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/configs/rotate/ppyoloe_r/README.md
+++ b/paddle_detection/configs/rotate/ppyoloe_r/README.md
@@ -0,0 +1,178 @@
+简体中文 | [English](README_en.md)
+
+# PP-YOLOE-R
+
+## 内容
+- [简介](#简介)
+- [模型库](#模型库)
+- [使用说明](#使用说明)
+- [预测部署](#预测部署)
+- [附录](#附录)
+- [引用](#引用)
+
+## 简介
+PP-YOLOE-R是一个高效的单阶段Anchor-free旋转框检测模型。基于PP-YOLOE, PP-YOLOE-R以极少的参数量和计算量为代价，引入了一系列有用的设计来提升检测精度。在DOTA 1.0数据集上，PP-YOLOE-R-l和PP-YOLOE-R-x在单尺度训练和测试的情况下分别达到了78.14和78.27 mAP，这超越了几乎所有的旋转框检测模型。通过多尺度训练和测试，PP-YOLOE-R-l和PP-YOLOE-R-x的检测精度进一步提升至80.02和80.73 mAP。在这种情况下，PP-YOLOE-R-x超越了所有的anchor-free方法并且和最先进的anchor-based的两阶段模型精度几乎相当。此外，PP-YOLOE-R-s和PP-YOLOE-R-m通过多尺度训练和测试可以达到79.42和79.71 mAP。考虑到这两个模型的参数量和计算量，其性能也非常卓越。在保持高精度的同时，PP-YOLOE-R避免使用特殊的算子，例如Deformable Convolution或Rotated RoI Align，以使其能轻松地部署在多种多样的硬件上。在1024x1024的输入分辨率下，PP-YOLOE-R-s/m/l/x在RTX 2080 Ti上使用TensorRT FP16分别能达到69.8/55.1/48.3/37.1 FPS，在Tesla V100上分别能达到114.5/86.8/69.7/50.7 FPS。更多细节可以参考我们的[**技术报告**](https://arxiv.org/abs/2211.02386)。
+
+<div align="center">
+  <img src="../../../docs/images/ppyoloe_r_map_fps.png" width=500 />
+</div>
+
+PP-YOLOE-R相较于PP-YOLOE做了以下几点改动：
+- Rotated Task Alignment Learning
+- 解耦的角度预测头
+- 使用DFL进行角度预测
+- 可学习的门控单元
+- [ProbIoU损失函数](https://arxiv.org/abs/2106.06072)
+
+## 模型库
+
+| 模型 | Backbone | mAP | V100 TRT FP16 (FPS) | RTX 2080 Ti TRT FP16 (FPS) | Params (M) | FLOPs (G) | 学习率策略 | 角度表示 | 数据增广 | GPU数目 | 每GPU图片数目 | 模型下载 | 配置文件 |
+|:---:|:--------:|:----:|:--------------------:|:------------------------:|:----------:|:---------:|:--------:|:----------:|:-------:|:------:|:-----------:|:--------:|:------:|
+| PP-YOLOE-R-s | CRN-s | 73.82 | 114.5 | 69.8 | 8.09 | 43.46 | 3x | oc | RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_s_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_s_3x_dota.yml) |
+| PP-YOLOE-R-s | CRN-s | 79.42 | 114.5 | 69.8 | 8.09 | 43.46 | 3x | oc | MS+RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_s_3x_dota_ms.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_s_3x_dota_ms.yml) |
+| PP-YOLOE-R-m | CRN-m | 77.64 | 86.8  | 55.1 | 23.96 |127.00 | 3x | oc | RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_m_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_m_3x_dota.yml) |
+| PP-YOLOE-R-m | CRN-m | 79.71 | 86.8  | 55.1 | 23.96 |127.00 | 3x | oc | MS+RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_m_3x_dota_ms.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_m_3x_dota_ms.yml) |
+| PP-YOLOE-R-l | CRN-l | 78.14 | 69.7  | 48.3 | 53.29 |281.65 | 3x | oc | RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml) |
+| PP-YOLOE-R-l | CRN-l | 80.02 | 69.7  | 48.3 | 53.29 |281.65 | 3x | oc | MS+RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota_ms.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota_ms.yml) |
+| PP-YOLOE-R-x | CRN-x | 78.28 | 50.7  | 37.1 | 100.27|529.82 | 3x | oc | RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_x_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_x_3x_dota.yml) |
