add static graph ResNet50 for benchmark

ppcls/static/program.py

@@ -406,7 +406,7 @@ def run(dataloader,
             if "time" in key else str(metric_dict[key].value)
             for key in metric_dict
         ])
-        ips_info = " ips: {:.5f} images/sec.".format(
+        ips_info = " ips: {:.5f} samples/sec.".format(
             batch_size / metric_dict["batch_time"].avg)
         fetchs_str += ips_info
 
@@ -433,8 +433,8 @@ def run(dataloader,
 
     end_str = ' '.join([str(m.mean) for m in metric_dict.values()] +
                        [metric_dict["batch_time"].total])
-    ips_info = "ips: {:.5f} images/sec.".format(batch_size /
-                                                metric_dict["batch_time"].avg)
+    ips_info = "ips: {:.5f} samples/sec.".format(batch_size /
+                                                 metric_dict["batch_time"].avg)
     if mode == 'eval':
         logger.info("END {:s} {:s} {:s}".format(mode, end_str, ips_info))
     else:

test_tipc/static/README.md

+# PaddleClas 下静态图benchmark模型执行说明
+
+静态图benchmark测试脚本说明
+
+# 目录说明
+# Docker 运行环境
+docker image: registry.baidubce.com/paddlepaddle/paddle:latest-dev-cuda11.2-cudnn8-gcc82
+paddle = 2.2.2
+python = 3.7
+# 运行benchmark测试步骤
+
+```shell
+git clone https://github.com/PaddlePaddle/PaddleClas.git
+cd PaddleClas
+```
+
+# 准备数据
+
+```shell
+bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
+```
+
+# 运行模型
+## 单卡（自动运行打开Profiling）
+
+```shell
+export CUDA_VISIBLE_DEVICES=0
+bash  test_tipc/static/${model_item}/N1C1/${shell_name}.sh
+```
+
+## 多卡
+
+```shell
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+bash test_tipc/static/${model_item}/N1C8/${shell_name}.sh
+```

test_tipc/static/ResNet50/N1C1/ResNet50_bs64_fp16_SingleP_DP.sh

+model_item=ResNet50
+bs_item=64
+fp_item=fp16
+run_process_type=SingleP
+run_mode=DP
+device_num=N1C1
+max_epochs=1
+num_workers=4
+
+# get data
+bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
+# run
+bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_process_type} ${run_mode} ${device_num} ${max_epochs} ${num_workers} 2>&1;

test_tipc/static/ResNet50/N1C1/ResNet50_bs64_fp32_SingleP_DP.sh

+model_item=ResNet50
+bs_item=64
+fp_item=fp32
+run_process_type=SingleP
+run_mode=DP
+device_num=N1C1
+max_epochs=1
+num_workers=4
+
+# get data
+bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
+# run
+bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_process_type} ${run_mode} ${device_num} ${max_epochs} ${num_workers} 2>&1;
+# run profiling
+sleep 10;
+export PROFILING=true
+bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_process_type} ${run_mode} ${device_num} ${max_epochs} ${num_workers} 2>&1;

test_tipc/static/ResNet50/N1C8/ResNet50_bs64_fp16_MultiP_DP.sh

+model_item=ResNet50
+bs_item=64
+fp_item=fp16
+run_process_type=MultiP
+run_mode=DP
+device_num=N1C8
+max_epochs=1
+num_workers=4
+
+# get data
+bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
+# run
+bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_process_type} ${run_mode} ${device_num} ${max_epochs} ${num_workers} 2>&1;

test_tipc/static/ResNet50/N1C8/ResNet50_bs64_fp32_MultiP_DP.sh

+model_item=ResNet50
+bs_item=64
+fp_item=fp32
+run_process_type=MultiP
+run_mode=DP
+device_num=N1C8
+max_epochs=1
+num_workers=4
+
+# get data
+bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
+# run
+bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_process_type} ${run_mode} ${device_num} ${max_epochs} ${num_workers} 2>&1;

test_tipc/static/ResNet50/benchmark_common/prepare.sh

+#!/bin/bash
+
+pip install -r requirements.txt
+cd dataset
+rm -rf ILSVRC2012
+wget -nc https://paddle-imagenet-models-name.bj.bcebos.com/data/ImageNet1k/ILSVRC2012_val.tar
+tar xf ILSVRC2012_val.tar
+ln -s ILSVRC2012_val ILSVRC2012
+cd ILSVRC2012
+ln -s val_list.txt  train_list.txt
+cd ../../

