Skip to content

S
S2T
Commit 478c694b by xuchen
Options

shell

parent e248f2f0
正在显示 包含 70 行增加 和 121 行删除
egs/mustc/asr/decode.sh
#!/usr/bin/env bash
gpu_num=1
gpu_num=0
data_dir=
test_subset=(dev tst-COMMON)
......@@ -15,12 +15,12 @@ ctc_infer=0
n_average=10
beam_size=5
len_penalty=1.0
max_tokens=80000
max_tokens=50000
dec_model=checkpoint_best.pt
cmd="./run.sh
--stage 3
--stop_stage 3
--stage 2
--stop_stage 2
--gpu_num ${gpu_num}
--exp_name ${exp_name}
--n_average ${n_average}
......
egs/mustc/asr/run.sh
......@@ -2,8 +2,7 @@
# Processing MuST-C Datasets
# Copyright 2021 Natural Language Processing Laboratory
# Xu Chen (xuchenneu@163.com)
# Copyright 2021 Chen Xu (xuchennlp@outlook.com)
# Set bash to 'debug' mode, it will exit on :
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
......@@ -16,22 +15,21 @@ eval=1
time=$(date "+%m%d_%H%M")
stage=1
stop_stage=4
stop_stage=2
######## hardware ########
# devices
######## Hardware ########
# Devices
device=(0)
gpu_num=8
update_freq=1
hdfs_get=0
root_dir=/opt/tiger
data_root_dir=/mnt/bn/nas-xc-1
code_dir=${root_dir}/s2t
pwd_dir=$PWD
root_dir=${ST_ROOT}
data_root_dir=${root_dir}
code_dir=${root_dir}/S2T
# dataset
# Dataset
src_lang=en
tgt_lang=de
dataset=must_c
......@@ -63,24 +61,22 @@ valid_split=dev
test_split=tst-COMMON
test_subset=dev,tst-COMMON
# exp
# Exp
sub_tag=
exp_prefix=$(date "+%m%d")
# exp_subfix=${ARNOLD_JOB_ID}_${ARNOLD_TASK_ID}_${ARNOLD_TRIAL_ID}
extra_tag=
extra_parameter=
exp_tag=baseline
exp_name=
# config
# Training Settings
train_config=base,ctc
data_config=config.yaml
# training setting
fp16=1
max_tokens=40000
step_valid=0
data_config=config.yaml
# decoding setting
# Decoding Settings
cer=0
ctc_infer=0
ctc_self_ensemble=0
......@@ -92,6 +88,7 @@ len_penalty=1.0
infer_score=0
infer_parameters=
# Parsing Options
if [[ ${speed_perturb} -eq 1 ]]; then
data_dir=${data_dir}_sp
exp_prefix=${exp_prefix}_sp
......@@ -124,19 +121,6 @@ if [[ ! -d ${data_dir} ]]; then
exit
fi
# setup nccl envs
export NCCL_IB_DISABLE=0
export NCCL_IB_HCA=$ARNOLD_RDMA_DEVICE:1
export NCCL_IB_GID_INDEX=3
export NCCL_SOCKET_IFNAME=eth0
HOSTS=$ARNOLD_WORKER_HOSTS
HOST=(${HOSTS//,/ })
HOST_SPLIT=(${HOST//:/ })
PORT=${HOST_SPLIT[1]}
INIT_METHOD="tcp://${ARNOLD_WORKER_0_HOST}:${ARNOLD_WORKER_0_PORT}"
DIST_RANK=$((ARNOLD_ID * ARNOLD_WORKER_GPU))
export PATH=$PATH:${code_dir}/scripts
. ./local/parse_options.sh || exit 1;
......@@ -150,21 +134,27 @@ if [[ -z ${exp_name} ]]; then
exp_name=${exp_name}_${exp_subfix}
fi
fi
model_dir=${code_dir}/checkpoints/${data_model_subfix}/${exp_name}
echo "stage: $stage"
echo "stop_stage: $stop_stage"
ckpt_dir=${root_dir}/checkpoints/
model_dir=${root_dir}/checkpoints/${data_model_subfix}/${sub_tag}/${exp_name}
# Start
cd ${code_dir}
echo "Start Stage: $stage"
echo "Stop Stage: $stop_stage"
if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then
echo "Default Stage: env configure"
pip3 install -e ${code_dir}
fi
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
echo "stage -1: Data Download"
# pass
echo "Stage -1: Data Download"