+| PP-YOLOE-R-x | CRN-x | 80.73 | 50.7  | 37.1 | 100.27|529.82 | 3x | oc | MS+RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_x_3x_dota_ms.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_x_3x_dota_ms.yml) |
+
+**注意:**
+
+- 如果**GPU卡数**或者**batch size**发生了改变，你需要按照公式 **lr<sub>new</sub> = lr<sub>default</sub> * (batch_size<sub>new</sub> * GPU_number<sub>new</sub>) / (batch_size<sub>default</sub> * GPU_number<sub>default</sub>)** 调整学习率。
+- 模型库中的模型默认使用单尺度训练单尺度测试。如果数据增广一栏标明MS，意味着使用多尺度训练和多尺度测试。如果数据增广一栏标明RR，意味着使用RandomRotate数据增广进行训练。
+- CRN表示在PP-YOLOE中提出的CSPRepResNet
+- PP-YOLOE-R的参数量和计算量是在重参数化之后计算得到，输入图像的分辨率为1024x1024
+- 速度测试使用TensorRT 8.2.3在DOTA测试集中测试2000张图片计算平均值得到。参考速度测试以复现[速度测试](#速度测试)
+
+## 使用说明
+
+参考[数据准备](../README.md#数据准备)准备数据。
+
+### 训练
+
+GPU单卡训练
+``` bash
+CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml
+```
+
+GPU多卡训练
+``` bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml
+```
+
+### 预测
+
+执行以下命令预测单张图片，图片预测结果会默认保存在`output`文件夹下面
+``` bash
+python tools/infer.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams --infer_img=demo/P0861__1.0__1154___824.png --draw_threshold=0.5
+```
+
+### DOTA数据集评估
+
+参考[DOTA Task](https://captain-whu.github.io/DOTA/tasks.html), 评估DOTA数据集需要生成一个包含所有检测结果的zip文件，每一类的检测结果储存在一个txt文件中，txt文件中每行格式为：`image_name score x1 y1 x2 y2 x3 y3 x4 y4`。将生成的zip文件提交到[DOTA Evaluation](https://captain-whu.github.io/DOTA/evaluation.html)的Task1进行评估。你可以执行以下命令得到test数据集的预测结果：
+``` bash
+python tools/infer.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams --infer_dir=/path/to/test/images --output_dir=output_ppyoloe_r --visualize=False --save_results=True
+```
+将预测结果处理成官网评估所需要的格式：
+``` bash
+python configs/rotate/tools/generate_result.py --pred_txt_dir=output_ppyoloe_r/ --output_dir=submit/ --data_type=dota10
+
+zip -r submit.zip submit
+```
+
+### 速度测试
+可以使用Paddle模式或者Paddle-TRT模式进行测速。当使用Paddle-TRT模式测速时，需要确保**TensorRT版本大于8.2, PaddlePaddle版本为develop版本**。使用Paddle-TRT进行测速，可以执行以下命令：
+
+``` bash
+# 导出模型
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams trt=True
+
+# 速度测试
+CUDA_VISIBLE_DEVICES=0 python configs/rotate/tools/inference_benchmark.py --model_dir output_inference/ppyoloe_r_crn_l_3x_dota/ --image_dir /path/to/dota/test/dir --run_mode trt_fp16
+```
+当只使用Paddle进行测速，可以执行以下命令：
+``` bash
+# 导出模型
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams
+
+# 速度测试
+CUDA_VISIBLE_DEVICES=0 python configs/rotate/tools/inference_benchmark.py --model_dir output_inference/ppyoloe_r_crn_l_3x_dota/ --image_dir /path/to/dota/test/dir --run_mode paddle
+```
+
+## 预测部署
+
+**使用Paddle**进行部署，执行以下命令：
+``` bash
+# 导出模型
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams
+
+# 预测图片
+python deploy/python/infer.py --image_file demo/P0072__1.0__0___0.png --model_dir=output_inference/ppyoloe_r_crn_l_3x_dota --run_mode=paddle --device=gpu
+```
+
+**使用Paddle-TRT进行部署**，执行以下命令：
+```
+# 导出模型
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams trt=True
+
+# 预测图片
+python deploy/python/infer.py --image_file demo/P0072__1.0__0___0.png --model_dir=output_inference/ppyoloe_r_crn_l_3x_dota --run_mode=trt_fp16 --device=gpu
+```
+
+**注意：**
+- 使用Paddle-TRT使用确保**PaddlePaddle版本为develop版本且TensorRT版本大于8.2**.
