提交 eb827bb3 编写于 作者: L LDOUBLEV

add test shell

上级 926dc47b
train_model_list: det;benchmark/benchmark_det.yml
gpu_list: -1|0|0,1
auto_cast_list: False|True
trainer_list: norm|quant|prune
python: python3.7
inference: python|C++
devices: cpu|gpu
use_mkldnn_list: True|False
cpu_threads_list: 1|6
rec_batch_size_list: 1|6
gpu_trt_list: True|False
gpu_precision_list: fp32|fp16|int8
img_dir: /paddle/OCR/test_set/benchmark_eval
epoch: 10
checkpoints: None
#!/bin/bash
FILENAME=$1
# MODE be one of ['lite_train_infer' 'whole_infer' 'whole_train_infer']
MODE=$2
# prepare pretrained weights and dataset
wget -nc -P ./pretrain_models/ https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/MobileNetV3_large_x0_5_pretrained.pdparams
if [ ${MODE} = "lite_train_infer" ];then
# pretrain lite train data
rm -rf ./train_data/icdar2015
wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/icdar2015_lite.tar
cd ./train_data/ && tar xf icdar2015_lite.tar &&
ln -s ./icdar2015_lite ./icdar2015
cd ../
elif [ ${MODE} = "whole_train_infer" ];then
rm -rf ./train_data/icdar2015
wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/icdar2015.tar
cd ./train_data/ && tar xf icdar2015.tar && cd ../
else
echo "Do Nothing"
fi
dataline=$(cat ${FILENAME})
# parser params
IFS=$'\n'
lines=(${dataline})
function func_parser(){
strs=$1
IFS=":"
array=(${strs})
tmp=${array[1]}
echo ${tmp}
}
IFS=$'\n'
# The training params
train_model_list=$(func_parser "${lines[0]}")
gpu_list=$(func_parser "${lines[1]}")
auto_cast_list=$(func_parser "${lines[2]}")
slim_trainer_list=$(func_parser "${lines[3]}")
python=$(func_parser "${lines[4]}")
# inference params
inference=$(func_parser "${lines[5]}")
devices=$(func_parser "${lines[6]}")
use_mkldnn_list=$(func_parser "${lines[7]}")
cpu_threads_list=$(func_parser "${lines[8]}")
rec_batch_size_list=$(func_parser "${lines[9]}")
gpu_trt_list=$(func_parser "${lines[10]}")
gpu_precision_list=$(func_parser "${lines[11]}")
img_dir=$(func_parser "${lines[12]}")
# train superparameters
epoch=$(func_parser "${lines[13]}")
checkpoints=$(func_parser "${lines[14]}")
for train_model in ${train_model_list[*]}; do
if [ ${train_model} = "det" ];then
model_name="det"
yml_file="configs/det/det_mv3_db.yml"
elif [ ${train_model} = "rec" ];then
model_name="rec"
yml_file="configs/rec/rec_mv3_none_bilstm_ctc.yml"
else
model_name="det"
yml_file="configs/det/det_mv3_db.yml"
fi
# array=(${train_model})
# for j in "${!array[@]}"; do
# model_name=${array[0]}
# yml_file=${array[1]}
# done
IFS="|"
for gpu in ${gpu_list[*]}; do
use_gpu=True
if [ ${gpu} = "-1" ];then
lanuch=""
use_gpu=False
elif [ ${#gpu} -le 1 ];then
launch=""
else
launch="-m paddle.distributed.launch --log_dir=./debug/ --gpus ${gpu}"
fi
# echo "model_name: ${model_name} yml_file: ${yml_file} launch: ${launch} gpu: ${gpu}"
for auto_cast in ${auto_cast_list[*]}; do
for slim_trainer in ${slim_trainer_list[*]}; do
if [ ${slim_trainer} = "norm" ]; then
trainer="tools/train.py"
export_model="tools/export_model.py"
elif [ ${slim_trainer} = "quant" ]; then