test_tipc/static/ResNet50/benchmark_common/run_benchmark.sh

+#!/usr/bin/env bash
+# Test training benchmark for a model.
+# Usage：bash run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_process_type} ${run_mode} ${device_num}
+function _set_params(){
+    model_item=${1:-"model_item"}   # (必选) 模型 item
+    base_batch_size=${2:-"2"}       # (必选) 如果是静态图单进程，则表示每张卡上的BS，需在训练时*卡数
+    fp_item=${3:-"fp32"}            # (必选) fp32|fp16
+    run_process_type=${4:-"SingleP"} # (必选) 单进程 SingleP|多进程 MultiP
+    run_mode=${5:-"DP"}             # (必选) MP模型并行|DP数据并行|PP流水线并行|混合并行DP1-MP1-PP1|DP1-MP4-PP1
+    device_num=${6:-"N1C1"}         # (必选) 使用的卡数量，N1C1|N1C8|N4C32 （4机32卡）
+    profiling=${PROFILING:-"false"}      # (必选) Profiling  开关，默认关闭，通过全局变量传递
+    model_repo="PaddleClas"          # (必选) 模型套件的名字
+    speed_unit="samples/sec"         # (必选)速度指标单位
+    skip_steps=10                  # (必选)解析日志，跳过模型前几个性能不稳定的step
+    keyword="ips:"                 # (必选)解析日志，筛选出性能数据所在行的关键字
+    convergence_key="loss:"        # (可选)解析日志，筛选出收敛数据所在行的关键字 如：convergence_key="loss:"
+    max_epochs=${7:-"1"}           # （可选）需保证模型执行时间在5分钟内，需要修改代码提前中断的直接提PR 合入套件；或使用max_epoch参数
+    num_workers=${8:-"4"}                  # (可选)
+#   以下为通用执行命令，无特殊可不用修改
+    model_name=${model_item}_bs${base_batch_size}_${fp_item}_${run_process_type}_${run_mode}  # (必填) 且格式不要改动,与竞品名称对齐
+    device=${CUDA_VISIBLE_DEVICES//,/ }
+    arr=(${device})
+    num_gpu_devices=${#arr[*]}
+    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # （必填） TRAIN_LOG_DIR  benchmark框架设置该参数为全局变量
+    profiling_log_path=${PROFILING_LOG_DIR:-$(pwd)}  # （必填） PROFILING_LOG_DIR benchmark框架设置该参数为全局变量
+    speed_log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}
+
+    train_log_file=${run_log_path}/${model_repo}_${model_name}_${device_num}_log
+    profiling_log_file=${profiling_log_path}/${model_repo}_${model_name}_${device_num}_profiling
+    speed_log_file=${speed_log_path}/${model_repo}_${model_name}_${device_num}_speed
+}
+function _train(){
+    batch_size=${base_batch_size}  # 如果模型跑多卡单进程时,请在_train函数中计算出多卡需要的bs
+    echo "current CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}, model_name=${model_name}, device_num=${device_num}, is profiling=${profiling}"
+
+    if [ ${fp_item} = "fp32" ]; then
+        config_file="-c ppcls/configs/ImageNet/ResNet/ResNet50.yaml"
+    else
+        config_file="-c ppcls/configs/ImageNet/ResNet/ResNet50_amp_O1.yaml"
+    fi
+    if [ ${profiling} = "false" ]; then
+	profiling_config=""
+	log_file=${train_log_file}
+    else
+	profiling_config="--profiler_options=\"batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile\""
+	log_file=${profiling_log_file}
+    fi
+
+    train_cmd="${config_file} -o DataLoader.Train.sampler.batch_size=${base_batch_size} -o Global.epochs=${max_epochs} -o DataLoader.Train.loader.num_workers=${num_workers} ${profiling_config}"
+#   以下为通用执行命令，无特殊可不用修改
+    case ${run_process_type} in
+    SingleP) 
+	   train_cmd="python ppcls/static/train.py ${train_cmd}";;
+    MultiP)
+	   train_cmd="python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 ppcls/static/train.py ${train_cmd}";;
+    *) echo "choose run_process_type(SingleP or MultiP)"; exit 1;
+    esac
+    echo "train_cmd: ${train_cmd}  log_file: ${log_file}"
+    timeout 5m ${train_cmd} > ${log_file} 2>&1
+    if [ $? -ne 0 ];then
+        echo -e "${model_name}, FAIL"
+    else
+        echo -e "${model_name}, SUCCESS"
+    fi
+    kill -9 `ps -ef|grep 'python'|awk '{print $2}'`
+    if [ ${run_process_type} = "MultiP" -a -d mylog ]; then
+        rm ${log_file}
+        cp mylog/workerlog.0 ${log_file}
+    fi
+    cd ../
+}
+# 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;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 $@
+_train       # 如果只产出训练log,不解析,可取消注释
+# _run     # 该函数在run_model.sh中,执行时会调用_train; 如果不联调只产出训练log可以注掉本行,提交时需打开