+
+**使用ONNX Runtime进行部署**，执行以下命令：
+```
+# 导出模型
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams export_onnx=True
+
+# 安装paddle2onnx
+pip install paddle2onnx
+
+# 转换成onnx模型
+paddle2onnx --model_dir output_inference/ppyoloe_r_crn_l_3x_dota --model_filename model.pdmodel --params_filename model.pdiparams --opset_version 11 --save_file ppyoloe_r_crn_l_3x_dota.onnx
+
+# 预测图片
+python configs/rotate/tools/onnx_infer.py --infer_cfg output_inference/ppyoloe_r_crn_l_3x_dota/infer_cfg.yml --onnx_file ppyoloe_r_crn_l_3x_dota.onnx --image_file demo/P0072__1.0__0___0.png
+
+```
+
+## 附录
+
+PP-YOLOE-R消融实验
+
+| 模型 | mAP | 参数量(M) | FLOPs(G) |
+| :-: | :-: | :------: | :------: |
+| Baseline | 75.61 | 50.65 | 269.09 |
+| +Rotated Task Alignment Learning | 77.24 | 50.65 | 269.09 |
+| +Decoupled Angle Prediction Head | 77.78 | 52.20 | 272.72 |
+| +Angle Prediction with DFL | 78.01 | 53.29 | 281.65 |
+| +Learnable Gating Unit for RepVGG | 78.14 | 53.29 | 281.65 |
+
+
+## 引用
+
+```
+@article{wang2022pp,
+  title={PP-YOLOE-R: An Efficient Anchor-Free Rotated Object Detector},
+  author={Wang, Xinxin and Wang, Guanzhong and Dang, Qingqing and Liu, Yi and Hu, Xiaoguang and Yu, Dianhai},
+  journal={arXiv preprint arXiv:2211.02386},
+  year={2022}
+}
+
+@article{xu2022pp,
+  title={PP-YOLOE: An evolved version of YOLO},
+  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},
+  journal={arXiv preprint arXiv:2203.16250},
+  year={2022}
+}
+
+@article{llerena2021gaussian,
+  title={Gaussian Bounding Boxes and Probabilistic Intersection-over-Union for Object Detection},
+  author={Llerena, Jeffri M and Zeni, Luis Felipe and Kristen, Lucas N and Jung, Claudio},
+  journal={arXiv preprint arXiv:2106.06072},
+  year={2021}
+}
+```
--- a/paddle_detection/configs/rotate/ppyoloe_r/README_en.md
+++ b/paddle_detection/configs/rotate/ppyoloe_r/README_en.md
@@ -0,0 +1,180 @@
+English | [简体中文](README.md)
+
+# PP-YOLOE-R
+
+## Content
+- [Introduction](#Introduction)
+- [Model Zoo](#Model-Zoo)
+- [Getting Start](#Getting-Start)
+- [Deployment](#Deployment)
+- [Appendix](#Appendix)
+- [Citations](#Citations)
+
+## Introduction
+PP-YOLOE-R is an efficient anchor-free rotated object detector. Based on PP-YOLOE, PP-YOLOE-R introduces a bag of useful tricks to improve detection precision at the expense of marginal parameters and computations.PP-YOLOE-R-l and PP-YOLOE-R-x achieve 78.14 and 78.27 mAP respectively on DOTA 1.0 dataset with single-scale training and testing, which outperform almost all other rotated object detectors. With multi-scale training and testing, the detection precision of PP-YOLOE-R-l and PP-YOLOE-R-x is further improved to 80.02 and 80.73 mAP. In this case, PP-YOLOE-R-x surpasses all anchor-free methods and demonstrates competitive performance to state-of-the-art anchor-based two-stage model. Moreover, PP-YOLOE-R-s and PP-YOLOE-R-m can achieve 79.42 and 79.71 mAP with multi-scale training and testing, which is an excellent result considering the parameters and GLOPS of these two models. While maintaining high precision, PP-YOLOE-R avoids using special operators, such as Deformable Convolution or Rotated RoI Align, to be deployed friendly on various hardware. At the input resolution of 1024$\times$1024, PP-YOLOE-R-s/m/l/x can reach 69.8/55.1/48.3/37.1 FPS on RTX 2080 Ti and 114.5/86.8/69.7/50.7 FPS on Tesla V100 GPU with TensorRT and FP16-precision. For more details, please refer to our [**technical report**](https://arxiv.org/abs/2211.02386).
