Skip to content

P
PaddleClas

PaddlePaddle / PaddleClas
大约 1 年 前同步成功

通知 115
Star 4999
Fork 1114
P
PaddleClas
提交 7990a12e 编写于 作者: H HydrogenSulfate
浏览文件
操作

Merge branch 'develop' into add_pact_chain

上级 1e281f8c 22531707
隐藏空白更改
内联 并排
Showing with 29 addition and 29 deletion
docs/zh_CN/PULC/PULC_vehicle_attribute.md
浏览文件 @ 7990a12e
......@@ -58,7 +58,7 @@
从表中可以看出，backbone 为 Res2Net200_vd_26w_4s 时精度较高，但是推理速度较慢。将 backbone 替换为轻量级模型 MobileNetV3_small_x0_35 后，速度可以大幅提升，但是精度下降明显。将 backbone 替换为 PPLCNet_x1_0 时，精度提升 2 个百分点，同时速度也提升 23% 左右。在此基础上，使用 SSLD 预训练模型后，在不改变推理速度的前提下，精度可以提升约 0.5 个百分点，进一步地，当融合EDA策略后，精度可以再提升 0.52 个百分点，最后，在使用 SKL-UGI 知识蒸馏后，精度可以继续提升 0.23 个百分点。此时，PPLCNet_x1_0 的精度与 Res2Net200_vd_26w_4s 仅相差 0.55 个百分点，但是速度快 32 倍。关于 PULC 的训练方法和推理部署方法将在下面详细介绍。
**备注：**
* 延时是基于 Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz 测试得到，开启 MKLDNN 加速策略，线程数为10。
* 关于PP-LCNet的介绍可以参考[PP-LCNet介绍](../models/PP-LCNet.md)，相关论文可以查阅[PP-LCNet paper](https://arxiv.org/abs/2109.15099)。
......@@ -178,7 +178,7 @@ from xml.dom.minidom import parse
vehicleids = []
def convert_annotation(input_fp, output_fp):
def convert_annotation(input_fp, output_fp, subdir):
in_file = open(input_fp)
list_file = open(output_fp, 'w')
tree = parse(in_file)
......@@ -201,12 +201,12 @@ def convert_annotation(input_fp, output_fp):
typeid = int (item.getAttribute("typeID"))
label[typeid+9] = '1'
label = ','.join(label)
list_file.write(os.path.join('image_train', name) + "\t" + label + "\n")
list_file.write(os.path.join(subdir, name) + "\t" + label + "\n")
list_file.close()
convert_annotation('train_label.xml', 'train_list.txt') #imagename vehiclenum colorid typeid
convert_annotation('test_label.xml', 'test_list.txt')
convert_annotation('train_label.xml', 'train_list.txt', 'image_train') #imagename vehiclenum colorid typeid
convert_annotation('test_label.xml', 'test_list.txt', 'image_test')
```
执行上述命令后，`VeRi`目录中具有以下数据：
......
test_tipc/docs/test_inference_cpp.md
浏览文件 @ 7990a12e
......@@ -248,20 +248,20 @@ bash test_tipc/prepare.sh test_tipc/config/ResNet/ResNet50_linux_gpu_normal_norm
测试方法如下所示，希望测试不同的模型文件，只需更换为自己的参数配置文件，即可完成对应模型的测试。
```shell
bash test_tipc/test_inference_cpp.sh ${your_params_file}
bash test_tipc/test_inference_cpp.sh ${your_params_file} cpp_infer
```
以`ResNet50`的`Linux GPU/CPU C++推理测试`为例，命令如下所示。
```shell
bash test_tipc/test_inference_cpp.sh test_tipc/config/ResNet/ResNet50_linux_gpu_normal_normal_infer_cpp_linux_gpu_cpu.txt
bash test_tipc/test_inference_cpp.sh test_tipc/config/ResNet/ResNet50_linux_gpu_normal_normal_infer_cpp_linux_gpu_cpu.txt cpp_infer
```
输出结果如下，表示命令运行成功。
```shell
Run successfully with command - ./deploy/cpp/build/clas_system -c inference_cls.yaml > ./test_tipc/output/ResNet50/cls_cpp_infer_gpu_usetrt_False_precision_fp32_batchsize_1.log 2>&1!
Run successfully with command - ./deploy/cpp/build/clas_system -c inference_cls.yaml > ./test_tipc/output/ResNet50/cls_cpp_infer_cpu_usemkldnn_False_threads_1_precision_fp32_batchsize_1.log 2>&1!
