{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. 模型简介\n",
    "PP-YOLOE+是PP-YOLOE的升级版本，从大规模的obj365目标检测预训练模型入手，在大幅提升收敛速度的同时，提升了模型在COCO数据集上的速度。同时，PP-YOLOE+大幅提升了包括数据预处理在内的端到端的预测速度。关于PP-YOLOE+的更多细节可以参考我们的[官方文档](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.5/configs/ppyoloe/README_cn.md)。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. 模型效果\n",
    "PP-YOLOE+_l在COCO test-dev2017达到了53.3的mAP, 同时其速度在Tesla V100上达到了78.1 FPS。如下图所示，PP-YOLOE+_s/m/x同样具有卓越的精度速度性价比。\n",
    "<div align=\"center\">\n",
    "  <img src=\"https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.5/docs/images/ppyoloe_plus_map_fps.png\" width=500 />\n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3. 模型如何使用\n",
    "首先克隆PaddleDetection，并将数据集存放在`dataset/coco/`目录下面"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "%cd ~/work\n",
    "!git clone https://gitee.com/paddlepaddle/PaddleDetection\n",
    "%cd PaddleDetection"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 3.1 训练\n",
    "执行以下指令训练PP-YOLOE+"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "# 单卡训练\n",
    "!python tools/train.py -c configs/ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml --eval --amp\n",
    "\n",
    "# 多卡训练\n",
    "!python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml --eval --amp"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**注意：**\n",
    "- 如果需要边训练边评估，请添加`--eval`.\n",
    "- PP-YOLOE支持混合精度训练，请添加`--amp`."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 3.2 推理部署\n",
    "#### 3.2.1 使用Paddle-TRT进行推理部署\n",
    "接下来，我们将介绍PP-YOLOE如何使用Paddle Inference在TensorRT FP16模式下部署\n",
    "\n",
    "首先，参考[Paddle Inference文档](https://www.paddlepaddle.org.cn/inference/master/user_guides/download_lib.html#python)，下载并安装与你的CUDA, CUDNN和TensorRT相应的wheel包。\n",
    "\n",
    "然后，运行以下命令导出模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "!python tools/export_model.py -c configs/ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco.pdparams trt=True"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "最后，使用TensorRT FP16进行推理"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "# 推理单张图片\n",
    "CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyoloe_plus_crn_l_80e_coco --image_file=demo/000000014439_640x640.jpg --device=gpu --run_mode=trt_fp16\n",
    "\n",
    "# 推理文件夹下的所有图片\n",
    "CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyoloe_plus_crn_l_80e_coco --image_dir=demo/ --device=gpu  --run_mode=trt_fp16"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**注意：**\n",
    "- TensorRT会根据网络的定义，执行针对当前硬件平台的优化，生成推理引擎并序列化为文件。该推理引擎只适用于当前软硬件平台。如果你的软硬件平台没有发生变化，你可以设置[enable_tensorrt_engine](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.4/deploy/python/infer.py#L745)的参数`use_static=True`，这样生成的序列化文件将会保存在`output_inference`文件夹下，下次执行TensorRT时将加载保存的序列化文件。\n",
    "- PaddleDetection release/2.4及其之后的版本将支持NMS调用TensorRT，需要依赖PaddlePaddle release/2.3及其之后的版本。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 3.2.2 使用PaddleInference进行推理部署\n",
    "对于一些不支持TensorRT的AI加速硬件，我们可以直接使用PaddleInference进行部署。首先运行以下命令导出模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "!python tools/export_model.py -c configs/ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco.pdparams"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "直接使用PaddleInference进行推理"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "# 推理单张图片\n",
    "CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyoloe_plus_crn_l_80e_coco --image_file=demo/000000014439_640x640.jpg --device=gpu --run_mode=paddle\n",
    "\n",
    "# 推理文件夹下的所有图片\n",
    "CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyoloe_plus_crn_l_80e_coco --image_dir=demo/ --device=gpu  --run_mode=paddle"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 3.2.3 使用ONNX-TensorRT进行速度测试\n",
    "**使用 ONNX 和 TensorRT** 进行测速，执行以下命令："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "# 导出模型\n",
    "!python tools/export_model.py -c configs/ppyoloe/ppyoloe_plus_crn_s_80e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_s_80e_coco.pdparams exclude_nms=True trt=True\n",
    "\n",
    "# 转化成ONNX格式\n",
    "!paddle2onnx --model_dir output_inference/ppyoloe_plus_crn_s_80e_coco --model_filename model.pdmodel --params_filename model.pdiparams --opset_version 12 --save_file ppyoloe_plus_crn_s_80e_coco.onnx\n",
    "\n",
    "# 测试速度，半精度，batch_size=1\n",
    "!trtexec --onnx=./ppyoloe_plus_crn_s_80e_coco.onnx --saveEngine=./ppyoloe_s_bs1.engine --workspace=1024 --avgRuns=1000 --shapes=image:1x3x640x640,scale_factor:1x2 --fp16\n",
    "\n",
    "# 测试速度，半精度，batch_size=32\n",
    "!trtexec --onnx=./ppyoloe_plus_crn_s_80e_coco.onnx --saveEngine=./ppyoloe_s_bs32.engine --workspace=1024 --avgRuns=1000 --shapes=image:32x3x640x640,scale_factor:32x2 --fp16\n",
    "\n",
    "# 使用上边的脚本, 在T4 和 TensorRT 7.2的环境下，PPYOLOE-plus-s模型速度如下\n",
    "# batch_size=1, 2.80ms, 357fps\n",
    "# batch_size=32, 67.69ms, 472fps"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 4. 模型原理\n",
    "PP-YOLOE+和PP-YOLOE的整体结构基本一致，其相对于PP-YOLOE的改进点如下：\n",
    "- 使用大规模数据集obj365预训练模型\n",
    "- 在backbone中block分支中增加alpha参数\n",
    "- 优化端到端推理速度，提升训练收敛速度"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 5.注意事项\n",
    "**所有的命令默认运行在AI Studio的`jupyter`上, 如果运行在终端上，去掉命令开头的符号%或!**"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 6. 相关论文及引用信息\n",
    "```\n",
    "@article{xu2022pp,\n",
    "  title={PP-YOLOE: An evolved version of YOLO},\n",
    "  author={Xu, Shangliang and Wang, Xinxin and Lv, Wenyu and Chang, Qinyao and Cui, Cheng and Deng, Kaipeng and Wang, Guanzhong and Dang, Qingqing and Wei, Shengyu and Du, Yuning and others},\n",
    "  journal={arXiv preprint arXiv:2203.16250},\n",
    "  year={2022}\n",
    "}\n",
    "```"
   ]
  }
 ],
 "metadata": {
  "language_info": {
   "name": "python"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}