提交源码

src/PaddleClas/benchmark/README.md

@@ -0,0 +1,27 @@
+# benchmark使用说明
+
+此目录所有shell脚本是为了测试PaddleClas中不同模型的速度指标，如单卡训练速度指标、多卡训练速度指标等。
+
+## 相关脚本说明
+
+一共有3个脚本：
+
+- `prepare_data.sh`: 下载相应的测试数据，并配置好数据路径
+- `run_benchmark.sh`: 执行单独一个训练测试的脚本，具体调用方式，可查看脚本注释
+- `run_all.sh`: 执行所有训练测试的入口脚本
+
+## 使用说明
+
+**注意**：为了跟PaddleClas中其他的模块的执行目录保持一致，此模块的执行目录为`PaddleClas`的根目录。
+
+### 1.准备数据
+
+```shell
+bash benchmark/prepare_data.sh
+```
+
+### 2.执行所有模型的测试
+
+```shell
+bash benchmark/run_all.sh
+```

src/PaddleClas/benchmark/prepare_data.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+dataset_url=$1
+
+package_check_list=(imageio tqdm Cython pycocotools tb_paddle scipy pandas wget h5py sklearn opencv-python visualdl)
+for package in ${package_check_list[@]}; do
+    if python -c "import ${package}" >/dev/null 2>&1; then
+        echo "${package} have already installed"
+    else
+        echo "${package} NOT FOUND"
+        pip install ${package}
+        echo "${package} installed"
+    fi
+done
+
+cd dataset
+rm -rf ILSVRC2012
+wget -nc ${dataset_url}
+tar xf ILSVRC2012_val.tar
+ln -s ILSVRC2012_val ILSVRC2012
+cd ILSVRC2012
+ln -s val_list.txt train_list.txt
+cd ../../

src/PaddleClas/benchmark/run_all.sh

@@ -0,0 +1,31 @@
+# 提供可稳定复现性能的脚本，默认在标准docker环境内py37执行： paddlepaddle/paddle:latest-gpu-cuda10.1-cudnn7  paddle=2.1.2  py=37
+# 执行目录：需说明
+# cd **
+# 1 安装该模型需要的依赖 (如需开启优化策略请注明)
+# pip install ...
+# 2 拷贝该模型需要数据、预训练模型
+# 3 批量运行（如不方便批量，1，2需放到单个模型中）
+log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}  # LOG_PATH_INDEX_DIR 后续QA设置参数
+model_mode_list=(MobileNetV1 MobileNetV2 MobileNetV3_large_x1_0 ShuffleNetV2_x1_0 HRNet_W48_C SwinTransformer_tiny_patch4_window7_224 alt_gvt_base)    # benchmark 监控模型列表
+#model_mode_list=(MobileNetV1 MobileNetV2 MobileNetV3_large_x1_0 EfficientNetB0 ShuffleNetV2_x1_0 DenseNet121 HRNet_W48_C SwinTransformer_tiny_patch4_window7_224 alt_gvt_base)   # 该脚本支持列表
+fp_item_list=(fp32)
+#bs_list=(32 64 96 128)
+for model_mode in ${model_mode_list[@]}; do
+      for fp_item in ${fp_item_list[@]}; do
+          if [ ${model_mode} = MobileNetV3_large_x1_0 ] || [ ${model_mode} = ShuffleNetV2_x1_0 ]; then
+              bs_list=(256)
+          else
+              bs_list=(64)
+          fi
+          for bs_item in ${bs_list[@]};do
+	    echo "index is speed, 1gpus, begin, ${model_name}"
+	    run_mode=sp
+	    CUDA_VISIBLE_DEVICES=0 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 1 ${model_mode} | tee ${log_path}/clas_${model_mode}_${run_mode}_bs${bs_item}_${fp_item}_1gpus 2>&1    #  (5min)
+	    sleep 10
+            echo "index is speed, 8gpus, run_mode is multi_process, begin, ${model_name}"
+            run_mode=mp
+            CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 1 ${model_mode}| tee ${log_path}/clas_${model_mode}_${run_mode}_bs${bs_item}_${fp_item}_8gpus8p 2>&1
+            sleep 10
+            done
+      done
+done