Run successfully with command - ResNet50 - ./deploy/cpp/build/clas_system -c inference_cls.yaml > ./test_tipc/output/ResNet50/cpp_infer/cpp_infer_gpu_usetrt_False_precision_fp32_batchsize_1.log 2>&1!
Run successfully with command - ResNet50 - ./deploy/cpp/build/clas_system -c inference_cls.yaml > ./test_tipc/output/ResNet50/cpp_infer/cpp_infer_cpu_usemkldnn_False_threads_1_precision_fp32_batchsize_1.log 2>&1!
```
最终log中会打印出结果，如下所示
......@@ -312,6 +312,6 @@ Current total inferen time cost: 5449.39 ms.
Top5: class_id: 265, score: 0.0420, label: toy poodle
```
详细log位于`./test_tipc/output/ResNet50/cls_cpp_infer_gpu_usetrt_False_precision_fp32_batchsize_1.log`和`./test_tipc/output/ResNet50/cls_cpp_infer_cpu_usemkldnn_False_threads_1_precision_fp32_batchsize_1.log`中。
详细log位于`./test_tipc/output/ResNet50/cpp_infer/cpp_infer_gpu_usetrt_False_precision_fp32_batchsize_1.log`和`./test_tipc/output/ResNet50/cpp_infer_cpu_usemkldnn_False_threads_1_precision_fp32_batchsize_1.log`中。
如果运行失败，也会在终端中输出运行失败的日志信息以及对应的运行命令。可以基于该命令，分析运行失败的原因。
test_tipc/prepare.sh
浏览文件 @ 7990a12e
......@@ -84,7 +84,12 @@ if [[ ${MODE} = "cpp_infer" ]]; then
fi
if [[ ! -d "./deploy/cpp/paddle_inference/" ]]; then
pushd ./deploy/cpp/
wget -nc https://paddle-inference-lib.bj.bcebos.com/2.2.2/cxx_c/Linux/GPU/x86-64_gcc8.2_avx_mkl_cuda10.1_cudnn7.6.5_trt6.0.1.5/paddle_inference.tgz
PADDLEInfer=$3
if [ "" = "$PADDLEInfer" ];then
wget -nc https://paddle-inference-lib.bj.bcebos.com/2.2.2/cxx_c/Linux/GPU/x86-64_gcc8.2_avx_mkl_cuda10.1_cudnn7.6.5_trt6.0.1.5/paddle_inference.tgz --no-check-certificate
else
wget -nc ${PADDLEInfer} --no-check-certificate
fi
tar xf paddle_inference.tgz
popd
fi
......
test_tipc/test_inference_cpp.sh
浏览文件 @ 7990a12e
......@@ -2,10 +2,17 @@
source test_tipc/common_func.sh
FILENAME=$1
GPUID=$2
MODE=$2
# set cuda device
GPUID=$3
if [[ ! $GPUID ]];then
GPUID=0
fi
env="export CUDA_VISIBLE_DEVICES=${GPUID}"
set CUDA_VISIBLE_DEVICES
eval $env
dataline=$(awk 'NR==1, NR==19{print}' $FILENAME)
# parser params
......@@ -30,7 +37,7 @@ cpp_benchmark_value=$(func_parser_value "${lines[16]}")
generate_yaml_cmd=$(func_parser_value "${lines[17]}")
transform_index_cmd=$(func_parser_value "${lines[18]}")
LOG_PATH="./test_tipc/output/${model_name}"
LOG_PATH="./test_tipc/output/${model_name}/${MODE}"
mkdir -p ${LOG_PATH}
status_log="${LOG_PATH}/results_cpp.log"
# generate_yaml_cmd="python3 test_tipc/generate_cpp_yaml.py"
......@@ -56,7 +63,7 @@ function func_shitu_cpp_inference(){
if [ ${use_mkldnn} = "False" ] && [ ${_flag_quant} = "True" ]; then
precison="int8"
fi
_save_log_path="${_log_path}/shitu_cpp_infer_cpu_usemkldnn_${use_mkldnn}_threads_${threads}_precision_${precision}_batchsize_${batch_size}.log"
_save_log_path="${_log_path}/cpp_infer_cpu_usemkldnn_${use_mkldnn}_threads_${threads}_precision_${precision}_batchsize_${batch_size}.log"
