[add]上传训练benchmark by z00560161
This commit is contained in:
liang_chaoming@huawei.com
@@ -0,0 +1,69 @@
|
||||
#!/bin/bash
|
||||
|
||||
rank_size=$1
|
||||
yamlPath=$2
|
||||
toolsPath=$3
|
||||
if [ -f /.dockerenv ];then
|
||||
CLUSTER=$4
|
||||
MPIRUN_ALL_IP="$5"
|
||||
export CLUSTER=${CLUSTER}
|
||||
fi
|
||||
currentDir=$(cd "$(dirname "$0")/.."; pwd)
|
||||
# 配置环境变量并调用 train 方法
|
||||
currtime=`date +%Y%m%d%H%M%S`
|
||||
mkdir -p ${currentDir%train*}/train/result/tf_resnet101/training_job_${currtime}/
|
||||
train_job_dir=${currentDir%train*}/train/result/tf_resnet101/training_job_${currtime}/
|
||||
echo "[`date +%Y%m%d-%H:%M:%S`] [INFO] ${train_job_dir} &"
|
||||
|
||||
# user env
|
||||
export HCCL_CONNECT_TIMEOUT=600
|
||||
export JOB_ID=9999001
|
||||
export SLOG_PRINT_TO_STDOUT=0
|
||||
export RANK_TABLE_FILE=${currentDir}/config/${rank_size}p.json
|
||||
|
||||
# 从 yaml 获取配置
|
||||
eval $(${toolsPath}/get_params_for_yaml.sh ${yamlPath} "tensorflow_config")
|
||||
|
||||
|
||||
# device 列表, 若无指定 device 根据 rank_size 顺序选择
|
||||
eval device_group=\$device_group_${rank_size}p
|
||||
if [ x"${device_group}" == x"" ] || [ ${rank_size} -ge 8 ];then
|
||||
device_group="$(seq 0 "$(expr $rank_size - 1)")"
|
||||
fi
|
||||
|
||||
# get last device id in device_group, hw log in performance from the dir named first_device_id
|
||||
device_group_str=`echo ${device_group} | sed 's/ //g'`
|
||||
first_device_id=`echo ${device_group_str: 0:1}`
|
||||
|
||||
|
||||
rank_id=0
|
||||
|
||||
if [ x"${CLUSTER}" == x"True" ];then
|
||||
# ln hw log
|
||||
ln -snf ${train_job_dir}/0/hw_resnet101.log ${train_job_dir}
|
||||
this_ip=$(hostname -I |awk '{print $1}')
|
||||
for ip in $MPIRUN_ALL_IP;do
|
||||
if [ x"$ip" != x"$this_ip" ];then
|
||||
scp $yamlPath root@$ip:$yamlPath
|
||||
fi
|
||||
done
|
||||
export PATH=$PATH:/usr/local/mpirun4.0/bin
|
||||
mpirun -H ${mpirun_ip} \
|
||||
--bind-to none -map-by slot\
|
||||
--allow-run-as-root \
|
||||
--mca btl_tcp_if_exclude lo,docker0,endvnic,virbr0,vethf40501b,docker_gwbridge,br-f42ac38052b4\
|
||||
--prefix /usr/local/mpirun4.0/ \
|
||||
${currentDir}/scripts/train.sh 0 $rank_size $yamlPath $currtime ${toolsPath} ${CLUSTER}
|
||||
else
|
||||
# ln hw log
|
||||
ln -snf ${train_job_dir}/${first_device_id}/hw_resnet101.log ${train_job_dir}
|
||||
for device_id in $device_group;do
|
||||
#echo "[`date +%Y%m%d-%H:%M:%S`] [INFO] start: train ${device_id} & " >> ${currentDir}/result/main.log
|
||||
${currentDir}/scripts/train.sh $device_id $rank_size $yamlPath $currtime ${toolsPath} $rank_id &
|
||||
let rank_id++
|
||||
done
|
||||
fi
|
||||
wait
|
||||
|
||||
|
||||
echo "[`date +%Y%m%d-%H:%M:%S`] [INFO] ${train_job_dir} &"
|
||||
+124
@@ -0,0 +1,124 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
device_id=$1
|
||||
rank_size=$2
|
||||
yamlPath=$3
|
||||
currtime=$4
|
||||
toolsPath=$5
|
||||
|
||||
currentDir=$(cd "$(dirname "$0")/.."; pwd)
|
||||
|
||||
export REMARK_LOG_FILE=hw_resnet101.log
|
||||
|
||||
mkdir -p ${currentDir%train*}/train/result/tf_resnet101/training_job_${currtime}/
|
||||
export train_job_dir=${currentDir%train*}/train/result/tf_resnet101/training_job_${currtime}/
|
||||
|
||||
|
||||
source ${currentDir}/config/npu_set_env.sh
|
||||
|
||||
benchmark_log_path=${currentDir%atlas_benchmark-master*}/atlas_benchmark-master/utils
|
||||
#atlasboost_path=${currentDir%atlas_benchmark-master*}/atlas_benchmark-master/utils/atlasboost
|
||||
code_dir_path=${currentDir}/code
|
||||
export PYTHONPATH=$PYTHONPATH:${benchmark_log_path}:${code_dir_path}
|
||||
|
||||
# 从 yaml 获取配置
|
||||
eval $(${toolsPath}/get_params_for_yaml.sh ${yamlPath} "tensorflow_config")
|
||||
|
||||
# user env
|
||||
export YAML_PATH=$3
|
||||
export HCCL_CONNECT_TIMEOUT=600
|
||||
export JOB_ID=9999002
|
||||
export RANK_TABLE_FILE=${currentDir}/config/${rank_size}p.json
