Initial commit

.gitignore

+**/__pycache__
+*.pth
+backup/
+proxy.sh
+datasets/annotations/
+datasets/train2017/
+datasets/val2017/
+logs/
+*.onnx
+*.log
\ No newline at end of file

LICENSE

+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

+# RT-DETR
+## 论文
+`DETRs Beat YOLOs on Real-time Object Detection`
+- https://arxiv.org/abs/2304.08069
+## 模型结构
+RT-DETR是一种先进的实时物体检测器，它建立在视觉转换器（Vision Transformer）架构之上：
+模型利用主干网络的最后三个阶段的输出特征{S3，S4，S5}作为编码器的输入；
+混合编码器通过度内特征交互（AIFI）和跨尺度特征融合模块（CCFM）将多尺度特征转换成图像特征序列；
+随后采用loU感知查询选择从编码器输出的特征序列中选择固定数量的特征，作为解码器的初始目标查询；
+最后解码器通过辅助预测头迭代优化目标查询，生成边界框和置信度得分。
+<div align=center>
+    <img src="./doc/RT-DETR.PNG"/>
+</div>
+
+## 算法原理
+1、主干网络：对于 backbone 部分，采用了经典的 ResNet 和可缩放的 HGNetv2 两种，两种 backbone 各训练了两个版本 ，以 HGNetv2 为 backbone 的 RT-DETR 包括 L 和 X 版本，以ResNet 为 backbone 的RT-DETR 则包括 RT-DETR-R50 和 RT-DETR-R101 。
+RT-DETR-R50/101 做主干方便和现有的 DETR 变体进行对比，而 RT-DETR-HGNet-L/X 则用来和现有的实时检测器进行对比，值得注意的是，HGNetv2是由百度自家研发的主干结构。
+与YOLO相似的地方在于，RT-DETR最终会输出三种不同尺寸的特征图，它们相对于输入图像的分辨率下采样倍数分别是 8 倍、16 倍和 32 倍。
+
+2、混合编码器：RT-DETR 采用了一层 Transformer 的 Encoder ，其包括度内特征交互（AIFI）和跨尺度特征融合模块（CCFM）两部分。
+它首先将二维的 S5 特征拉成向量，然后交给 AIFI 模块处理，其数学过程就是多头自注意力与 FFN，随后，再将输出 Reshape 回二维，记作 F5，以便去完成后续的所谓的“跨尺度特征融合”。
+CCFM模块是由 2 个 1×1 卷积和 N 个 RepBlock 构成的，通过调整 CCFM 中 RepBlock 的数量和 Encoder 的编码维度分别控制 Hybrid Encoder 的深度和宽度，同时对 backbone 进行相应的调整即可实现检测器的缩放。
+
+3、loU：过在训练期间约束检测器对高 IoU 的特征产生高分类分数，对低 IoU 的特征产生低分类分数。从而使得模型根据分类分数选择的 Top-K 特征对应的预测框同时具有高分类分数和高 IoU 分数。
+
+4、解码器：支持不同层数的灵活推理，无需重训练。
+
+<div align=center>
+    <img src="./doc/RT-DETR.PNG"/>
+</div>
+
+## 代码改动说明
+项目要求torch==2.0.1、torchvision==0.15.2，dcu的torch==2.1.0、torchvision==0.16.0版本过高。
+问题主要集中在torchvison.datapoints、torchvison.transformers等库的调用中，其中torchvision的datapoints依赖库在高版本中完全被移除了。
+因此进行了代码适配修改，根据api替换了一遍，由于修改过多不再一一展示，可在仓库内搜索"TODO"查看。
+ps：仓库中是改动后的代码，不需再次修改
+
+## 环境配置
+### Docker（方法一）
+```
+docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-ubuntu20.04-dtk24.04.2-py3.10
+docker run -it --name=RT-DETR --network=host --privileged=true --device=/dev/kfd --device=/dev/dri --shm-size=16G --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v /path/your_code_data:/home/RT-DETR -v /opt/hyhal/:/opt/hyhal/:ro <imageID> bash  # <imageID>为以上拉取的docker的镜像ID替换
+
+cd RT-DETR
+# 安装依赖
+pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
+pip install -r requirements.txt
+```
+### Dockerfile（方法二）
+```
+docker build --no-cache -t rtdetr:latest .
+docker run -it --name=RT-DETR --network=host --privileged=true --device=/dev/kfd --device=/dev/dri --shm-size=16G --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v /path/your_code_data:/home/RT-DETR -v /opt/hyhal/:/opt/hyhal/:ro rtdetr /bin/bash
+
+cd RT-DETR
+# 安装依赖
+pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
+pip install -r requirements.txt
+```
+### Anaconda（方法三）
+1、关于本项目DCU显卡所需的特殊深度学习库可从光合开发者社区下载安装: https://developer.hpccube.com/tool/
+```
+DTK软件栈：dtk24.04.2
+python：python3.10
+pytorch：2.1.0
+torchvision：0.16.0
+onnxruntime：1.15.0
+```
+`Tips：以上dtk软件栈、python、pytorch等DCU相关工具版本需要严格一一对应`
+
+2、其他非特殊库直接按照下面步骤进行安装
+```
+cd RT-DETR
+# 安装依赖
+pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
+pip install -r requirements.txt
+```
+
+## 数据集
+### 训练数据集
+`Coco2017`
+仅需要annotations_trainval2017.zip、val2017.zip、train2017.zip作为数据集。可通过[scnet](http://113.200.138.88:18080/aidatasets/coco2017) 或[官网链接](https://cocodataset.org/#download)进行下载，下载后的压缩包需要解压缩。通过官网下载和解压数据集的代码如下：
+
+ps：本仓库内准备了小数据集供训练测试，位于。。。。。
+
+```
+cd RT-DETR/datasets
+wget -i url.txt
+# 解压
+apt-get update
+apt-get install unzip
+unzip annotations_trainval2017.zip;unzip val2017.zip;unzip train2017.zip
+```
+数据集目录结构如下：
+```
+RT-DETR/datasets:
+ ── train2017
+    │   └── ...
+ ── val2017
+    │   └── ...
+ ── annotations
+    │   └── ...
+```
+### 推理数据集
+推理测试所用数据已保存在RT-DETR/datasets/000000033109.jpg
+
+## 训练
+运行代码时会默认下载辅助模型，存储目录为/root/.cache/torch/hub/checkpoints/
+
+### 单机多卡
+```
+cd RT-DETR
+HIP_VISIBLE_DEVICES=0,1,2,3;torchrun --nproc_per_node=4 rtdetr_pytorch/tools/train.py -c rtdetr_pytorch/configs/rtdetr/rtdetr_r18vd_6x_coco.yml
+# -c 配置文件路径
+# -r 恢复训练的权重加载路径
+```
+注：如果挂载目录不是"/home/RT-DETR",需要修改RT-DETR/rtdetr_pytorch/configs/dataset/coco_detection.yml和RT-DETR/rtdetr_pytorch/configs/rtdetr/rtdetr_r18vd_6x_coco.yml中的对应路径。
+
+## 推理
+权重可通过[scnet](http://113.200.138.88:18080/aimodels/findsource-dependency/rt-detrv1/-/tree/main/pytorch) 或 [官网链接](https://github.com/lyuwenyu/storage/releases/tag/v0.1)进行下载，通过官网下载的代码如下：
+```
+cd RT-DETR
+wget -i model/url.txt -P model/
+```
+onnx_infer.py内可以修改onnx模型目录和数据集目录，推理代码如下：
+```
+cd RT-DETR
+
+# 1、导出onnx文件，以r18vd_6x为例
+HIP_VISIBLE_DEVICES=0 python rtdetr_pytorch/tools/export_onnx.py \
+	-c rtdetr_pytorch/configs/rtdetr/rtdetr_r101vd_6x_coco.yml \
+	-r model/rtdetr_r101vd_6x_coco_from_paddle.pth \
+	-f model/onnx/rtdetr_r101vd_6x_coco.onnx \
+	--check
+# -c 模型配置文件的存储目录
+# -r 源模型的目录
+# -f onnx模型的导出目录
+# --check 检查onnx模型是否导出完成
+
+# 2、运行推理
+python onnx_infer.py
+```
+注：如果挂载目录不是"/home/RT-DETR",需要修改RT-DETR/rtdetr_pytorch/configs/dataset/coco_detection.yml和RT-DETR/rtdetr_pytorch/configs/rtdetr/rtdetr_r18vd_6x_coco.yml中的对应路径。
+
+## 评估
+评估代码如下：
+```
+CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 rtdetr_pytorch/tools/train.py \
+	-c rtdetr_pytorch/configs/rtdetr/rtdetr_r101vd_6x_coco.yml \
+	-r model/rtdetr_r101vd_6x_coco_from_paddle.pth \
+	--test-only
+# -c 配置文件路径
+# -r 权重加载路径
+# --test-only 只进行评估
+```
+
+## result
+默认推理结果为：
+<div align=center>
+    <img src="./doc/inference_result.png"/>
+</div>
+
+### 精度
+|                              | 测试参数                        | 软件栈     | final loss |
+| ---------------------------- | ------------------------------- | ---------- | ---------- |
+| A800 * 4 （80G，1410 Mhz）   | config=rtdetr_r18vd_6x_coco.yml | cuda11.8   | 9.390466   |
+| k100ai * 4 （64G，1400 Mhz） | config=rtdetr_r18vd_6x_coco.yml | dtk24.04.2 | 9.400760   |
+
+## 应用场景
+### 算法类别
+`目标检测`
+### 热点应用行业
+`制造,交通,电商,广媒,医疗`
+## 预训练权重
+- http://113.200.138.88:18080/aimodels/findsource-dependency/rt-detrv1/-/tree/main/pytorch
+- https://github.com/lyuwenyu/storage/releases/tag/v0.1
+## 源码仓库及问题反馈
+- https://developer.sourcefind.cn/codes/modelzoo/rt-detr_pytorch
+## 参考资料
+- https://github.com/huangb23/VTimeLLM
\ No newline at end of file

