_time < 1.0 ? 0:1);" && \
    echo "Invalid --realign-times option $realign_times: elements must be strictly between 0 and 1.";
  # the next formula is based on the one for mix_up_iter above.
  realign_iter=$(perl -e '($j,$k,$n,$p)=@ARGV; print int(0.5 + ($j==$k ? $n*$p : $n*(sqrt((1-$p)*$j*$j+$p*$k*$k)-$j)/($k-$j))); ' $num_jobs_initial $num_jobs_final $num_iters $realign_time) || exit 1;
  realign_this_iter[$realign_iter]=$realign_time
done

cur_egs_dir=$egs_dir

while [ $x -lt $num_iters ]; do
  [ $x -eq $exit_stage ] && echo "$0: Exiting early due to --exit-stage $exit_stage" && exit 0;

  this_num_jobs=$(perl -e "print int(0.5+$num_jobs_initial+($num_jobs_final-$num_jobs_initial)*$x/$num_iters);")

  ilr=$initial_effective_lrate; flr=$final_effective_lrate; np=$num_archives_processed; nt=$num_archives_to_process;
  this_learning_rate=$(perl -e "print (($x + 1 >= $num_iters ? $flr : $ilr*exp($np*log($flr/$ilr)/$nt))*$this_num_jobs);");

  echo "On iteration $x, learning rate is $this_learning_rate."

  if [ ! -z "${realign_this_iter[$x]}" ]; then
    prev_egs_dir=$cur_egs_dir
    cur_egs_dir=$dir/egs_${realign_this_iter[$x]}
  fi

  if [ $x -ge 0 ] && [ $stage -le $x ]; then
    if [ ! -z "${realign_this_iter[$x]}" ]; then
      time=${realign_this_iter[$x]}

      echo "Getting average posterior for purposes of adjusting the priors."
      # Note: this just uses CPUs, using a smallish subset of data.
      # always use the first egs archive, which makes the script simpler;
      # we're using different random subsets of it.
      rm $dir/post.$x.*.vec 2>/dev/null
      $cmd JOB=1:$num_jobs_compute_prior $dir/log/get_post.$x.JOB.log \
        nnet-copy-egs --srand=JOB --frame=random ark:$prev_egs_dir/egs.1.ark ark:- \| \
        nnet-subset-egs --srand=JOB --n=$prior_subset_size ark:- ark:- \| \
        nnet-compute-from-egs "nnet-to-raw-nnet $dir/$x.mdl -|" ark:- ark:- \| \
        matrix-sum-rows ark:- ark:- \| vector-sum ark:- $dir/post.$x.JOB.vec || exit 1;

      sleep 3;  # make sure there is time for $dir/post.$x.*.vec to appear.

      $cmd $dir/log/vector_sum.$x.log \
        vector-sum $dir/post.$x.*.vec $dir/post.$x.vec || exit 1;
      rm $dir/post.$x.*.vec;

      echo "Re-adjusting priors based on computed posteriors"
      $cmd $dir/log/adjust_priors.$x.log \
        nnet-adjust-priors $dir/$x.mdl $dir/post.$x.vec $dir/$x.mdl || exit 1;

      sleep 2

      steps/nnet2/align.sh --nj $num_jobs_align --cmd "$align_cmd" --use-gpu $align_use_gpu \
        --transform-dir "$transform_dir" --online-ivector-dir "$online_ivector_dir" \
        --iter $x $data $lang $dir $dir/ali_$time || exit 1

      lid/nnet2/relabel_egs2.sh --cmd "$cmd" --iter $x $dir/ali_$time \
        $prev_egs_dir $cur_egs_dir || exit 1

      if $cleanup && [[ $prev_egs_dir =~ $dir/egs* ]]; then
        steps/nnet2/remove_egs.sh $prev_egs_dir
      fi
    fi

    # Set off jobs doing some diagnostics, in the background.
    # Use the egs dir from the previous iteration for the diagnostics
    $cmd $dir/log/compute_prob_valid.$x.log \
      nnet-compute-prob $dir/$x.mdl ark:$cur_egs_dir/valid_diagnostic.egs &
    $cmd $dir/log/compute_prob_train.$x.log \
      nnet-compute-prob $dir/$x.mdl ark:$cur_egs_dir/train_diagnostic.egs &
    if [ $x -gt 0 ] && [ ! -f $dir/log/mix_up.$[$x-1].log ]; then
      $cmd $dir/log/progress.$x.log \
        nnet-show-progress --use-gpu=no $dir/$[$x-1].mdl $dir/$x.mdl \
        ark:$cur_egs_dir/train_diagnostic.egs '&&' \
        nnet-am-info $dir/$x.mdl &
    fi

    echo "Training neural net (pass $x)"