trainer="deploy/slim/quantization/quant.py"
export_model="deploy/slim/quantization/export_model.py"
elif [ ${slim_trainer} = "prune" ]; then
trainer="deploy/slim/prune/sensitivity_anal.py"
export_model="deploy/slim/prune/export_prune_model.py"
elif [ ${slim_trainer} = "distill" ]; then
trainer="deploy/slim/distill/train_dml.py"
export_model="deploy/slim/distill/export_distill_model.py"
else
trainer="tools/train.py"
export_model="tools/export_model.py"
fi
# dataset="Train.dataset.data_dir=${train_dir} Train.dataset.label_file_list=${train_label_file} Eval.dataset.data_dir=${eval_dir} Eval.dataset.label_file_list=${eval_label_file}"
save_log=${log_path}/${model_name}_${slim_trainer}_autocast_${auto_cast}_gpuid_${gpu}
echo ${python} ${launch} ${trainer} -c ${yml_file} -o Global.auto_cast=${auto_cast} Global.save_model_dir=${save_log} Global.use_gpu=${use_gpu} Global.epoch=${epoch}
echo ${python} ${export_model} -c ${yml_file} -o Global.pretrained_model=${save_log}/best_accuracy Global.save_inference_dir=${save_log}/export_inference/
if [ "${model_name}" = "det" ]; then
export rec_batch_size_list=( "1" )
inference="tools/infer/predict_det.py"
elif [ "${model_name}" = "rec" ]; then
inference="tools/infer/predict_rec.py"
fi
# inference
for device in ${devices[*]}; do
if [ ${device} = "cpu" ]; then
for use_mkldnn in ${use_mkldnn_list[*]}; do
for threads in ${cpu_threads_list[*]}; do
for rec_batch_size in ${rec_batch_size_list[*]}; do
echo ${python} ${inference} --enable_mkldnn=${use_mkldnn} --use_gpu=False --cpu_threads=${threads} --benchmark=True --det_model_dir=${save_log}/export_inference/ --rec_batch_num=${rec_batch_size} --rec_model_dir=${rec_model_dir} --image_dir=${img_dir} --save_log_path=${log_path}/${model_name}_${slim_trainer}_cpu_usemkldnn_${use_mkldnn}_cputhreads_${threads}_recbatchnum_${rec_batch_size}_infer.log
# ${python} ${inference} --enable_mkldnn=${use_mkldnn} --use_gpu=False --cpu_threads=${threads} --benchmark=True --det_model_dir=${save_log}/export_inference/ --rec_batch_num=${rec_batch_size} --rec_model_dir=${rec_model_dir} --image_dir=${img_dir} 2>&1 | tee ${log_path}/${model_name}_${slim_trainer}_cpu_usemkldnn_${use_mkldnn}_cputhreads_${threads}_recbatchnum_${rec_batch_size}_infer.log
done
done
done
else
for use_trt in ${gpu_trt_list[*]}; do
for precision in ${gpu_precision_list[*]}; do
if [ ${use_trt} = "False" ] && [ ${precision} != "fp32" ]; then
continue
fi
for rec_batch_size in ${rec_batch_size_list[*]}; do
# echo "${model_name} ${det_model_dir} ${rec_model_dir}, use_trt: ${use_trt} use_fp16: ${use_fp16}"
echo ${python} ${inference} --use_gpu=True --use_tensorrt=${use_trt} --precision=${precision} --benchmark=True --det_model_dir=${save_log}/export_inference/ --rec_batch_num=${rec_batch_size} --rec_model_dir=${rec_model_dir} --image_dir=${img_dir} --save_log_path=${log_path}/${model_name}_${slim_trainer}_gpu_usetensorrt_${use_trt}_usefp16_${precision}_recbatchnum_${rec_batch_size}_infer.log
done
done
done
fi
done
done
done
done
done
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册