README_cn.md

+简体中文 | [English](README.md)
+
+# RT-DETR 
+
+This is the official implementation of the paper "[DETRs Beat YOLOs on Real-time Object Detection](https://arxiv.org/abs/2304.08069)".
+
+<div align="center">
+  <img src="https://github.com/lyuwenyu/RT-DETR/assets/77494834/0ede1dc1-a854-43b6-9986-cf9090f11a61" width=500 >
+</div>
+
+## 最新动态
+
+- 发布RT-DETR-R50, RT-DETR-R101模型
+- 发布RT-DETR-R50-m模型（scale模型的范例）
+- 发布RT-DETR-R34, RT-DETR-R18模型
+- 发布RT-DETR-L, RT-DETR-X模型
+
+
+## 代码仓库
+- [RT-DETR-paddle](./rtdetr_paddle)
+- [RT-DETR--pytorch](./rtdetr_pytorch)
+
+
+## 简介
+<!-- We propose a **R**eal-**T**ime **DE**tection **TR**ansformer (RT-DETR), the first real-time end-to-end object detector to our best knowledge. Specifically, we design an efficient hybrid encoder to efficiently process multi-scale features by decoupling the intra-scale interaction and cross-scale fusion, and propose IoU-aware query selection to improve the initialization of object queries. In addition, our proposed detector supports flexibly adjustment of the inference speed by using different decoder layers without the need for retraining, which facilitates the practical application of real-time object detectors. Our RT-DETR-L achieves 53.0% AP on COCO val2017 and 114 FPS on T4 GPU, while RT-DETR-X achieves 54.8% AP and 74 FPS, outperforming all YOLO detectors of the same scale in both speed and accuracy. Furthermore, our RT-DETR-R50 achieves 53.1% AP and 108 FPS, outperforming DINO-Deformable-DETR-R50 by 2.2% AP in accuracy and by about 21 times in FPS.  -->
+RT-DETR是第一个实时端到端目标检测器。具体而言，我们设计了一个高效的混合编码器，通过解耦尺度内交互和跨尺度融合来高效处理多尺度特征，并提出了IoU感知的查询选择机制，以优化解码器查询的初始化。此外，RT-DETR支持通过使用不同的解码器层来灵活调整推理速度，而不需要重新训练，这有助于实时目标检测器的实际应用。RT-DETR-R50在COCO val2017上实现了53.1%的AP，在T4 GPU上实现了108FPS，RT-DETR-R101实现了54.3%的AP和74FPS，在速度和精度方面都优于相同规模的所有YOLO检测器。使用Objects365预训练之后, RT-DETR-R50 和 RT-DETR-R101 分别实现了 55.3% 和 56.2% AP的精度.
+若要了解更多细节，请参考我们的论文[paper](https://arxiv.org/abs/2304.08069).
+
+<div align="center">
+  <img src="https://github.com/lyuwenyu/RT-DETR/assets/77494834/c211a164-ddce-4084-8b71-fb73f29f363b" width=500 >
+</div>
+
+## 引用RT-DETR
+如果需要在你的研究中使用RT-DETR，请通过以下方式引用我们的论文：
+```
+@misc{lv2023detrs,
+      title={DETRs Beat YOLOs on Real-time Object Detection},
+      author={Yian Zhao and Wenyu Lv and Shangliang Xu and Jinman Wei and Guanzhong Wang and Qingqing Dang and Yi Liu and Jie Chen},
+      year={2023},
+      eprint={2304.08069},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+```