fi
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
### Task dependent. You have to make data the following preparation part by yourself.
### But you can utilize Kaldi recipes in most cases
echo "stage 0: ASR Data Preparation"
echo "Stage 0: ASR Data Preparation"
if [[ ! -e ${data_dir} ]]; then
mkdir -p ${data_dir}
fi
......@@ -205,32 +195,8 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
[[ $eval -eq 1 ]] && eval ${cmd}
fi
if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then
echo "default stage: env configure"
pip3 install -e ${code_dir} -i https://bytedpypi.byted.org/simple --no-build-isolation --default-timeout=10000
fi
if [[ -d /mnt/bn/nas-xc-1/checkpoints && ! -d ${code_dir}/checkpoints ]]; then
ln -s /mnt/bn/nas-xc-1/checkpoints ${code_dir}
fi
# if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
if [ ${hdfs_get} -eq 1 ]; then
ln_data_dir=`echo ${data_dir} | sed -e "s#${data_root_dir}#${code_dir}#"`
echo ${ln_data_dir}
mkdir -p ${ln_data_dir}
ln -s ${data_dir}/../* ${ln_data_dir}
rm -r ${ln_data_dir}
hdfs_path=`echo ${data_dir} | sed -e "s#${data_root_dir}#hdfs://haruna/home/byte_arnold_lq_mlnlc/user/xuchen/#"`
hdfs dfs -get ${hdfs_path} ${ln_data_dir}
sed -i -e "s#${data_root_dir}#${code_dir}#" ${ln_data_dir}/config*
data_dir=${ln_data_dir}
fi
# fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
echo "stage 2: ASR Network Training"
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "Stage 1: Network Training"
[[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
......@@ -240,6 +206,7 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
export CUDA_VISIBLE_DEVICES=${device}
fi
echo -e "data=${data_dir} model=${model_dir}"
......@@ -327,22 +294,17 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
echo -e "\033[34mRun command: \n${cmd} \033[0m"
# save info
log=./history.log
log=${ckpt_dir}/history.log
echo "${time} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
tail -n 50 ${log} > tmp.log
mv tmp.log $log
# export CUDA_VISIBLE_DEVICES=${device}
log=${model_dir}/train.log
cmd="${cmd} 2>&1 | tee -a ${log}"
#cmd="nohup ${cmd} >> ${log} 2>&1 &"
if [[ $eval -eq 1 ]]; then
# tensorboard
if [[ -z ${ARNOLD_TENSORBOARD_CURRENT_PORT} ]]; then
port=6666
else
port=${ARNOLD_TENSORBOARD_CURRENT_PORT}
fi
port=6666
tensorboard --logdir ${model_dir} --port ${port} --bind_all &
echo "${cmd}" > ${model_dir}/cmd
......@@ -352,8 +314,8 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
fi
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
echo "stage 3: ASR Decoding"
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
echo "Stage 2: Decoding"
if [[ ${n_average} -ne 1 ]]; then
# Average models
dec_model=avg_${n_average}_checkpoint.pt
......@@ -377,18 +339,18 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
export CUDA_VISIBLE_DEVICES=${device}
fi
# export CUDA_VISIBLE_DEVICES=${device}
suffix=beam${beam_size}_alpha${len_penalty}_tokens${max_tokens}
if [[ ${n_average} -ne 1 ]]; then
suffix=${suffix}_${n_average}
fi
if [[ -n ${cer} && ${cer} -eq 1 ]]; then
suffix=${suffix}_cer
else
suffix=${suffix}_wer
fi
if [[ ${n_average} -ne 1 ]]; then
suffix=${suffix}_${n_average}
fi
if [[ ${infer_score} -eq 1 ]]; then
suffix=${suffix}_score
fi
......@@ -435,9 +397,9 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
cd ${code_dir}
if [[ $eval -eq 1 ]]; then
src_ctc_file=translation-${subset}.txt.ctc
if [[ -f ${model_dir}/${src_ctc_file} ]]; then
rm ${model_dir}/${src_ctc_file}
ctc_file=translation-${subset}.ctc
if [[ ${ctc_infer} -eq 1 && -f ${model_dir}/${ctc_file} ]]; then
rm ${model_dir}/${ctc_file}
fi
eval $cmd
......@@ -448,33 +410,34 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
cd ${pwd_dir}
if [[ -f ${model_dir}/enc_dump ]]; then
mv ${model_dir}/enc_dump ${model_dir}/${subset}-${suffix}-enc-dump
mv ${model_dir}/enc_dump ${model_dir}/dump-${subset}-enc-${suffix}
fi
if [[ -f ${model_dir}/dec_dump ]]; then
mv ${model_dir}/dec_dump ${model_dir}/${subset}-${suffix}-dec-dump
mv ${model_dir}/dec_dump ${model_dir}/dump-${subset}-dec-${suffix}
fi
trans_file=translation-${subset}-${suffix}.txt
if [[ ${ctc_infer} -eq 1 && -f ${model_dir}/${src_ctc_file} ]]; then
if [[ ${ctc_infer} -eq 1 && -f ${model_dir}/${ctc_file} ]]; then
ref_file=${model_dir}/${subset}.${src_lang}
if [[ ! -f ${ref_file} ]]; then
python3 ./local/extract_txt_from_tsv.py ${data_dir}/${subset}.tsv ${ref_file} "src_text"
fi
if [[ -f ${ref_file} ]]; then
src_ctc=$(mktemp -t temp.record.XXXXXX)
ctc=$(mktemp -t temp.record.XXXXXX)
cd ./local
./cal_wer.sh ${model_dir} ${subset} ${trans_file} ${src_ctc_file} ${ref_file} > ${src_ctc}
./cal_wer.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} > ${ctc}
cd ..
