更换文档检测模型

2024-08-27 14:42:45 +08:00
parent aea6f19951
commit 1514e09c40
2072 changed files with 254336 additions and 4967 deletions
--- a/paddle_detection/benchmark/README.md
+++ b/paddle_detection/benchmark/README.md
@@ -0,0 +1,47 @@
+# 通用检测benchmark测试脚本说明
+
+```
+├── benchmark
+│   ├── analysis_log.py
+│   ├── prepare.sh
+│   ├── README.md
+│   ├── run_all.sh
+│   ├── run_benchmark.sh
+```
+
+## 脚本说明
+
+### prepare.sh
+相关数据准备脚本，完成数据、模型的自动下载
+### run_all.sh
+主要运行脚本，可完成所有相关模型的测试方案
+### run_benchmark.sh
+单模型运行脚本，可完成指定模型的测试方案
+
+## Docker 运行环境
+* docker image: registry.baidubce.com/paddlepaddle/paddle:2.1.2-gpu-cuda10.2-cudnn7
+* paddle = 2.1.2
+* python = 3.7
+
+## 运行benchmark测试
+
+### 运行所有模型
+```
+git clone https://github.com/PaddlePaddle/PaddleDetection.git
+cd PaddleDetection
+bash benchmark/run_all.sh
+```
+
+### 运行指定模型
+* Usage：bash run_benchmark.sh ${run_mode} ${batch_size} ${fp_item} ${max_epoch} ${model_name}
+* model_name: faster_rcnn, fcos, deformable_detr, gfl, hrnet, higherhrnet, solov2, jde, fairmot
+```
+git clone https://github.com/PaddlePaddle/PaddleDetection.git
+cd PaddleDetection
+bash benchmark/prepare.sh
+
+# 单卡
+CUDA_VISIBLE_DEVICES=0 bash benchmark/run_benchmark.sh sp 2 fp32 1 faster_rcnn
+# 多卡
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash benchmark/run_benchmark.sh mp 2 fp32 1 faster_rcnn
+```
--- a/paddle_detection/benchmark/configs/faster_rcnn_r50_fpn_1x_coco.yml
+++ b/paddle_detection/benchmark/configs/faster_rcnn_r50_fpn_1x_coco.yml
@@ -0,0 +1,48 @@
+_BASE_: [
+  '../../configs/datasets/coco_detection.yml',
+  '../../configs/runtime.yml',
+  '../../configs/faster_rcnn/_base_/optimizer_1x.yml',
+  '../../configs/faster_rcnn/_base_/faster_rcnn_r50_fpn.yml',
+]
+weights: output/faster_rcnn_r50_fpn_1x_coco/model_final
+
+worker_num: 2
+TrainReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 2, target_size: [800, 1333], keep_ratio: True}
+  - RandomFlip: {prob: 0.5}
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - Permute: {}
+  batch_transforms:
+  - PadBatch: {pad_to_stride: 32}
+  batch_size: 1
+  shuffle: true
+  drop_last: true
+  collate_batch: false
+
+
+EvalReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 2, target_size: [800, 1333], keep_ratio: True}
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - Permute: {}
+  batch_transforms:
+  - PadBatch: {pad_to_stride: 32}
+  batch_size: 1
+  shuffle: false
+  drop_last: false
+
+
+TestReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 2, target_size: [800, 1333], keep_ratio: True}
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - Permute: {}
+  batch_transforms:
+  - PadBatch: {pad_to_stride: 32}
+  batch_size: 1
+  shuffle: false
+  drop_last: false
--- a/paddle_detection/benchmark/prepare.sh
+++ b/paddle_detection/benchmark/prepare.sh
@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+
+pip install -U pip Cython
+pip install -r requirements.txt
+
+mv ./dataset/coco/download_coco.py . && rm -rf ./dataset/coco/* && mv ./download_coco.py ./dataset/coco/
+# prepare lite train data
+wget -nc -P ./dataset/coco/ https://paddledet.bj.bcebos.com/data/coco_benchmark.tar
+cd ./dataset/coco/ && tar -xvf coco_benchmark.tar && mv -u coco_benchmark/* .