README_origin.md

+English | [简体中文](README_cn.md)
+
+
+<h2 align="center">RT-DETR: DETRs Beat YOLOs on Real-time Object Detection</h2>
+<p align="center">
+    <!-- <a href="https://github.com/lyuwenyu/RT-DETR/blob/main/LICENSE">
+        <img alt="license" src="https://img.shields.io/badge/LICENSE-Apache%202.0-blue">
+    </a> -->
+    <a href="https://github.com/lyuwenyu/RT-DETR/blob/main/LICENSE">
+        <img alt="license" src="https://img.shields.io/github/license/lyuwenyu/RT-DETR">
+    </a>
+    <a href="https://github.com/lyuwenyu/RT-DETR/pulls">
+        <img alt="prs" src="https://img.shields.io/github/issues-pr/lyuwenyu/RT-DETR">
+    </a>
+    <a href="https://github.com/lyuwenyu/RT-DETR/issues">
+        <img alt="issues" src="https://img.shields.io/github/issues/lyuwenyu/RT-DETR?color=pink">
+    </a>
+    <a href="https://github.com/lyuwenyu/RT-DETR">
+        <img alt="issues" src="https://img.shields.io/github/stars/lyuwenyu/RT-DETR">
+    </a>
+    <a href="https://arxiv.org/abs/2304.08069">
+        <img alt="arXiv" src="https://img.shields.io/badge/arXiv-2304.08069-red">
+    </a>
+    <a href="mailto: lyuwenyu@foxmail.com">
+        <img alt="emal" src="https://img.shields.io/badge/contact_me-email-yellow">
+    </a>
+</p>
+
+---
+
+<div align="center">
+  <img src="https://github.com/lyuwenyu/RT-DETR/assets/77494834/0ede1dc1-a854-43b6-9986-cf9090f11a61" width=500 >
+</div>
+
+
+
+
+This is the official implementation of the paper "[DETRs Beat YOLOs on Real-time Object Detection](https://arxiv.org/abs/2304.08069)".
+
+
+## 🚀 Updates
+- \[2024.02.27\] Our work has been accepted to CVPR 2024!
+- \[2024.01.23\] Fix difference on data augmentation with paper in rtdetr_pytorch [#84](https://github.com/lyuwenyu/RT-DETR/commit/5dc64138e439247b4e707dd6cebfe19d8d77f5b1).
+- \[2023.11.07\] Add pytorch ✅ *rtdetr_r34vd* for requests [#107](https://github.com/lyuwenyu/RT-DETR/issues/107), [#114](https://github.com/lyuwenyu/RT-DETR/issues/114).
+- \[2023.11.05\] Upgrade the logic of `remap_mscoco_category` to facilitate training of custom datasets, see detils in [*Train custom data*](./rtdetr_pytorch/) part. [#81](https://github.com/lyuwenyu/RT-DETR/commit/95fc522fd7cf26c64ffd2ad0c622c392d29a9ebf).
+- \[2023.10.23\] Add [*discussion for deployments*](https://github.com/lyuwenyu/RT-DETR/issues/95), supported onnxruntime, TensorRT, openVINO.
+- \[2023.10.12\] Add tuning code for pytorch version, now you can tuning rtdetr based on pretrained weights.
+- \[2023.09.19\] Upload ✅ [*pytorch weights*](https://github.com/lyuwenyu/RT-DETR/issues/42) convert from paddle version.
+- \[2023.08.24] Release RT-DETR-R18 pretrained models on objects365. *49.2 mAP* and *217 FPS*.
+- \[2023.08.22\] Upload ✅ [*rtdetr_pytorch*](./rtdetr_pytorch/) source code. Please enjoy it!
+- \[2023.08.15\] Release RT-DETR-R101 pretrained models on objects365. *56.2 mAP* and *74 FPS*.
+- \[2023.07.30\] Release RT-DETR-R50 pretrained models on objects365. *55.3 mAP* and *108 FPS*.
+- \[2023.07.28\] Fix some bugs, and add some comments. [1](https://github.com/lyuwenyu/RT-DETR/pull/14), [2](https://github.com/lyuwenyu/RT-DETR/commit/3b5cbcf8ae3b907e6b8bb65498a6be7c6736eabc).
+- \[2023.07.13\] Upload ✅ [*training logs on coco*](https://github.com/lyuwenyu/RT-DETR/issues/8).
+- \[2023.05.17\] Release RT-DETR-R18, RT-DETR-R34, RT-DETR-R50-m（example for scaled).
+- \[2023.04.17\] Release RT-DETR-R50, RT-DETR-R101, RT-DETR-L, RT-DETR-X.
+
+## 📍 Implementations
+- 🔥 RT-DETR paddle: [code](./rtdetr_paddle), [weights](./rtdetr_paddle)
+- 🔥 RT-DETR pytorch: [code](./rtdetr_pytorch), [weights](./rtdetr_pytorch)
+
+
+| Model | Epoch | Input shape | Dataset | $AP^{val}$ | $AP^{val}_{50}$| Params(M) | FLOPs(G) | T4 TensorRT FP16(FPS)