eval $transform_index_cmd
command="${generate_yaml_cmd} --type shitu --batch_size ${batch_size} --mkldnn ${use_mkldnn} --gpu ${use_gpu} --cpu_thread ${threads} --tensorrt False --precision ${precision} --data_dir ${_img_dir} --benchmark True --cls_model_dir ${cpp_infer_model_dir} --det_model_dir ${cpp_det_infer_model_dir} --gpu_id ${GPUID}"
eval $command
......@@ -80,7 +87,7 @@ function func_shitu_cpp_inference(){
continue
fi
for batch_size in ${cpp_batch_size_list[*]}; do
_save_log_path="${_log_path}/shitu_cpp_infer_gpu_usetrt_${use_trt}_precision_${precision}_batchsize_${batch_size}.log"
_save_log_path="${_log_path}/cpp_infer_gpu_usetrt_${use_trt}_precision_${precision}_batchsize_${batch_size}.log"
eval $transform_index_cmd
command="${generate_yaml_cmd} --type shitu --batch_size ${batch_size} --mkldnn False --gpu ${use_gpu} --cpu_thread 1 --tensorrt ${use_trt} --precision ${precision} --data_dir ${_img_dir} --benchmark True --cls_model_dir ${cpp_infer_model_dir} --det_model_dir ${cpp_det_infer_model_dir} --gpu_id ${GPUID}"
eval $command
......@@ -118,7 +125,7 @@ function func_cls_cpp_inference(){
if [ ${use_mkldnn} = "False" ] && [ ${_flag_quant} = "True" ]; then
precison="int8"
fi
_save_log_path="${_log_path}/cls_cpp_infer_cpu_usemkldnn_${use_mkldnn}_threads_${threads}_precision_${precision}_batchsize_${batch_size}.log"
_save_log_path="${_log_path}/cpp_infer_cpu_usemkldnn_${use_mkldnn}_threads_${threads}_precision_${precision}_batchsize_${batch_size}.log"
command="${generate_yaml_cmd} --type cls --batch_size ${batch_size} --mkldnn ${use_mkldnn} --gpu ${use_gpu} --cpu_thread ${threads} --tensorrt False --precision ${precision} --data_dir ${_img_dir} --benchmark True --cls_model_dir ${cpp_infer_model_dir} --gpu_id ${GPUID}"
eval $command
......@@ -142,7 +149,7 @@ function func_cls_cpp_inference(){
continue
fi
for batch_size in ${cpp_batch_size_list[*]}; do
_save_log_path="${_log_path}/cls_cpp_infer_gpu_usetrt_${use_trt}_precision_${precision}_batchsize_${batch_size}.log"
_save_log_path="${_log_path}/cpp_infer_gpu_usetrt_${use_trt}_precision_${precision}_batchsize_${batch_size}.log"
command="${generate_yaml_cmd} --type cls --batch_size ${batch_size} --mkldnn False --gpu ${use_gpu} --cpu_thread 1 --tensorrt ${use_trt} --precision ${precision} --data_dir ${_img_dir} --benchmark True --cls_model_dir ${cpp_infer_model_dir} --gpu_id ${GPUID}"
eval $command
command="${_script} > ${_save_log_path} 2>&1"
......@@ -235,18 +242,6 @@ cd ../../../
# cd ../../
echo "################### build PaddleClas demo finished ###################"
# set cuda device
GPUID=$3
if [ ${#GPUID} -le 0 ];then
env="export CUDA_VISIBLE_DEVICES=0"
else
env="export CUDA_VISIBLE_DEVICES=${GPUID}"
fi
set CUDA_VISIBLE_DEVICES
eval $env
echo "################### run test ###################"
export Count=0
IFS="|"
......
Markdown is supported
0% 或 .
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 或