+
+<div align="center">
+  <img src="../../../docs/images/ppyoloe_r_map_fps.png" width=500 />
+</div>
+
+Compared with PP-YOLOE, PP-YOLOE-R has made the following changes:
+- Rotated Task Alignment Learning
+- Decoupled Angle Prediction Head
+- Angle Prediction with DFL
+- Learnable Gating Unit for RepVGG
+- [ProbIoU Loss](https://arxiv.org/abs/2106.06072)
+
+## Model Zoo
+| Model | Backbone | mAP  | V100 TRT FP16 (FPS) | RTX 2080 Ti TRT FP16 (FPS) | Params (M) | FLOPs (G) | Lr Scheduler | Angle | Aug | GPU Number | images/GPU | download | config |
+|:-----:|:--------:|:----:|:-------------------:|:--------------------------:|:-----------:|:---------:|:--------:|:-----:|:---:|:----------:|:----------:|:--------:|:------:|
+| PP-YOLOE-R-s | CRN-s | 73.82 | 114.5 | 69.8 | 8.09 | 43.46 | 3x | oc | RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_s_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_s_3x_dota.yml) |
+| PP-YOLOE-R-s | CRN-s | 79.42 | 114.5 | 69.8 | 8.09 | 43.46 | 3x | oc | MS+RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_s_3x_dota_ms.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_s_3x_dota_ms.yml) |
+| PP-YOLOE-R-m | CRN-m | 77.64 | 86.8  | 55.1 | 23.96 |127.00 | 3x | oc | RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_m_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_m_3x_dota.yml) |
+| PP-YOLOE-R-m | CRN-m | 79.71 | 86.8  | 55.1 | 23.96 |127.00 | 3x | oc | MS+RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_m_3x_dota_ms.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_m_3x_dota_ms.yml) |
+| PP-YOLOE-R-l | CRN-l | 78.14 | 69.7  | 48.3 | 53.29 |281.65 | 3x | oc | RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml) |
+| PP-YOLOE-R-l | CRN-l | 80.02 | 69.7  | 48.3 | 53.29 |281.65 | 3x | oc | MS+RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota_ms.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota_ms.yml) |
+| PP-YOLOE-R-x | CRN-x | 78.28 | 50.7  | 37.1 | 100.27|529.82 | 3x | oc | RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_x_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_x_3x_dota.yml) |
+| PP-YOLOE-R-x | CRN-x | 80.73 | 50.7  | 37.1 | 100.27|529.82 | 3x | oc | MS+RR | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_x_3x_dota_ms.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/ppyoloe_r/ppyoloe_r_crn_x_3x_dota_ms.yml) |
+
+**Notes:**
+
+- if **GPU number** or **mini-batch size** is changed, **learning rate** should be adjusted according to the formula **lr<sub>new</sub> = lr<sub>default</sub> * (batch_size<sub>new</sub> * GPU_number<sub>new</sub>) / (batch_size<sub>default</sub> * GPU_number<sub>default</sub>)**.
+- Models in model zoo is trained and tested with single scale by default. If `MS` is indicated in the data augmentation column, it means that multi-scale training and multi-scale testing are used. If `RR` is indicated in the data augmentation column, it means that RandomRotate data augmentation is used for training.
+- CRN denotes CSPRepResNet proposed in PP-YOLOE
+- The parameters and GLOPs of PP-YOLOE-R are calculated after re-parameterization, and the resolution of the input image is 1024x1024
+- Speed is calculated and averaged by testing 2000 images on the DOTA test dataset. Refer to [Speed testing](#Speed-testing) to reproduce the results.
+
+## Getting Start
+
+Refer to [Data-Preparation](../README_en.md#Data-Preparation) to prepare data.
+
+### Training
+
+Single GPU Training
+``` bash
+CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml
+```
+
+Multiple GPUs Training
+``` bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml
+```
+
+### Inference
+
+Run the follow command to infer single image, the result of inference will be saved in `output` directory by default.
