{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. PP-YOLOv2模型简介\n",
    "YOLO系列作为目标检测的重要算法,采用单阶段(one-stage)方法使得检测速度大幅提升,但是速度的提升也牺牲了部分准确率作为代价。因此，如何在提升YOLOv3的准确性的同时保持推理速度成为了其实际应用时的关键问题。为同时满足准确性与高效性，PP-YOLOv2作者团队做了大量优化工作，PP-YOLOv2（R50）在COCO test数据集mAP从45.9%达到了49.5%，相较v1提升了3.6个百分点，FP32 FPS高达68.9FPS，FP16 FPS高达106.5FPS，超越了YOLOv4甚至YOLOv5！如果使用RestNet101作为骨架网络，PP-YOLOv2（R101）的mAP更高达50.3%，并且比同等精度下的YOLOv5x快15.9%！\n",
    "\n",
    "PP-YOLO模型由飞桨官方出品，是PaddleDetection优化和改进的YOLOv3的模型。\n",
    "更多关于PaddleDetection可以点击https://github.com/PaddlePaddle/PaddleDetection 进行了解。\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. 模型效果及应用场景\n",
    "### 2.1 目标检测任务：\n",
    "\n",
    "#### 2.1.1 数据集：\n",
    "\n",
    "数据集以COCO格式为主，分为训练集和测试集。\n",
    "\n",
    "#### 2.1.2 模型效果速览：\n",
    "\n",
    "PP-YOLOv2在图片上的检测效果为：\n",
    "\n",
    "<div align=\"center\">\n",
    "<img src=\"https://user-images.githubusercontent.com/23690325/198869600-b7a549db-2cc6-49b1-8009-937fb5abe992.png\"  width = \"80%\"  />\n",
    "</div>\n",
    "\n",
    "<div align=\"center\">\n",
    "<img src=\"https://user-images.githubusercontent.com/23690325/198869611-451eda5f-eda6-4717-902c-b9b06070bc72.png\"  width = \"80%\"  />\n",
    "</div>\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3. 模型如何使用\n",
    "\n",
    "### 3.1 模型推理：\n",
    "* 下载 \n",
    "\n",
    "（不在Jupyter Notebook上运行时需要将\"！\"或者\"%\"去掉。）\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "scrolled": true,
    "tags": []
   },
   "outputs": [],
   "source": [
    "# 克隆PaddleDetection仓库\n",
    "%mkdir -p ~/work\n",
    "%cd ~/work/\n",
    "!git clone https://github.com/PaddlePaddle/PaddleDetection.git\n",
    "\n",
    "# 安装其他依赖\n",
    "%cd PaddleDetection\n",
    "%mkdir -p demo_input demo_output\n",
    "!pip install -r requirements.txt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 验证是否安装成功\n",
    "如果报错，只需执行上一步操作。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "# 测试是否安装成功\n",
    "!python ppdet/modeling/tests/test_architectures.py"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 快速体验\n",
    "\n",
    "恭喜！ 您已经成功安装了PaddleDetection，接下来快速体验目标检测效果"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true,
    "tags": []
   },
   "outputs": [],
   "source": [
    "# 在GPU上预测一张图片\n",
    "!export CUDA_VISIBLE_DEVICES=0\n",
    "!wget -P demo_input -N https://paddledet.bj.bcebos.com/modelcenter/images/General/000000014439.jpg\n",
    "!python tools/infer.py -c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml -o use_gpu=true weights=https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams --infer_img=demo_input/000000014439.jpg"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "会在output文件夹下生成一个画有预测结果的同名图像。\n",
    "\n",
    "结果如下图：\n",
    "<div align=\"center\">\n",
    "<img src=\"https://bj.bcebos.com/v1/paddledet/modelcenter/images/ppyolov2_infer.jpg\"  width = \"60%\"  />\n",
    "</div>\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 3.2 模型训练：\n",
    "* 克隆PaddleDetection仓库（详见3.1）"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 准备数据集\n",
    "\n",
    "这里需要开发者自行准备数据集，以下举例假设开发者已经准备好wider_face数据集并且解压到 PaddleDetection/dataset/wider_face/下\n",
    "通过以下命令确认数据集已经准备完成。\n",
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 修改yaml配置文件\n",
    "\n",
    "\n",
    "\n",
    "修改配置文件``` configs/runtime.yml```\n",
    "\n",
    "```\n",
    "use_gpu: true # 是否使用GPU\n",
    "log_iter: 20  # 每多少个迭代次显示\n",
    "save_dir: output # 模型保存目录\n",
    "snapshot_epoch: 1 # 多少个epoch保存一次\n",
    "print_flops: false\n",
    "\n",
    "```\n",
    "修改配置文件``` configs/datasets/coco_detection.yml```\n",
    "\n",
    "```\n",
    "metric: COCO    \n",
    "num_classes: 1  # 分类个数 \n",
    "\n",
    "TrainDataset:\n",
    "  !COCODataSet\n",
    "    image_dir: WIDER_train/images   # 训练图像数据基于数据集根目录的相对路径\n",
    "    anno_path: WIDERFaceTrainCOCO.json  # 训练标注文件基于数据集根目录的相对路径\n",
    "    dataset_dir: dataset/wider_face # 数据集根目录\n",
    "    data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']\n",
    "\n",
    "EvalDataset:\n",
    "  !COCODataSet\n",
    "    image_dir: WIDER_val/images     # 测试图像数据基于数据集根目录的相对路径\n",