+|:---:|:---:|:---:| :---:|:---:|:---:|:---:|:---:|:---:|
+| RT-DETR-R18 | 6x | 640 | COCO | 46.5 | 63.8 | 20 | 60 | 217 |
+| RT-DETR-R34 | 6x | 640 | COCO | 48.9 | 66.8 | 31 | 92 | 161 |
+| RT-DETR-R50-m | 6x | 640 | COCO | 51.3 | 69.6 | 36 | 100 | 145 |
+| RT-DETR-R50 | 6x |  640 | COCO | 53.1 | 71.3 | 42 | 136 | 108 |
+| RT-DETR-R101 | 6x | 640 | COCO | 54.3 | 72.7 | 76 | 259 | 74 |
+| RT-DETR-HGNetv2-L | 6x | 640 | COCO | 53.0 | 71.6 | 32 | 110 | 114 |
+| RT-DETR-HGNetv2-X | 6x | 640 | COCO | 54.8 | 73.1 | 67 | 234 | 74 |
+| RT-DETR-R18 | 5x | 640 | COCO + Objects365 | **49.2** | **66.6** | 20 | 60 | **217** |
+| RT-DETR-R50 | 2x | 640 | COCO + Objects365 | **55.3** | **73.4** | 42 | 136 | **108** |
+| RT-DETR-R101 | 2x | 640 | COCO + Objects365 | **56.2** | **74.6** | 76 | 259 | **74** |
+
+**Notes:**
+- `COCO + Objects365` in the table means finetuned model on COCO using pretrained weights trained on Objects365.
+
+
+## 💡 Introduction
+We propose a **R**eal-**T**ime **DE**tection **TR**ansformer (RT-DETR, aka RTDETR), the first real-time end-to-end object detector to our best knowledge. Our RT-DETR-R50 / R101 achieves 53.1% / 54.3% AP on COCO and 108 / 74 FPS on T4 GPU, outperforming previously advanced YOLOs in both speed and accuracy. Furthermore, RT-DETR-R50 outperforms DINO-R50 by 2.2% AP in accuracy and about 21 times in FPS. After pre-training with Objects365, RT-DETR-R50 / R101 achieves 55.3% / 56.2% AP.
+<div align="center">
+  <img src="https://github.com/lyuwenyu/RT-DETR/assets/77494834/c211a164-ddce-4084-8b71-fb73f29f363b" width=500 >
+</div>
+
+## 🦄 Performance
+
+### 🏕️ Complex Scenarios
+<div align="center">
+  <img src="https://github.com/lyuwenyu/RT-DETR/assets/77494834/52743892-68c8-4e53-b782-9f89221739e4" width=500 >
+</div>
+
+### 🌋 Difficult Conditions
+<div align="center">
+  <img src="https://github.com/lyuwenyu/RT-DETR/assets/77494834/213cf795-6da6-4261-8549-11947292d3cb" width=500 >
+</div>
+
+## Citation
+If you use `RT-DETR` in your work, please use the following BibTeX entries:
+```
+@misc{lv2023detrs,
+      title={DETRs Beat YOLOs on Real-time Object Detection},
+      author={Yian Zhao and Wenyu Lv and Shangliang Xu and Jinman Wei and Guanzhong Wang and Qingqing Dang and Yi Liu and Jie Chen},
+      year={2023},
+      eprint={2304.08069},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+```

benchmark/README.md

+# 论文测速使用的部分代码和工具
+
+
+## 测试YOLO系列的速度 [in progress]
+以[yolov8](https://github.com/ultralytics/ultralytics)为例
+
+<details open>
+<summary>1. 转onnx </summary>  
+
+执行`yolov8_onnx.py`中的`export_onnx`函数，新增代码主要涉及输出格式的转换
+</details>
+
+
+<details>
+<summary>2. 插入nms </summary>
+
+使用`utils.py`中的`yolo_insert_nms`函数，导出onnx模型后使用[Netron](https://netron.app/)查看结构. <img width="924" alt="image" src="https://github.com/lyuwenyu/RT-DETR/assets/17582080/cb466483-d3a3-4f23-a68d-7ab8825059c8">
+</details>
+
+
+<details>
+<summary>3. 转tensorrt </summary>
+
+可以使用`trtexec.md`中的的脚本转换，或者使用`utils.py`中的Python代码转换
+```bash
+# trtexec -h
+trtexec --onnx=./yolov8l_w_nms.onnx --saveEngine=yolov8l_w_nms.engine --buildOnly --fp16
+```
+</details>
+
+
+<details>
+<summary>4. trtexec测速 </summary>
+
+可以使用`trtexec.md`中的的脚本转换，去掉`--buildOnly`参数
+
+</details>
+
+
+
+<details>
+<summary>5. profile分析（可选） </summary>
+
+在4的基础之上加以下命令
+```bash
+nsys profile --force-overwrite=true  -t 'nvtx,cuda,osrt,cudnn' -c cudaProfilerApi -o yolov8l_w_nms 
+```
+可以使用nsys可视化分析
+<img width="1090" alt="image" src="https://github.com/lyuwenyu/RT-DETR/assets/17582080/507d8bde-9e7c-4ae5-b571-976c540ef2c6">
+
+</details>
+
+
+<details>
+<summary>6. Python测速或者部署   </summary>
+
+在Coco val数据集上测模型的平均速度使用`trtinfer.py`中的代码推理
+
+</details>