+
+``` bash
+python tools/infer.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams --infer_img=demo/P0861__1.0__1154___824.png --draw_threshold=0.5
+```
+
+### Evaluation on DOTA Dataset
+Refering to [DOTA Task](https://captain-whu.github.io/DOTA/tasks.html), You need to submit a zip file containing results for all test images for evaluation. The detection results of each category are stored in a txt file, each line of which is in the following format
+`image_id score x1 y1 x2 y2 x3 y3 x4 y4`. To evaluate, you should submit the generated zip file to the Task1 of [DOTA Evaluation](https://captain-whu.github.io/DOTA/evaluation.html). You can run the following command to get the inference results of test dataset:
+``` bash
+python tools/infer.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams --infer_dir=/path/to/test/images --output_dir=output_ppyoloe_r --visualize=False --save_results=True
+```
+Process the prediction results into the format required for the official website evaluation:
+``` bash
+python configs/rotate/tools/generate_result.py --pred_txt_dir=output_ppyoloe_r/ --output_dir=submit/ --data_type=dota10
+
+zip -r submit.zip submit
+```
+
+### Speed testing
+
+You can use Paddle mode or Paddle-TRT mode for speed testing. When using Paddle-TRT for speed testing, make sure that **the version of TensorRT is larger than 8.2 and the version of PaddlePaddle is the develop version**. Using Paddle-TRT to test speed, run following command
+
+``` bash
+# export inference model
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams trt=True
+
+# speed testing
+CUDA_VISIBLE_DEVICES=0 python configs/rotate/tools/inference_benchmark.py --model_dir output_inference/ppyoloe_r_crn_l_3x_dota/ --image_dir /path/to/dota/test/dir --run_mode trt_fp16
+```
+Using Paddle to test speed, run following command
+``` bash
+# export inference model
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams
+
+# speed testing
+CUDA_VISIBLE_DEVICES=0 python configs/rotate/tools/inference_benchmark.py --model_dir output_inference/ppyoloe_r_crn_l_3x_dota/ --image_dir /path/to/dota/test/dir --run_mode paddle
+
+```
+
+## Deployment
+
+**Using Paddle** to for deployment, run following command
+
+``` bash
+# export inference model
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams
+
+# inference single image
+python deploy/python/infer.py --image_file demo/P0072__1.0__0___0.png --model_dir=output_inference/ppyoloe_r_crn_l_3x_dota --run_mode=paddle --device=gpu
+```
+
+**Using Paddle-TRT** for deployment, run following command
+
+``` bash
+# export inference model
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams trt=True
+
+# inference single image
+python deploy/python/infer.py --image_file demo/P0072__1.0__0___0.png --model_dir=output_inference/ppyoloe_r_crn_l_3x_dota --run_mode=trt_fp16 --device=gpu
+```
+**Notes:**
+- When using Paddle-TRT for speed testing, make sure that **the version of TensorRT is larger than 8.2 and the version of PaddlePaddle is the develop version**
+
+**Using ONNX Runtime** for deployment, run following command
+
+``` bash
+# export inference model
+python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams export_onnx=True
+
+# install paddle2onnx
+pip install paddle2onnx
+
+# convert to onnx model
+paddle2onnx --model_dir output_inference/ppyoloe_r_crn_l_3x_dota --model_filename model.pdmodel --params_filename model.pdiparams --opset_version 11 --save_file ppyoloe_r_crn_l_3x_dota.onnx
+
+# inference single image
+python configs/rotate/tools/onnx_infer.py --infer_cfg output_inference/ppyoloe_r_crn_l_3x_dota/infer_cfg.yml --onnx_file ppyoloe_r_crn_l_3x_dota.onnx --image_file demo/P0072__1.0__0___0.png
+```
+
+## Appendix
+
+Ablation experiments of PP-YOLOE-R
+
+| Model | mAP | Params(M) | FLOPs(G) |
+| :-: | :-: | :------: | :------: |
+| Baseline | 75.61 | 50.65 | 269.09 |
+| +Rotated Task Alignment Learning | 77.24 | 50.65 | 269.09 |
+| +Decoupled Angle Prediction Head | 77.78 | 52.20 | 272.72 |
+| +Angle Prediction with DFL | 78.01 | 53.29 | 281.65 |
+| +Learnable Gating Unit for RepVGG | 78.14 | 53.29 | 281.65 |
+
+## Citations
+
+```
+@article{wang2022pp,
+  title={PP-YOLOE-R: An Efficient Anchor-Free Rotated Object Detector},
+  author={Wang, Xinxin and Wang, Guanzhong and Dang, Qingqing and Liu, Yi and Hu, Xiaoguang and Yu, Dianhai},
+  journal={arXiv preprint arXiv:2211.02386},
+  year={2022}
+}
+
+@article{xu2022pp,
+  title={PP-YOLOE: An evolved version of YOLO},
+  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},
+  journal={arXiv preprint arXiv:2203.16250},
+  year={2022}
+}
+
+@article{llerena2021gaussian,
+  title={Gaussian Bounding Boxes and Probabilistic Intersection-over-Union for Object Detection},
+  author={Llerena, Jeffri M and Zeni, Luis Felipe and Kristen, Lucas N and Jung, Claudio},
+  journal={arXiv preprint arXiv:2106.06072},
+  year={2021}
+}
+```
--- a/paddle_detection/configs/rotate/ppyoloe_r/_base_/optimizer_3x.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/_base_/optimizer_3x.yml
@@ -0,0 +1,19 @@
+epoch: 36
+
+LearningRate:
+  base_lr: 0.008
+  schedulers:
+    - !CosineDecay
+      max_epochs: 44
+    - !LinearWarmup
+      start_factor: 0.