    if [ $x -gt 0 ] && \
      [ $x -le $[($num_hidden_layers-1)*$add_layers_period] ] && \
      [ $[$x%$add_layers_period] -eq 0 ]; then
      do_average=false # if we've just mixed up, don't do averaging take the best.
      cur_num_hidden_layers=$[$x/$add_layers_period];
      mdl="nnet-init --srand=$x $dir/hidden_${cur_num_hidden_layers}.config - | nnet-insert $dir/$x.mdl - - | nnet-am-copy --learning-rate=$this_learning_rate - -|"
    else
      do_average=true
      if [ $x -eq 0 ]; then do_average=false; fi # on iteration 0, pick the best, don't average.
      mdl="nnet-am-copy --learning-rate=$this_learning_rate $dir/$x.mdl -|"
    fi
    if $do_average; then
      this_minibatch_size=$minibatch_size
    else
      # on iteration zero or when we just added a layer, use a smaller minibatch
      # size and just one job: the model-averaging doesn't seem to be helpful
      # when the model is changing too fast (i.e. it worsens the objective
      # function), and the smaller minibatch size will help to keep
      # the update stable.
      this_minibatch_size=$[$minibatch_size/2];
    fi

    rm $dir/.error 2>/dev/null


    ( # this sub-shell is so that when we "wait" below,
      # we only wait for the training jobs that we just spawned,
      # not the diagnostic jobs that we spawned above.

      # We can't easily use a single parallel SGE job to do the main training,
      # because the computation of which archive and which --frame option
      # to use for each job is a little complex, so we spawn each one separately.
      for n in $(seq $this_num_jobs); do
        k=$[$num_archives_processed + $n - 1]; # k is a zero-based index that we'll derive
                                               # the other indexes from.
        archive=$[($k%$num_archives)+1]; # work out the 1-based archive index.
        frame=$[(($k/$num_archives)%$frames_per_eg)]; # work out the 0-based frame
        # index; this increases more slowly than the archive index because the
        # same archive with different frame indexes will give similar gradients,
        # so we want to separate them in time.

        $cmd $parallel_opts $dir/log/train.$x.$n.log \
          nnet-train$parallel_suffix $parallel_train_opts \
          --minibatch-size=$this_minibatch_size --srand=$x "$mdl" \
          "ark:nnet-copy-egs --frame=$frame ark:$cur_egs_dir/egs.$archive.ark ark:-|nnet-shuffle-egs --buffer-size=$shuffle_buffer_size --srand=$x ark:- ark:-|" \
          $dir/$[$x+1].$n.mdl || touch $dir/.error &
      done
      wait
    )
    # the error message below is not that informative, but $cmd will
    # have printed a more specific one.
    [ -f $dir/.error ] && echo "$0: error on iteration $x of training" && exit 1;

    nnets_list=
    for n in `seq 1 $this_num_jobs`; do
      nnets_list="$nnets_list $dir/$[$x+1].$n.mdl"
    done

    if $do_average; then
      # average the output of the different jobs.
      $cmd $dir/log/average.$x.log \
        nnet-am-average $nnets_list $dir/$[$x+1].mdl || exit 1;
    else
      # choose the best from the different jobs.
      n=$(perl -e '($nj,$pat)=@ARGV; $best_n=1; $best_logprob=-1.0e+10; for ($n=1;$n<=$nj;$n++) {
          $fn = sprintf($pat,$n); open(F, "<$fn") || die "Error opening log file $fn";
          undef $logprob; while (<F>) { if (m/log-prob-per-frame=(\S+)/) { $logprob=$1; } }
          close(F); if (defined $logprob && $logprob > $best_logprob) { $best_logprob=$logprob;
          $best_n=$n; } } print "$best_n\n"; ' $num_jobs_nnet $dir/log/train.$x.%d.log) || exit 1;
      [ -z "$n" ] && echo "Error getting best model" && exit 1;
      cp $dir/$[$x+1].$n.mdl $dir/$[$x+1].mdl || exit 1;
    fi

    if [ "$mix_up" -gt 0 ] && [ $x -eq $mix_up_iter ]; then
      # mix up.
      echo Mixing up from $num_leaves to $mix_up components
      $cmd $dir/log/mix_up.$x.log \
        nnet-am-mixup --min-count=10 --num-mixtures=$mix_up \
        $dir/$[$x+1].mdl $dir/$[$x+1].mdl || exit 1;
    fi
    rm $nnets_list
    [ ! -f $dir/$[$x+1].mdl ] && exit 1;
    if [ -f $dir/$[$x-1].mdl ] && $cleanup && \
       [ $[($x-1)%100] -ne 0  ] && [ $[$x-1] -lt $first_model_combine ]; then
      rm $dir/$[$x-1].mdl
    fi
  fi
  x=$[$x+1]
  num_archives_processed=$[$num_archives_processed+$this_num_jobs]
done


if [ $stage -le $num_iters ]; then
  echo "Doing final combination to produce final.mdl"