benchmark/dataset.py

+'''by lyuwenyu
+'''
+
+import os
+import glob
+from PIL import Image
+
+import torch
+import torch.utils.data as data
+import torchvision
+import torchvision.transforms as T 
+import torchvision.transforms.functional as F 
+
+
+class ToTensor(T.ToTensor):
+    def __init__(self) -> None:
+        super().__init__()
+
+    def __call__(self, pic):
+        if isinstance(pic, torch.Tensor):
+            return pic 
+        return super().__call__(pic)
+
+class PadToSize(T.Pad):
+    def __init__(self, size, fill=0, padding_mode='constant'):
+        super().__init__(0, fill, padding_mode)
+        self.size = size
+        self.fill = fill
+
+    def __call__(self, img):
+        """
+        Args:
+            img (PIL Image or Tensor): Image to be padded.
+
+        Returns:
+            PIL Image or Tensor: Padded image.
+        """
+        w, h = F.get_image_size(img)
+        padding = (0, 0, self.size[0] - w, self.size[1] - h)
+        return F.pad(img, padding, self.fill, self.padding_mode)
+
+
+class Dataset(data.Dataset):
+    def __init__(self, img_dir: str='', preprocess: T.Compose=None, device='cuda:0') -> None:
+        super().__init__()
+
+        self.device = device
+        self.size = 640
+
+        self.im_path_list = list(glob.glob(os.path.join(img_dir, '*.jpg')))
+
+        if preprocess is None: 
+            self.preprocess = T.Compose([
+                    T.Resize(size=639, max_size=640),
+                    PadToSize(size=(640, 640), fill=114),
+                    ToTensor(),
+                    T.ConvertImageDtype(torch.float),
+            ])
+        else:
+            self.preprocess = preprocess
+
+    def __len__(self, ):
+        return len(self.im_path_list)
+
+    def __getitem__(self, index):
+        # im = Image.open(self.img_path_list[index]).convert('RGB')
+        im = torchvision.io.read_file(self.im_path_list[index])
+        im = torchvision.io.decode_jpeg(im, mode=torchvision.io.ImageReadMode.RGB, device=self.device)
+        _, h, w = im.shape # c,h,w
+
+        im = self.preprocess(im)
+
+        blob = {
+            'image': im, 
+            'im_shape': torch.tensor([self.size, self.size]).to(im.device),
+            'scale_factor': torch.tensor([self.size / h, self.size / w]).to(im.device),
+            'orig_size': torch.tensor([w, h]).to(im.device),
+        }
+
+        return blob
+
+    @staticmethod
+    def post_process():
+        pass
+
+    @staticmethod
+    def collate_fn():
+        pass
+
+
+def draw_nms_result(blob, outputs, draw_score_threshold=0.25, name=''):
+    '''show result
+    Keys:
+        'num_dets', 'det_boxes', 'det_scores', 'det_classes'
+    '''    
+    for i in range(blob['image'].shape[0]):
+        det_scores = outputs['det_scores'][i]
+        det_boxes = outputs['det_boxes'][i][det_scores > draw_score_threshold]
+        
+        im = (blob['image'][i] * 255).to(torch.uint8)
+        im = torchvision.utils.draw_bounding_boxes(im, boxes=det_boxes, width=2)
+        Image.fromarray(im.permute(1, 2, 0).cpu().numpy()).save(f'test_{name}_{i}.jpg')

benchmark/trtexec.md

+
+```bash
+# build tensorrt engine 
+trtexec --onnx=./yolov8l_w_nms.onnx --saveEngine=yolov8l_w_nms.engine --buildOnly --fp16
+
+# using dynamic shapes
+# --explicitBatch --minShapes=image:1x3x640x640 --optShapes=image:8x3x640x640  --maxShapes=image:16x3x640x640 --shapes=image:8x3x640x640
+
+# timeline 
+nsys profile --force-overwrite=true  -t 'nvtx,cuda,osrt,cudnn' -c cudaProfilerApi -o yolov8l_w_nms  trtexec --loadEngine=./yolov8l_w_nms.engine --fp16 --avgRuns=10 --loadInputs='image:input_tensor.bin'
+
+# https://forums.developer.nvidia.com/t/about-loadinputs-in-trtexec/218880
+```