echo "CTC WER" >> ${result_file}
tail -n 2 ${src_ctc} >> ${result_file}
tail -n 2 ${ctc} >> ${result_file}
src_bleu=$(mktemp -t temp.record.XXXXXX)
cd local
./cal_ctc_bleu.sh ${model_dir} ${subset} ${trans_file} ${src_ctc_file} ${ref_file} ${tokenizer} ${src_lang} > ${src_bleu}
./cal_ctc_bleu.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} ${tokenizer} ${src_lang} > ${src_bleu}
cd ..
cat ${src_bleu} >> ${result_file}
rm ${src_ctc} ${src_bleu}
rm ${ctc} ${src_bleu}
else
echo "No reference for source language."
fi
......@@ -484,11 +447,3 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
echo
cat ${result_file}
fi
# if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# cd ${fairseq_dir}
# echo "Stage 4: Upload model and log"
# echo "Path: hdfs://haruna/home/byte_arnold_lq_mlnlc/user/xuchen/s2t/checkpoints/${data_model_subfix}/${exp_name}"
# hdfs dfs -mkdir -p hdfs://haruna/home/byte_arnold_lq_mlnlc/user/xuchen/s2t/checkpoints/${data_model_subfix}
# hdfs dfs -put -f ${model_dir} hdfs://haruna/home/byte_arnold_lq_mlnlc/user/xuchen/s2t/checkpoints/${data_model_subfix}
# fi
egs/mustc/st/run.sh
......@@ -2,7 +2,7 @@
# Processing MuST-C Datasets
# Copyright 2021 Chen Xu (xuchenneu@163.com)
# Copyright 2021 Chen Xu (xuchennlp@outlook.com)
# Set bash to 'debug' mode, it will exit on :
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
......@@ -17,19 +17,19 @@ time=$(date "+%m%d_%H%M")
stage=1
stop_stage=2
######## hardware ########
# devices
######## Hardware ########
# Devices
device=(0)
gpu_num=8
update_freq=1
pwd_dir=$PWD
root_dir=${pwd_dir}/../../../..
data_root_dir=${root_dir}/data
root_dir=${pwd_dir}/../../../../
data_root_dir=${root_dir}
code_dir=${root_dir}/S2T
# dataset
# Dataset
src_lang=en
tgt_lang=de
dataset=must_c
......@@ -63,7 +63,7 @@ valid_split=dev
test_split=tst-COMMON
test_subset=dev,tst-COMMON
# exp
# Exp
sub_tag=
exp_prefix=$(date "+%m%d")
extra_tag=
......@@ -71,16 +71,14 @@ extra_parameter=
exp_tag=baseline
exp_name=
# config
# Training Settings
train_config=base,ctc
# training setting
fp16=1
max_tokens=40000
step_valid=0
bleu_valid=0
# decoding setting
# Decoding Settings
sacrebleu=1
dec_model=checkpoint_best.pt
ctc_infer=0
......@@ -90,6 +88,7 @@ len_penalty=1.0
infer_score=0
infer_parameters=
# Parsing Options
if [[ ${share_dict} -eq 1 ]]; then
data_config=config_share.yaml
else
......@@ -136,12 +135,14 @@ if [[ -z ${exp_name} ]]; then
exp_name=${exp_name}_${exp_subfix}
fi
fi
ckpt_dir=${code_dir}/checkpoints/
model_dir=${code_dir}/checkpoints/${data_model_subfix}/${sub_tag}/${exp_name}
# Start
cd ${code_dir}
echo "Start Stage: $stage"
echo "Stop Stage: $stop_stage"
cd ${code_dir}
if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then
echo "Default Stage: env configure"
......@@ -150,12 +151,10 @@ fi
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
echo "Stage -1: Data Download"
# pass
fi
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
### Task dependent. You have to make data the following preparation part by yourself.
### But you can utilize Kaldi recipes in most cases
echo "Stage 0: ASR Data Preparation"
if [[ ! -e ${data_dir} ]]; then
mkdir -p ${data_dir}
......@@ -255,6 +254,7 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
export CUDA_VISIBLE_DEVICES=${device}
fi
echo -e "data=${data_dir} model=${model_dir}"
......@@ -308,11 +308,6 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
cmd="${cmd}
--distributed-world-size $gpu_num
--ddp-backend no_c10d"
if [[ ${DIST_RANK} -ne 0 ]]; then
cmd="${cmd}
--distributed-init-method ${INIT_METHOD}
--distributed-rank ${DIST_RANK}"
fi
fi
if [[ $fp16 -eq 1 ]]; then
cmd="${cmd}
......@@ -362,7 +357,6 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "${time} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
tail -n 50 ${log} > tmp.log
mv tmp.log $log
# export CUDA_VISIBLE_DEVICES=${device}
log=${model_dir}/train.log
cmd="${cmd} 2>&1 | tee -a ${log}"
......@@ -404,8 +398,8 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
export CUDA_VISIBLE_DEVICES=${device}
fi
# export CUDA_VISIBLE_DEVICES=${device}
suffix=beam${beam_size}_alpha${len_penalty}_tokens${max_tokens}
if [[ ${n_average} -ne 1 ]]; then
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论