|
||||
export RANK_SIZE=${rank_size}
|
||||
export RANK_INDEX=0
|
||||
export SLOG_PRINT_TO_STDOUT=0
|
||||
export DEVICE_ID=$1
|
||||
DEVICE_INDEX=$(( DEVICE_ID + RANK_INDEX * 8 ))
|
||||
export DEVICE_INDEX=${DEVICE_INDEX}
|
||||
|
||||
cd ${train_job_dir}
|
||||
curd_dir=${currentDir%atlas_benchmark-master*}/atlas_benchmark-master/utils/atlasboost
|
||||
export PYTHONPATH=$PYTHONPATH:${curd_dir}
|
||||
|
||||
if [ x"$6" != x"True" ];then
|
||||
rank_id=$6
|
||||
export RANK_ID=$6
|
||||
else
|
||||
device_id_mo=$(python3.7 -c "import src.tensorflow.mpi_ops as atlasboost;atlasboost.init(); \
|
||||
device_id = atlasboost.local_rank();cluster_device_id = str(device_id); \
|
||||
atlasboost.set_device_id(device_id);print(atlasboost.rank())")
|
||||
device_id_mo=`echo $device_id_mo`
|
||||
rank_id=${device_id_mo##* }
|
||||
export RANK_ID=${rank_id}
|
||||
device=${device_id_mo##*deviceid = }
|
||||
device_id=${device%% phyid=*}
|
||||
export DEVICE_ID=${device_id}
|
||||
hccljson=${train_job_dir}/*.json
|
||||
cp ${hccljson} ${currentDir}/config/${rank_size}p.json
|
||||
fi
|
||||
|
||||
export MODEL_CKPT_PATH=${train_job_dir}/${device_id}/ckpt${device_id}
|
||||
|
||||
#mkdir exec path
|
||||
mkdir -p ${train_job_dir}/${device_id}
|
||||
cd ${train_job_dir}/${device_id}
|
||||
|
||||
startTime=`date +%Y%m%d-%H:%M:%S`
|
||||
startTime_s=`date +%s`
|
||||
|
||||
# 根据单卡/多卡区分调用参数
|
||||
|
||||
if [ x"$6" == x"True" ];then
|
||||
export CLUSTER=True
|
||||
echo "run cluster"
|
||||
python3.7 ${currentDir}/code/official/r1/resnet/imagenet_main.py \
|
||||
--device_count=${rank_size} \
|
||||
--resnet_size=101 \
|
||||
--batch_size=${batch_size} \
|
||||
--num_gpus=1 \
|
||||
--max_train_steps=${max_train_steps} \
|
||||
--cosine_lr=True \
|
||||
--dtype=fp16 \
|
||||
--label_smoothing=0.1 \
|
||||
--loss_scale=512 \
|
||||
--train_epochs=${epoches} \
|
||||
--epochs_between_evals=${epochs_between_evals} \
|
||||
--hooks=ExamplesPerSecondHook,loggingtensorhook,loggingmetrichook \
|
||||
--data_dir=${data_url} \
|
||||
--model_dir=${MODEL_CKPT_PATH} > ${train_job_dir}/train_${device_id}.log 2>&1
|
||||
else
|
||||
python3.7 ${currentDir}/code/official/r1/resnet/imagenet_main.py \
|
||||
--device_count=${rank_size} \
|
||||
--resnet_size=101 \
|
||||
--batch_size=${batch_size} \
|
||||
--num_gpus=1 \
|
||||
--max_train_steps=${max_train_steps} \
|
||||
--cosine_lr=True \
|
||||
--dtype=fp16 \
|
||||
--label_smoothing=0.1 \
|
||||
--loss_scale=512 \
|
||||
--train_epochs=${epoches} \
|
||||
--epochs_between_evals=${epochs_between_evals} \
|
||||
--hooks=ExamplesPerSecondHook,loggingtensorhook,loggingmetrichook \
|
||||
--data_dir=${data_url} \
|
||||
--model_dir=${MODEL_CKPT_PATH} > ${train_job_dir}/train_${device_id}.log 2>&1
|
||||
fi
|
||||
|
||||
if [ $? -eq 0 ] ;then
|
||||
echo ":::ABK 1.0.0 resnet101 train success"
|
||||
echo ":::ABK 1.0.0 resnet101 train success" >> ${train_job_dir}/train_${device_id}.log
|
||||
echo ":::ABK 1.0.0 resnet101 train success" >> ${train_job_dir}/${device_id}/hw_resnet101.log
|
||||
else
|
||||
echo ":::ABK 1.0.0 resnet101 train failed"
|
||||
echo ":::ABK 1.0.0 resnet101 train failed" >> ${train_job_dir}/train_${device_id}.log
|
||||
echo ":::ABK 1.0.0 resnet101 train failed" >> ${train_job_dir}/${device_id}/hw_resnet101.log
|
||||
fi
|
||||
|
||||
endTime=`date +%Y%m%d-%H:%M:%S`
|
||||
endTime_s=`date +%s`
|
||||
sumTime=$[ $endTime_s - $startTime_s ]
|
||||
hour=$(( $sumTime/3600 ))
|
||||
min=$(( ($sumTime-${hour}*3600)/60 ))
|
||||
sec=$(( $sumTime-${hour}*3600-${min}*60 ))
|
||||
echo ":::ABK 1.0.0 resnet101 train total time ${hour}:${min}:${sec}"
|
||||
echo ":::ABK 1.0.0 resnet101 train total time ${hour}:${min}:${sec}" >> ${train_job_dir}/${device_id}/hw_resnet101.log
|
||||
Reference in New Issue
Block a user