#!/usr/bin/env bash

# This training script is modified from the original run.sh

# Copyright 2019 Mobvoi Inc. All Rights Reserved.
. ./path.sh || exit 1;

# Use this to control how many gpu you use, It's 1-gpu training if you specify
# just 1gpu, otherwise it's is multiple gpu training based on DDP in pytorch
export CUDA_VISIBLE_DEVICES="0,1,2,3"
# The NCCL_SOCKET_IFNAME variable specifies which IP interface to use for nccl
# communication. More details can be found in
# https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html
# export NCCL_SOCKET_IFNAME=ens4f1
export NCCL_DEBUG=INFO
stage=4 # start from 0 if you need to start from data preparation
stop_stage=4

# The num of machines(nodes) for multi-machine training, 1 is for one machine.
# NFS is required if num_nodes > 1.
num_nodes=1

# The rank of each node or machine, which ranges from 0 to `num_nodes - 1`.
# You should set the node_rank=0 on the first machine, set the node_rank=1
# on the second machine, and so on.
node_rank=0
# The aishell dataset location, please change this to your own path
# make sure of using absolute path. DO-NOT-USE relatvie path!
data=/data/conformer/train/
data_url=www.openslr.org/resources/33

dict=data/dict/lang_char.txt

# data_type can be `raw` or `shard`. Typically, raw is used for small dataset,
# `shard` is used for large dataset which is over 1k hours, and `shard` is
# faster on reading data and training.
data_type=raw
num_utts_per_shard=1000

train_set=train
# Optional train_config
# 1. conf/train_transformer.yaml: Standard transformer
# 2. conf/train_conformer.yaml: Standard conformer
# 3. conf/train_unified_conformer.yaml: Unified dynamic chunk causal conformer
# 4. conf/train_unified_transformer.yaml: Unified dynamic chunk transformer
# 5. conf/train_u2++_conformer.yaml: U2++ conformer
# 6. conf/train_u2++_transformer.yaml: U2++ transformer
train_config=conf/train_conformer.yaml
cmvn=true
dir=exp/conformer
checkpoint=

# use average_checkpoint will get better result
average_checkpoint=false
decode_checkpoint=$dir/final.pt
average_num=30
decode_modes="attention_rescoring"
log_dir=

. tools/parse_options.sh || exit 1;

if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
  mkdir -p $dir
  # You have to rm `INIT_FILE` manually when you resume or restart a
  # multi-machine training.
  INIT_FILE=$dir/ddp_init
  init_method=file://$(readlink -f $INIT_FILE)
  echo "$0: init method is $init_method"
  num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
  # Use "nccl" if it works, otherwise use "gloo"
  dist_backend="nccl"
  world_size=`expr $num_gpus \* $num_nodes`
  echo "total gpus is: $world_size"
  cmvn_opts=
  $cmvn && cp data/${train_set}/global_cmvn $dir
  $cmvn && cmvn_opts="--cmvn ${dir}/global_cmvn"

  # train.py rewrite $train_config to $dir/train.yaml with model input
  # and output dimension, and $dir/train.yaml will be used for inference
  # and export.
  for ((i = 0; i < $num_gpus; ++i)); do
  {
    gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
    # Rank of each gpu/process used for knowing whether it is
    # the master of a worker.
    rank=`expr $node_rank \* $num_gpus + $i`
    test_dir=$dir/test_${decode_modes}
    python wenet/bin/train.py --gpu $gpu_id \
      --config $train_config \
      --data_type $data_type \
      --symbol_table $dict \
      --train_data data/$train_set/data.list \
      --cv_data data/dev/data.list \
      ${checkpoint:+--checkpoint $checkpoint} \
      --model_dir $dir \
      --ddp.init_method $init_method \
      --ddp.world_size $world_size \
      --ddp.rank $rank \
      --ddp.dist_backend $dist_backend \
      --num_workers 1 \
      $cmvn_opts \
      --pin_memory \
      --val_ref_file data/test/text \
      --val_hyp_file $test_dir/text \
      --log_dir $log_dir
  } &
  done
  wait
fi