+      steps: 1000
+
+OptimizerBuilder:
+  clip_grad_by_norm: 35.
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0005
+    type: L2
--- a/paddle_detection/configs/rotate/ppyoloe_r/_base_/ppyoloe_r_crn.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/_base_/ppyoloe_r_crn.yml
@@ -0,0 +1,49 @@
+architecture: YOLOv3
+norm_type: sync_bn
+use_ema: true
+ema_decay: 0.9998
+
+YOLOv3:
+  backbone: CSPResNet
+  neck: CustomCSPPAN
+  yolo_head: PPYOLOERHead
+  post_process: ~
+
+CSPResNet:
+  layers: [3, 6, 6, 3]
+  channels: [64, 128, 256, 512, 1024]
+  return_idx: [1, 2, 3]
+  use_large_stem: True
+  use_alpha: True
+
+CustomCSPPAN:
+  out_channels: [768, 384, 192]
+  stage_num: 1
+  block_num: 3
+  act: 'swish'
+  spp: true
+  use_alpha: True
+
+PPYOLOERHead:
+  fpn_strides: [32, 16, 8]
+  grid_cell_offset: 0.5
+  use_varifocal_loss: true
+  static_assigner_epoch: -1
+  loss_weight: {class: 1.0, iou: 2.5, dfl: 0.05}
+  static_assigner:
+    name: FCOSRAssigner
+    factor: 12
+    threshold: 0.23
+    boundary: [[512, 10000], [256, 512], [-1, 256]]
+  assigner:
+    name: RotatedTaskAlignedAssigner
+    topk: 13
+    alpha: 1.0
+    beta: 6.0
+  nms:
+    name: MultiClassNMS
+    nms_top_k: 2000
+    keep_top_k: -1
+    score_threshold: 0.1
+    nms_threshold: 0.1
+    normalized: False
--- a/paddle_detection/configs/rotate/ppyoloe_r/_base_/ppyoloe_r_reader.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/_base_/ppyoloe_r_reader.yml
@@ -0,0 +1,46 @@
+worker_num: 4
+image_height: &image_height 1024
+image_width: &image_width 1024
+image_size: &image_size [*image_height, *image_width]
+
+TrainReader:
+  sample_transforms:
+    - Decode: {}
+    - Poly2Array: {}
+    - RandomRFlip: {}
+    - RandomRRotate: {angle_mode: 'value', angle: [0, 90, 180, -90]}
+    - RandomRRotate: {angle_mode: 'value', angle: [30, 60], rotate_prob: 0.5}
+    - RResize: {target_size: *image_size, keep_ratio: True, interp: 2}
+    - Poly2RBox: {filter_threshold: 2, filter_mode: 'edge', rbox_type: 'oc'}
+  batch_transforms:
+    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
+    - Permute: {}
+    - PadRGT: {}
+    - PadBatch: {pad_to_stride: 32}
+  batch_size: 2
+  shuffle: true
+  drop_last: true
+  use_shared_memory: true
+  collate_batch: true
+
+EvalReader:
+  sample_transforms:
+    - Decode: {}
+    - Poly2Array: {}
+    - RResize: {target_size: *image_size, keep_ratio: True, interp: 2}
+    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
+    - Permute: {}
+  batch_transforms:
+    - PadBatch: {pad_to_stride: 32}
+  batch_size: 2
+  collate_batch: false
+
+TestReader:
+  sample_transforms:
+    - Decode: {}
+    - Resize: {target_size: *image_size, keep_ratio: True, interp: 2}
+    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
+    - Permute: {}
+  batch_transforms:
+    - PadBatch: {pad_to_stride: 32}
+  batch_size: 2
--- a/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml
@@ -0,0 +1,15 @@
+_BASE_: [
+  '../../datasets/dota.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/ppyoloe_r_reader.yml',
+  '_base_/ppyoloe_r_crn.yml'
+]
+
+log_iter: 50
+snapshot_epoch: 1
+weights: output/ppyoloe_r_crn_l_3x_dota/model_final
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_l_pretrained.pdparams
+depth_mult: 1.0
+width_mult: 1.0
--- a/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota_ms.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota_ms.yml