+rm -rf coco_benchmark/
+
+cd ../../
+rm -rf ./dataset/mot/*
+# prepare mot mini train data
+wget -nc -P ./dataset/mot/ https://paddledet.bj.bcebos.com/data/mot_benchmark.tar
+cd ./dataset/mot/ && tar -xvf mot_benchmark.tar && mv -u mot_benchmark/* .
+rm -rf mot_benchmark/
--- a/paddle_detection/benchmark/run_all.sh
+++ b/paddle_detection/benchmark/run_all.sh
@@ -0,0 +1,47 @@
+# Use docker: paddlepaddle/paddle:latest-gpu-cuda10.1-cudnn7  paddle=2.1.2  python3.7
+#
+# Usage:
+#   git clone https://github.com/PaddlePaddle/PaddleDetection.git
+#   cd PaddleDetection
+#   bash benchmark/run_all.sh
+log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}  #  benchmark系统指定该参数,不需要跑profile时,log_path指向存speed的目录
+
+# run prepare.sh
+bash benchmark/prepare.sh
+
+model_name_list=(faster_rcnn fcos deformable_detr gfl hrnet higherhrnet solov2 jde fairmot)
+fp_item_list=(fp32)
+max_epoch=2
+
+for model_item in ${model_name_list[@]}; do
+      for fp_item in ${fp_item_list[@]}; do
+          case ${model_item} in
+              faster_rcnn) bs_list=(1 8) ;;
+              fcos) bs_list=(2) ;;
+              deformable_detr) bs_list=(2) ;;
+              gfl) bs_list=(2) ;;
+              hrnet) bs_list=(64) ;;
+              higherhrnet) bs_list=(20) ;;
+              solov2) bs_list=(2) ;;
+              jde) bs_list=(4) ;;
+              fairmot) bs_list=(6) ;;
+              *) echo "wrong model_name"; exit 1;
+          esac
+          for bs_item in ${bs_list[@]}
+            do
+            run_mode=sp
+            log_name=detection_${model_item}_bs${bs_item}_${fp_item}   # 如:clas_MobileNetv1_mp_bs32_fp32_8
+            echo "index is speed, 1gpus, begin, ${log_name}"
+            CUDA_VISIBLE_DEVICES=0 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} \
+             ${fp_item} ${max_epoch} ${model_item} | tee ${log_path}/${log_name}_speed_1gpus 2>&1
+            sleep 60
+
+            run_mode=mp
+            log_name=detection_${model_item}_bs${bs_item}_${fp_item}   # 如:clas_MobileNetv1_mp_bs32_fp32_8
+            echo "index is speed, 8gpus, run_mode is multi_process, begin, ${log_name}"
+            CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash benchmark/run_benchmark.sh ${run_mode} \
+             ${bs_item} ${fp_item} ${max_epoch} ${model_item}| tee ${log_path}/${log_name}_speed_8gpus8p 2>&1
+            sleep 60
+            done
+      done
+done
--- a/paddle_detection/benchmark/run_benchmark.sh
+++ b/paddle_detection/benchmark/run_benchmark.sh
@@ -0,0 +1,92 @@
+#!/usr/bin/env bash
+set -xe
+# Usage：CUDA_VISIBLE_DEVICES=0 bash benchmark/run_benchmark.sh ${run_mode} ${batch_size} ${fp_item} ${max_epoch} ${model_name}
+python="python3.7"
+# Parameter description
+function _set_params(){
+    run_mode=${1:-"sp"}            # sp|mp
+    batch_size=${2:-"2"}
+    fp_item=${3:-"fp32"}           # fp32|fp16
+    max_epoch=${4:-"1"}
+    model_item=${5:-"model_item"}
+    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}
+# 添加日志解析需要的参数