  # Now do combination.
  nnets_list=()
  # the if..else..fi statement below sets 'nnets_list'.
  if [ $max_models_combine -lt $num_models_combine ]; then
    # The number of models to combine is too large, e.g. > 20.  In this case,
    # each argument to nnet-combine-fast will be an average of multiple models.
    cur_offset=0 # current offset from first_model_combine.
    for n in $(seq $max_models_combine); do
      next_offset=$[($n*$num_models_combine)/$max_models_combine]
      sub_list=""
      for o in $(seq $cur_offset $[$next_offset-1]); do
        iter=$[$first_model_combine+$o]
        mdl=$dir/$iter.mdl
        [ ! -f $mdl ] && echo "Expected $mdl to exist" && exit 1;
        sub_list="$sub_list $mdl"
      done
      nnets_list[$[$n-1]]="nnet-am-average $sub_list - |"
      cur_offset=$next_offset
    done
  else
    nnets_list=
    for n in $(seq 0 $[num_models_combine-1]); do
      iter=$[$first_model_combine+$n]
      mdl=$dir/$iter.mdl
      [ ! -f $mdl ] && echo "Expected $mdl to exist" && exit 1;
      nnets_list[$n]=$mdl
    done
  fi


  # Below, use --use-gpu=no to disable nnet-combine-fast from using a GPU, as
  # if there are many models it can give out-of-memory error; set num-threads to 8
  # to speed it up (this isn't ideal...)
  num_egs=`nnet-copy-egs ark:$cur_egs_dir/combine.egs ark:/dev/null 2>&1 | tail -n 1 | awk '{print $NF}'`
  mb=$[($num_egs+$combine_num_threads-1)/$combine_num_threads]
  [ $mb -gt 512 ] && mb=512
  # Setting --initial-model to a large value makes it initialize the combination
  # with the average of all the models.  It's important not to start with a
  # single model, or, due to the invariance to scaling that these nonlinearities
  # give us, we get zero diagonal entries in the fisher matrix that
  # nnet-combine-fast uses for scaling, which after flooring and inversion, has
  # the effect that the initial model chosen gets much higher learning rates
  # than the others.  This prevents the optimization from working well.
  $cmd $combine_parallel_opts $dir/log/combine.log \
    nnet-combine-fast --initial-model=100000 --num-lbfgs-iters=40 --use-gpu=no \
      --num-threads=$combine_num_threads \
      --verbose=3 --minibatch-size=$mb "${nnets_list[@]}" ark:$cur_egs_dir/combine.egs \
      $dir/final.mdl || exit 1;

  # Normalize stddev for affine or block affine layers that are followed by a
  # pnorm layer and then a normalize layer.
  $cmd $dir/log/normalize.log \
    nnet-normalize-stddev $dir/final.mdl $dir/final.mdl || exit 1;

  # Compute the probability of the final, combined model with
  # the same subset we used for the previous compute_probs, as the
  # different subsets will lead to different probs.
  $cmd $dir/log/compute_prob_valid.final.log \
    nnet-compute-prob $dir/final.mdl ark:$cur_egs_dir/valid_diagnostic.egs &
  $cmd $dir/log/compute_prob_train.final.log \
    nnet-compute-prob $dir/final.mdl ark:$cur_egs_dir/train_diagnostic.egs &
fi

if [ $stage -le $[$num_iters+1] ]; then
  echo "Getting average posterior for purposes of adjusting the priors."
  # Note: this just uses CPUs, using a smallish subset of data.
  rm $dir/post.$x.*.vec 2>/dev/null
  $cmd JOB=1:$num_jobs_compute_prior $dir/log/get_post.$x.JOB.log \
    nnet-copy-egs --frame=random --srand=JOB ark:$cur_egs_dir/egs.1.ark ark:- \| \
    nnet-subset-egs --srand=JOB --n=$prior_subset_size ark:- ark:- \| \
    nnet-compute-from-egs "nnet-to-raw-nnet $dir/final.mdl -|" ark:- ark:- \| \
    matrix-sum-rows ark:- ark:- \| vector-sum ark:- $dir/post.$x.JOB.vec || exit 1;

  sleep 3;  # make sure there is time for $dir/post.$x.*.vec to appear.

  $cmd $dir/log/vector_sum.$x.log \
   vector-sum $dir/post.$x.*.vec $dir/post.$x.vec || exit 1;

  rm $dir/post.$x.*.vec;

  echo "Re-adjusting priors based on computed posteriors"
  $cmd $dir/log/adjust_priors.final.log \
    nnet-adjust-priors $dir/final.mdl $dir/post.$x.vec $dir/final.mdl || exit 1;
fi


if [ ! -f $dir/final.mdl ]; then
  echo "$0: $dir/final.mdl does not exist."
  # we don't want to clean up if the training didn't succeed.
  exit 1;
fi

sleep 2

echo Done

if $cleanup; then
  echo Cleaning up data
  if $remove_egs && [[ $cur_egs_dir =~ $dir/egs* ]]; then
    steps/nnet2/remove_egs.sh $cur_egs_dir
  fi

  echo Removing most of the models
  for x in `seq 0 $num_iters`; do
    if [ $[$x%100] -ne 0 ] && [ $x -ne $num_iters ] && [ -f $dir/$x.mdl ]; then
       # delete all but every 100th model; don't delete the ones which combine to form the final model.
      rm $dir/$x.mdl
    fi
  done
fi