@@ -0,0 +1,15 @@
+_BASE_: [
+  '../../datasets/dota_ms.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/ppyoloe_r_reader.yml',
+  '_base_/ppyoloe_r_crn.yml'
+]
+
+log_iter: 50
+snapshot_epoch: 1
+weights: output/ppyoloe_r_crn_l_3x_dota/model_final
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_l_pretrained.pdparams
+depth_mult: 1.0
+width_mult: 1.0
--- a/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_m_3x_dota.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_m_3x_dota.yml
@@ -0,0 +1,15 @@
+_BASE_: [
+  '../../datasets/dota.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/ppyoloe_r_reader.yml',
+  '_base_/ppyoloe_r_crn.yml'
+]
+
+log_iter: 50
+snapshot_epoch: 1
+weights: output/ppyoloe_r_crn_m_3x_dota/model_final
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_m_pretrained.pdparams
+depth_mult: 0.67
+width_mult: 0.75
--- a/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_m_3x_dota_ms.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_m_3x_dota_ms.yml
@@ -0,0 +1,15 @@
+_BASE_: [
+  '../../datasets/dota_ms.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/ppyoloe_r_reader.yml',
+  '_base_/ppyoloe_r_crn.yml'
+]
+
+log_iter: 50
+snapshot_epoch: 1
+weights: output/ppyoloe_r_crn_m_3x_dota/model_final
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_m_pretrained.pdparams
+depth_mult: 0.67
+width_mult: 0.75
--- a/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_s_3x_dota.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_s_3x_dota.yml
@@ -0,0 +1,15 @@
+_BASE_: [
+  '../../datasets/dota.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/ppyoloe_r_reader.yml',
+  '_base_/ppyoloe_r_crn.yml'
+]
+
+log_iter: 50
+snapshot_epoch: 1
+weights: output/ppyoloe_r_crn_s_3x_dota/model_final
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_s_pretrained.pdparams
+depth_mult: 0.33
+width_mult: 0.50
--- a/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_s_3x_dota_ms.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_s_3x_dota_ms.yml
@@ -0,0 +1,15 @@
+_BASE_: [
+  '../../datasets/dota_ms.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/ppyoloe_r_reader.yml',
+  '_base_/ppyoloe_r_crn.yml'
+]
+
+log_iter: 50
+snapshot_epoch: 1
+weights: output/ppyoloe_r_crn_s_3x_dota/model_final
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_s_pretrained.pdparams
+depth_mult: 0.33
+width_mult: 0.50
--- a/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_x_3x_dota.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_x_3x_dota.yml
@@ -0,0 +1,15 @@
+_BASE_: [
+  '../../datasets/dota.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/ppyoloe_r_reader.yml',
+  '_base_/ppyoloe_r_crn.yml'
+]
+
+log_iter: 50
+snapshot_epoch: 1
+weights: output/ppyoloe_r_crn_x_3x_dota/model_final
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_x_pretrained.pdparams
+depth_mult: 1.33
+width_mult: 1.25
--- a/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_x_3x_dota_ms.yml
+++ b/paddle_detection/configs/rotate/ppyoloe_r/ppyoloe_r_crn_x_3x_dota_ms.yml
@@ -0,0 +1,15 @@
+_BASE_: [
+  '../../datasets/dota_ms.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/ppyoloe_r_reader.yml',
+  '_base_/ppyoloe_r_crn.yml'
+]
+
+log_iter: 50
+snapshot_epoch: 1
+weights: output/ppyoloe_r_crn_x_3x_dota/model_final
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_x_pretrained.pdparams
+depth_mult: 1.33
+width_mult: 1.25