    "    anno_path: WIDERFaceValCOCO.json     # 测试标注文件基于数据集根目录的相对路径\n",
    "    dataset_dir: dataset/wider_face\n",
    "\n",
    "TestDataset:\n",
    "  !ImageFolder\n",
    "    anno_path: WIDERFaceValCOCO.json\n",
    "    \n",
    "```\n",
    "修改配置文件``` configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml```\n",
    "\n",
    "```\n",
    "_BASE_: [\n",
    "  '../datasets/coco_detection.yml',\n",
    "  '../runtime.yml',\n",
    "  './_base_/ppyolov2_r50vd_dcn.yml',\n",
    "  './_base_/optimizer_365e.yml',\n",
    "  './_base_/ppyolov2_reader.yml',\n",
    "]\n",
    "\n",
    "snapshot_epoch: 8   #每训练多少个epoch保存一次模型\n",
    "weights: output/ppyolov2_r50vd_dcn_365e_coco/model_final\n",
    "```\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 训练模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true,
    "tags": []
   },
   "outputs": [],
   "source": [
    "%cd ~/work/PaddleDetection/\n",
    "%env CUDA_VISIBLE_DEVICES=0\n",
    "#开始训练\n",
    "!python  tools/train.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml  --use_vdl=true "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 模型评估\n",
    "\n",
    "我们提供了```configs/ppyolo/ppyolo_test.yml```用于评估COCO test-dev2017数据集的效果，评估COCO test-dev2017数据集的效果须先从COCO数据集下载页下载test-dev2017数据集，解压到```configs/ppyolo/ppyolo_test.yml```中EvalReader.dataset中配置的路径，并使用如下命令进行评估（需附上模型平均精度AP和AR评估指标，提供图片或者表格）。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "%cd ~/work/PaddleDetection/\n",
    "%env CUDA_VISIBLE_DEVICES=0\n",
    "#训练完以后，进行评估\n",
    "!python tools/eval.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml  -o use_gpu=true"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 4. 模型原理"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 采用 Path Aggregation Network（路径聚合网络）设计 Detection Net\n",
    "\n",
    "PP-YOLOv2 采用了 FPN 的变形之一—PAN（Path Aggregation Network）来从上至下的聚合特征信息。\n",
    "\n",
    "<div align=\"center\">\n",
    "<img src=\"https://user-images.githubusercontent.com/23690325/198870025-f7f0e976-0fb0-4496-b359-b0e2a751cab1.png\"  width = \"80%\"  />\n",
    "</div>\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 采用 Mish 激活函数\n",
    "\n",
    "PP-YOLOv2的mish 激活函数应用在了 detection neck 而不是骨架网络上。\n",
    "\n",
    "* 更大的输入尺寸\n",
    "\n",
    "增加输入尺寸直接带来了目标面积的扩大。这样，网络可以更容易捕捉到小尺幅目标的信息，得到更高的性能。然而，更大的输入会带来更多的内存占用。所以在使用这个策略的同时，PP-YOLOv2同时减少 Batch Size。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 5. 注意事项\n",
    "\n",
    "不管是 PP-YOLO 还是 PP-YOLOv2，都是在寻找在产业实践中最高性价比的目标检测方案，而不是单纯的以提升单阶段目标检测的精度去堆网络和策略。关于PP-YOLOv2的论文中也特别提到，是以实验报告的角度来为业界开发者展示更多网络优化的方法，这些策略也可以被应用在其他网络的优化上，希望在给业界开发者带来更好的网络的同时，也带来更多的算法优化启发。同时，在使用PP-YOLO系列时也应当注意：\n",
    "\n",
    "\n",
    "* PP-YOLO模型使用COCO数据集中的train2017作为训练集，使用val2017和test-dev2017作为测试集，Box APtest为mAP(IoU=0.5:0.95)评估结果。\n",
    "* PP-YOLO模型训练过程中使用8 GPUs，每GPU batch size为24进行训练，如训练GPU数和batch size不使用上述配置，须参考FAQ调整学习率和迭代次数。\n",
    "* PP-YOLO模型推理速度测试采用单卡V100，batch size=1进行测试，使用CUDA 10.2, CUDNN 7.5.1，TensorRT推理速度测试使用TensorRT 5.1.2.2。\n",
    "* PP-YOLO模型FP32的推理速度测试数据为使用tools/export_model.py脚本导出模型后，使用deploy/python/infer.py脚本中的--run_benchnark参数使用Paddle预测库进行推理速度benchmark测试结果, 且测试的均为不包含数据预处理和模型输出后处理(NMS)的数据(与YOLOv4(AlexyAB)测试方法一致)。\n",
    "* TensorRT FP16的速度测试相比于FP32去除了yolo_box(bbox解码)部分耗时，即不包含数据预处理，bbox解码和NMS(与YOLOv4(AlexyAB)测试方法一致)。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 6. 相关论文以及引用信息\n",
    "（如果本模型有相关论文发表，或者是基于某些论文的结果，可以在这里\n",
    "提供Bibtex格式的参考文献。）\n",
    "```\n",
    "@article{huang2021pp,\n",
    "  title={PP-YOLOv2: A Practical Object Detector},\n",
    "  author={Huang, Xin and Wang, Xinxin and Lv, Wenyu and Bai, Xiaying and Long, Xiang and Deng, Kaipeng and Dang, Qingqing and Han, Shumin and Liu, Qiwen and Hu, Xiaoguang and others},\n",
    "  journal={arXiv preprint arXiv:2104.10419},\n",
    "  year={2021}\n",
    "}\n",
    "```\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3.8.13 ('paddle_env')",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.13"
  },
  "vscode": {
   "interpreter": {
    "hash": "864bc28e4d94d9c1c4bd0747e4313c0ab41718ab445ced17dbe1a405af5ecc64"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}