+    base_batch_size=${batch_size}
+    mission_name="目标检测"
+    direction_id="0"
+    ips_unit="images/s"
+    skip_steps=10                     # 解析日志，有些模型前几个step耗时长，需要跳过                                    (必填)
+    keyword="ips:"                 # 解析日志，筛选出数据所在行的关键字                                             (必填)
+    index="1"
+    model_name=${model_item}_bs${batch_size}_${fp_item}
+
+    device=${CUDA_VISIBLE_DEVICES//,/ }
+    arr=(${device})
+    num_gpu_devices=${#arr[*]}
+    log_file=${run_log_path}/${model_item}_${run_mode}_bs${batch_size}_${fp_item}_${num_gpu_devices}
+}
+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"
+
+    # set runtime params
+    set_optimizer_lr_sp=" "
+    set_optimizer_lr_mp=" "
+    # parse model_item
+    case ${model_item} in
+        faster_rcnn) model_yml="benchmark/configs/faster_rcnn_r50_fpn_1x_coco.yml"
+            set_optimizer_lr_sp="LearningRate.base_lr=0.001" ;;
+        fcos) model_yml="configs/fcos/fcos_r50_fpn_1x_coco.yml"
+            set_optimizer_lr_sp="LearningRate.base_lr=0.001" ;;
+        deformable_detr) model_yml="configs/deformable_detr/deformable_detr_r50_1x_coco.yml" ;;
+        gfl) model_yml="configs/gfl/gfl_r50_fpn_1x_coco.yml"
+            set_optimizer_lr_sp="LearningRate.base_lr=0.001" ;;
+        hrnet) model_yml="configs/keypoint/hrnet/hrnet_w32_256x192.yml" ;;
+        higherhrnet) model_yml="configs/keypoint/higherhrnet/higherhrnet_hrnet_w32_512.yml" ;;
+        solov2) model_yml="configs/solov2/solov2_r50_fpn_1x_coco.yml" ;;
+        jde) model_yml="configs/mot/jde/jde_darknet53_30e_1088x608.yml" ;;
+        fairmot) model_yml="configs/mot/fairmot/fairmot_dla34_30e_1088x608.yml" ;;
+        *) echo "Undefined model_item"; exit 1;
+    esac
+
+    set_batch_size="TrainReader.batch_size=${batch_size}"
+    set_max_epoch="epoch=${max_epoch}"
+    set_log_iter="log_iter=1"
+    if [ ${fp_item} = "fp16" ]; then
+        set_fp_item="--fp16"
+    else
+        set_fp_item=" "
+    fi
+
+    case ${run_mode} in
+        sp) train_cmd="${python} -u tools/train.py -c ${model_yml} ${set_fp_item} \
+            -o ${set_batch_size} ${set_max_epoch} ${set_log_iter} ${set_optimizer_lr_sp}" ;;
+        mp) rm -rf mylog
+            train_cmd="${python} -m paddle.distributed.launch --log_dir=./mylog \
+            --gpus=${CUDA_VISIBLE_DEVICES} tools/train.py -c ${model_yml} ${set_fp_item} \
+            -o ${set_batch_size} ${set_max_epoch} ${set_log_iter} ${set_optimizer_lr_mp}"
+            log_parse_file="mylog/workerlog.0" ;;
+        *) echo "choose run_mode(sp or mp)"; exit 1;
+    esac
+
+    timeout 15m ${train_cmd} > ${log_file} 2>&1
+    if [ $? -ne 0 ];then
+        echo -e "${train_cmd}, FAIL"
+        export job_fail_flag=1
+    else
+        echo -e "${train_cmd}, 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;该脚本在联调时可从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 $@
+# _train       # 如果只想产出训练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; 如果不联调只想要产出训练log可以注掉本行,提交时需打开
+