benchmark/trtinfer.py

+'''by lyuwenyu
+'''
+
+import time 
+import contextlib
+from collections import namedtuple, OrderedDict
+
+import torch
+import numpy as np
+import tensorrt as trt
+
+from utils import TimeProfiler
+
+class TRTInference(object):
+    def __init__(self, engine_path, device='cuda:0', backend='torch', max_batch_size=32, verbose=False):
+        self.engine_path = engine_path
+        self.device = device
+        self.backend = backend
+        self.max_batch_size = max_batch_size
+        
+        self.logger = trt.Logger(trt.Logger.VERBOSE) if verbose else trt.Logger(trt.Logger.INFO)  
+
+        self.engine = self.load_engine(engine_path)
+
+        self.context = self.engine.create_execution_context()
+
+        self.bindings = self.get_bindings(self.engine, self.context, self.max_batch_size, self.device)
+        self.bindings_addr = OrderedDict((n, v.ptr) for n, v in self.bindings.items())
+
+        self.input_names = self.get_input_names()
+        self.output_names = self.get_output_names()
+        
+        if self.backend == 'cuda':
+            self.stream = cuda.Stream()
+
+        self.time_profile = TimeProfiler()
+
+    def init(self, ):
+        self.dynamic = False 
+
+    def load_engine(self, path):
+        '''load engine
+        '''
+        trt.init_libnvinfer_plugins(self.logger, '')
+        with open(path, 'rb') as f, trt.Runtime(self.logger) as runtime:
+            return runtime.deserialize_cuda_engine(f.read())
+    
+    def get_input_names(self, ):
+        names = []
+        for _, name in enumerate(self.engine):
+            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
+                names.append(name)
+        return names
+    
+    def get_output_names(self, ):
+        names = []
+        for _, name in enumerate(self.engine):
+            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.OUTPUT:
+                names.append(name)
+        return names
+
+    def get_bindings(self, engine, context, max_batch_size=32, device=None):
+        '''build binddings
+        '''
+        Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr'))
+        bindings = OrderedDict()
+        # max_batch_size = 1
+
+        for i, name in enumerate(engine):
+            shape = engine.get_tensor_shape(name)
+            dtype = trt.nptype(engine.get_tensor_dtype(name))
+
+            if shape[0] == -1:
+                dynamic = True 
+                shape[0] = max_batch_size
+                if engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:  # dynamic
+                    context.set_input_shape(name, shape)
+
+            if self.backend == 'cuda':
+                if engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
+                    data = np.random.randn(*shape).astype(dtype)
+                    ptr = cuda.mem_alloc(data.nbytes)
+                    bindings[name] = Binding(name, dtype, shape, data, ptr) 
+                else:
+                    data = cuda.pagelocked_empty(trt.volume(shape), dtype)
+                    ptr = cuda.mem_alloc(data.nbytes)
+                    bindings[name] = Binding(name, dtype, shape, data, ptr) 
+
+            else:
+                data = torch.from_numpy(np.empty(shape, dtype=dtype)).to(device)
+                bindings[name] = Binding(name, dtype, shape, data, data.data_ptr())
+
+        return bindings
+
+    def run_torch(self, blob):
+        '''torch input
+        '''
+        for n in self.input_names:
+            if self.bindings[n].shape != blob[n].shape:
+                self.context.set_input_shape(n, blob[n].shape) 
+                self.bindings[n] = self.bindings[n]._replace(shape=blob[n].shape)
+
+        self.bindings_addr.update({n: blob[n].data_ptr() for n in self.input_names})
+        self.context.execute_v2(list(self.bindings_addr.values()))
+        outputs = {n: self.bindings[n].data for n in self.output_names}
+
+        return outputs
+
+
+    def async_run_cuda(self, blob):
+        '''numpy input
+        '''
+        for n in self.input_names:
+            cuda.memcpy_htod_async(self.bindings_addr[n], blob[n], self.stream)
+        
+        bindings_addr = [int(v) for _, v in self.bindings_addr.items()]
+        self.context.execute_async_v2(bindings=bindings_addr, stream_handle=self.stream.handle)
+        
+        outputs = {}
+        for n in self.output_names:
+            cuda.memcpy_dtoh_async(self.bindings[n].data, self.bindings[n].ptr, self.stream)
+            outputs[n] = self.bindings[n].data
+        
+        self.stream.synchronize()
+        
+        return outputs
+    
+    def __call__(self, blob):
+        if self.backend == 'torch':
+            return self.run_torch(blob)
+
+        elif self.backend == 'cuda':
+            return self.async_run_cuda(blob)
+
+    def synchronize(self, ):
+        if self.backend == 'torch' and torch.cuda.is_available():
+            torch.cuda.synchronize()
+
+        elif self.backend == 'cuda':
+            self.stream.synchronize()
+    
+    def warmup(self, blob, n):
+        for _ in range(n):
+            _ = self(blob)
+
+    def speed(self, blob, n):
+        self.time_profile.reset()
+        for _ in range(n):
+            with self.time_profile:
+                _ = self(blob)
+
+        return self.time_profile.total / n 
+

benchmark/utils.py

+'''by lyuwenyu
+'''
+
+import time 
+import contextlib
+import numpy as np
+from PIL import Image
+from collections import OrderedDict
+
+import onnx
+import torch 
+import onnx_graphsurgeon
+
+
+def to_binary_data(path, size=(640, 640), output_name='input_tensor.bin'):
+    '''--loadInputs='image:input_tensor.bin'
+    '''
+    im = Image.open(path).resize(size)
+    data = np.asarray(im, dtype=np.float32).transpose(2, 0, 1)[None] / 255.
+    data.tofile(output_name)
+
+
+def yolo_insert_nms(path, score_threshold=0.01, iou_threshold=0.7, max_output_boxes=300, simplify=False):
+    '''
+    http://www.xavierdupre.fr/app/onnxcustom/helpsphinx/api/onnxops/onnx__EfficientNMS_TRT.html
+    https://huggingface.co/spaces/muttalib1326/Punjabi_Character_Detection/blob/3dd1e17054c64e5f6b2254278f96cfa2bf418cd4/utils/add_nms.py
+    '''
+    onnx_model = onnx.load(path)
+
+    if simplify:
+        from onnxsim import simplify
+        onnx_model, _ = simplify(onnx_model,  overwrite_input_shapes={'image': [1, 3, 640, 640]})
+
+    graph = onnx_graphsurgeon.import_onnx(onnx_model)
+    graph.toposort()
+    graph.fold_constants()
+    graph.cleanup()
+
+    topk = max_output_boxes
+    attrs = OrderedDict(plugin_version='1',
+                        background_class=-1,
+                        max_output_boxes=topk,
+                        score_threshold=score_threshold,
+                        iou_threshold=iou_threshold,
+                        score_activation=False,
+                        box_coding=0, )
+
+    outputs = [onnx_graphsurgeon.Variable('num_dets', np.int32, [-1, 1]),
+               onnx_graphsurgeon.Variable('det_boxes', np.float32, [-1, topk, 4]),
+               onnx_graphsurgeon.Variable('det_scores', np.float32, [-1, topk]),
+               onnx_graphsurgeon.Variable('det_classes', np.int32, [-1, topk])]
+
+    graph.layer(op='EfficientNMS_TRT', 
+                name="batched_nms", 
+                inputs=[graph.outputs[0], 
+                        graph.outputs[1]], 
+                outputs=outputs, 
+                attrs=attrs, )
+
+    graph.outputs = outputs
+    graph.cleanup().toposort()
+
+    onnx.save(onnx_graphsurgeon.export_onnx(graph), f'yolo_w_nms.onnx')
+
+
+class TimeProfiler(contextlib.ContextDecorator):
+    def __init__(self, ):
+        self.total = 0
+        
+    def __enter__(self, ):
+        self.start = self.time()
+        return self 
+    
+    def __exit__(self, type, value, traceback):
+        self.total += self.time() - self.start
+    
+    def reset(self, ):
+        self.total = 0
+    
+    def time(self, ):
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()
+        return time.time()