src/PaddleClas/benchmark/run_benchmark.sh

@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+set -xe
+# 运行示例：CUDA_VISIBLE_DEVICES=0 bash run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 500 ${model_mode}
+# 参数说明
+function _set_params(){
+    run_mode=${1:-"sp"}          # 单卡sp|多卡mp
+    batch_size=${2:-"64"}
+    fp_item=${3:-"fp32"}        # fp32|fp16
+    epochs=${4:-"2"}       # 可选，如果需要修改代码提前中断
+    model_item=${5:-"model_item"}
+    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # TRAIN_LOG_DIR 后续QA设置该参数
+
+    index=1
+    mission_name="图像分类"           # 模型所属任务名称，具体可参考scripts/config.ini                                （必填）
+    direction_id=0                    # 任务所属方向，0：CV，1：NLP，2：Rec。                                          (必填)
+    skip_steps=8                      # 解析日志，有些模型前几个step耗时长，需要跳过                                  (必填)
+    keyword="ips:"              # 解析日志，筛选出数据所在行的关键字                                            (必填)
+    keyword_loss="loss:"       #选填
+    model_mode=-1                      # 解析日志，具体参考scripts/analysis.py.                                        (必填)
+    ips_unit="images/s"
+    base_batch_size=$batch_size
+#   以下不用修改   
+    device=${CUDA_VISIBLE_DEVICES//,/ }
+    arr=(${device})
+    num_gpu_devices=${#arr[*]}
+    log_file=${run_log_path}/clas_${model_item}_${run_mode}_bs${batch_size}_${fp_item}_${num_gpu_devices}
+    model_name=${model_item}_bs${batch_size}_${fp_item}    # model_item 用于yml匹配，model_name用于入库
+}
+function _train(){
+    echo "Train on ${num_gpu_devices} GPUs"
+    echo "current CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES, gpus=$num_gpu_devices, batch_size=$batch_size"
+
+    if [ ${fp_item} = "fp32" ];then
+        model_config=`find ppcls/configs/ImageNet -name ${model_item}.yaml`
+    else
+        model_config=`find ppcls/configs/ImageNet -name ${model_item}_fp16.yaml`
+    fi
+
+    train_cmd="-c ${model_config} -o DataLoader.Train.sampler.batch_size=${batch_size} -o Global.epochs=${epochs}  -o Global.eval_during_train=False -o Global.print_batch_step=2"
+    case ${run_mode} in
+    sp) train_cmd="python -u tools/train.py ${train_cmd}" ;;
+    mp)
+        train_cmd="python -m paddle.distributed.launch --log_dir=./mylog --gpus=$CUDA_VISIBLE_DEVICES tools/train.py ${train_cmd}"
+        log_parse_file="mylog/workerlog.0" ;;
+    *) echo "choose run_mode(sp or mp)"; exit 1;
+    esac
+    rm -rf mylog
+# 以下不用修改
+    timeout 5m ${train_cmd} > ${log_file} 2>&1
+    if [ $? -ne 0 ];then
+        echo -e "${model_name}, FAIL"
+        export job_fail_flag=1
+    else
+        echo -e "${model_name}, SUCCESS"
+        export job_fail_flag=0
+    fi
+    kill -9 `ps -ef|grep 'python'|awk '{print $2}'`
+ 
+    if [ $run_mode = "mp" -a -d mylog ]; then
+        rm ${log_file}
+        cp mylog/workerlog.0 ${log_file}
+    fi
+}
+
+source ${BENCHMARK_ROOT}/scripts/run_model.sh # 在该脚本中会对符合benchmark规范的log使用analysis.py 脚本进行性能数据解析;该脚本在连调时可从benchmark repo中下载https://github.com/PaddlePaddle/benchmark/blob/master/scripts/run_model.sh;如果不联调只想要产出训练log可以注掉本行,提交时需打开
+_set_params $@
+_run
+#_train