benchmark/yolov8_onnx.py

+'''by lyuwenyu
+'''
+
+import torch 
+import torchvision
+
+import numpy as np 
+import onnxruntime as ort 
+
+from utils import yolo_insert_nms
+
+class YOLOv8(torch.nn.Module):
+    def __init__(self, name) -> None:
+        super().__init__()
+        from ultralytics import YOLO
+        # Load a model
+        # build a new model from scratch
+        # model = YOLO(f'{name}.yaml')  
+
+        # load a pretrained model (recommended for training)
+        model = YOLO(f'{name}.pt')  
+        self.model = model.model
+
+    def forward(self, x):
+        '''https://github.com/ultralytics/ultralytics/blob/main/ultralytics/nn/tasks.py#L216
+        '''
+        pred: torch.Tensor = self.model(x)[0] # n 84 8400,
+        pred = pred.permute(0, 2, 1)
+        boxes, scores = pred.split([4, 80], dim=-1)
+        boxes = torchvision.ops.box_convert(boxes, in_fmt='cxcywh', out_fmt='xyxy')
+
+        return boxes, scores
+
+
+
+def export_onnx(name='yolov8n'):
+    '''export onnx
+    '''
+    m = YOLOv8(name)
+
+    x = torch.rand(1, 3, 640, 640)
+    dynamic_axes = {
+        'image': {0: '-1'}
+    }
+    torch.onnx.export(m, x, f'{name}.onnx', 
+                      input_names=['image'], 
+                      output_names=['boxes', 'scores'], 
+                      opset_version=13, 
+                      dynamic_axes=dynamic_axes)
+
+    data = np.random.rand(1, 3, 640, 640).astype(np.float32)
+    sess = ort.InferenceSession(f'{name}.onnx')
+    _ = sess.run(output_names=None, input_feed={'image': data})
+
+
+if __name__ == '__main__':
+
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--name', type=str, default='yolov8l')
+    parser.add_argument('--score_threshold', type=float, default=0.001)
+    parser.add_argument('--iou_threshold', type=float, default=0.7)
+    parser.add_argument('--max_output_boxes', type=int, default=300)
+    args = parser.parse_args()
+
+    export_onnx(name=args.name)
+    
+    yolo_insert_nms(path=f'{args.name}.onnx', 
+                    score_threshold=args.score_threshold, 
+                    iou_threshold=args.iou_threshold, 
+                    max_output_boxes=args.max_output_boxes, )
+

datasets/url.txt

+http://images.cocodataset.org/zips/val2017.zip
+http://images.cocodataset.org/zips/train2017.zip
+http://images.cocodataset.org/annotations/annotations_trainval2017.zip
\ No newline at end of file

model.properties

+# 模型唯一标识
+modelCode = 1123
+# 模型名称
+modelName=rt-detr_pytorch
+# 模型描述
+modelDescription=RT-DETR是百度开发的一种端到端的实时物体检测器，可在保持高精度的同时提供实时性能
+# 应用场景
+appScenario=推理,训练,目标检测,制造,交通,电商,广媒,医疗
+# 框架类型
+frameType=pytorch
\ No newline at end of file

model/url.txt

+https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r18vd_dec3_6x_coco_from_paddle.pth
+https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r34vd_dec4_6x_coco_from_paddle.pth
+https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r50vd_m_6x_coco_from_paddle.pth
+https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r50vd_6x_coco_from_paddle.pth
+https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r101vd_6x_coco_from_paddle.pth
+# COCO + Objects365表示使用在 Objects365 上训练的预训练权重在 COCO 上微调的模型。
+https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r18vd_5x_coco_objects365_from_paddle.pth
+https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r50vd_2x_coco_objects365_from_paddle.pth
+https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r101vd_2x_coco_objects365_from_paddle.pth
\ No newline at end of file

onnx_infer.py

+import onnxruntime as ort 
+from PIL import Image, ImageDraw
+from torchvision.transforms import ToTensor
+import torch
+import time
+# print(onnx.helper.printable_graph(mm.graph))
+
+im = Image.open('datasets/000000033109.jpg').convert('RGB') # TODO:修改推理图片路径
+im = im.resize((640, 640))
+im_data = ToTensor()(im)[None]
+print(im_data.shape)
+size = torch.tensor([[640, 640]])
+sess = ort.InferenceSession("model/onnx/rtdetr_r101vd_6x_coco.onnx")  # TODO:修改onnx模型路径
+start=time.time()
+output = sess.run(
+    # output_names=['labels', 'boxes', 'scores'],
+    output_names=None,
+    input_feed={'images': im_data.data.numpy(), "orig_target_sizes": size.data.numpy()}
+)
+end=time.time()
+print(end-start,1.0/(end-start))
+
+# print(type(output))
+# print([out.shape for out in output])
+
+labels, boxes, scores = output
+print("labels shape = ",labels)
+print("boxes shape = ",boxes)
+print("scores shape = ",scores)
+draw = ImageDraw.Draw(im)
+thrh = 0.6
+
+for i in range(im_data.shape[0]):
+
+    scr = scores[i]
+    lab = labels[i][scr > thrh]
+    box = boxes[i][scr > thrh]
+
+    print(i, sum(scr > thrh))
+
+    for b in box:
+        draw.rectangle(list(b), outline='red',)
+        draw.text((b[0], b[1]), text=str(lab[i]), fill='blue', )
+
+im.save('result/test.jpg')  # TODO:修改推理结果存储路径
\ No newline at end of file

requirements.txt

+# torch==2.0.1
+# torchvision==0.15.2
+onnx==1.14.0
+onnxruntime==1.15.1
+pycocotools
+PyYAML
+scipy

rtdetr_pytorch/README.md

+## TODO
+<details>
+<summary> see details </summary>
+
+- [x] Training
+- [x] Evaluation
+- [x] Export onnx
+- [x] Upload source code
+- [x] Upload weight convert from paddle, see [*links*](https://github.com/lyuwenyu/RT-DETR/issues/42)
+- [x] Align training details with the [*paddle version*](../rtdetr_paddle/)
+- [x] Tuning rtdetr based on [*pretrained weights*](https://github.com/lyuwenyu/RT-DETR/issues/42)
+
+</details>
+
+
+## Model Zoo
+
+| Model | Dataset | Input Size | AP<sup>val</sup> | AP<sub>50</sub><sup>val</sup> | #Params(M) | FPS |  checkpoint |
+| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+rtdetr_r18vd | COCO | 640 | 46.4 | 63.7 | 20 | 217 | [url<sup>*</sup>](https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r18vd_dec3_6x_coco_from_paddle.pth)
+rtdetr_r34vd | COCO | 640 | 48.9 | 66.8 | 31 | 161 | [url<sup>*</sup>](https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r34vd_dec4_6x_coco_from_paddle.pth)
+rtdetr_r50vd_m | COCO | 640 | 51.3 | 69.5 | 36 | 145 | [url<sup>*</sup>](https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r50vd_m_6x_coco_from_paddle.pth)
+rtdetr_r50vd | COCO | 640 | 53.1 | 71.2| 42 | 108 | [url<sup>*</sup>](https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r50vd_6x_coco_from_paddle.pth)
+rtdetr_r101vd | COCO | 640 | 54.3 | 72.8 | 76 | 74 | [url<sup>*</sup>](https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r101vd_6x_coco_from_paddle.pth)
+rtdetr_18vd | COCO+Objects365 | 640 | 49.0 | 66.5 | 20 | 217 | [url<sup>*</sup>](https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r18vd_5x_coco_objects365_from_paddle.pth)
+rtdetr_r50vd | COCO+Objects365 | 640 | 55.2 | 73.4 | 42 | 108 | [url<sup>*</sup>](https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r50vd_2x_coco_objects365_from_paddle.pth)
+rtdetr_r101vd | COCO+Objects365 | 640 | 56.2 | 74.5 | 76 | 74 | [url<sup>*</sup>](https://github.com/lyuwenyu/storage/releases/download/v0.1/rtdetr_r101vd_2x_coco_objects365_from_paddle.pth)
+
+Notes
+- `COCO + Objects365` in the table means finetuned model on `COCO` using pretrained weights trained on `Objects365`.
+- `url`<sup>`*`</sup> is the url of pretrained weights convert from paddle model for save energy. *It may have slight differences between this table and paper*
+<!-- - `FPS` is evaluated on a single T4 GPU with $batch\\_size = 1$ and $tensorrt\\_fp16$ mode -->
+
+## Quick start
+
+<details>
+<summary>Install</summary>
+
+```bash
+pip install -r requirements.txt
+```
+
+</details>
+
+
+<details>
+<summary>Data</summary>
+
+- Download and extract COCO 2017 train and val images.
+```
+path/to/coco/
+  annotations/  # annotation json files
+  train2017/    # train images
+  val2017/      # val images
+```
+- Modify config [`img_folder`, `ann_file`](configs/dataset/coco_detection.yml)
+</details>
+
+
+
+<details>
+<summary>Training & Evaluation</summary>
+
+- Training on a Single GPU:
+
+```shell
+# training on single-gpu
+export CUDA_VISIBLE_DEVICES=0
+python tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml
+```
+
+- Training on Multiple GPUs:
+
+```shell
+# train on multi-gpu
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+torchrun --nproc_per_node=4 tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml
+```
+
+- Evaluation on Multiple GPUs:
+
+```shell
+# val on multi-gpu
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+torchrun --nproc_per_node=4 tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml -r path/to/checkpoint --test-only
+```
+
+</details>
+
+
+
+<details>
+<summary>Export</summary>
+
+```shell
+python tools/export_onnx.py -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml -r path/to/checkpoint --check
+```
+</details>
+
+
+
+
+<details open>
+<summary>Train custom data</summary>
+
+1. set `remap_mscoco_category: False`. This variable only works for ms-coco dataset. If you want to use `remap_mscoco_category` logic on your dataset, please modify variable [`mscoco_category2name`](https://github.com/lyuwenyu/RT-DETR/blob/main/rtdetr_pytorch/src/data/coco/coco_dataset.py#L154) based on your dataset.
+
+2. add `-t path/to/checkpoint` (optinal) to tuning rtdetr based on pretrained checkpoint. see [training script details](./tools/README.md).
+</details>