Introduction

YOLOX is an anchor-free version of YOLO, with a simpler design but better performance! It aims to bridge the gap between research and industrial communities. For more details, please refer to our report on Arxiv.

This repo is an implementation of PyTorch version YOLOX, there is also a MegEngine implementation.

Updates!!

• 【2023/02/28】 We support assignment visualization tool, see doc here.
• 【2022/04/14】 We support jit compile op.
• 【2021/08/19】 We optimize the training process with 2x faster training and ~1% higher performance! See notes for more details.
• 【2021/08/05】 We release MegEngine version YOLOX.
• 【2021/07/28】 We fix the fatal error of memory leak
• 【2021/07/26】 We now support MegEngine deployment.
• 【2021/07/20】 We have released our technical report on Arxiv.

Benchmark

Standard Models.

Model	size	mAPval 0.5:0.95	mAPtest 0.5:0.95	Speed V100 (ms)	Params (M)	FLOPs (G)	weights
YOLOX-s	640	40.5	40.5	9.8	9.0	26.8	github
YOLOX-m	640	46.9	47.2	12.3	25.3	73.8	github
YOLOX-l	640	49.7	50.1	14.5	54.2	155.6	github
YOLOX-x	640	51.1	51.5	17.3	99.1	281.9	github
YOLOX-Darknet53	640	47.7	48.0	11.1	63.7	185.3	github

Legacy models

Model	size	mAPtest 0.5:0.95	Speed V100 (ms)	Params (M)	FLOPs (G)	weights
YOLOX-s	640	39.6	9.8	9.0	26.8	onedrive/github
YOLOX-m	640	46.4	12.3	25.3	73.8	onedrive/github
YOLOX-l	640	50.0	14.5	54.2	155.6	onedrive/github
YOLOX-x	640	51.2	17.3	99.1	281.9	onedrive/github
YOLOX-Darknet53	640	47.4	11.1	63.7	185.3	onedrive/github

Light Models.

Model	size	mAPval 0.5:0.95	Params (M)	FLOPs (G)	weights
YOLOX-Nano	416	25.8	0.91	1.08	github
YOLOX-Tiny	416	32.8	5.06	6.45	github

Legacy models

Model	size	mAPval 0.5:0.95	Params (M)	FLOPs (G)	weights
YOLOX-Nano	416	25.3	0.91	1.08	github
YOLOX-Tiny	416	32.8	5.06	6.45	github

Quick Start

Installation

Step1. Install YOLOX from source.

git clone git@github.com:Megvii-BaseDetection/YOLOX.git
cd YOLOX
pip3 install -v -e .  # or  python3 setup.py develop

Demo

Step1. Download a pretrained model from the benchmark table.

Step2. Use either -n or -f to specify your detector's config. For example:

python tools/demo.py image -n yolox-s -c /path/to/your/yolox_s.pth --path assets/dog.jpg --conf 0.25 --nms 0.45 --tsize 640 --save_result --device [cpu/gpu]

or

python tools/demo.py image -f exps/default/yolox_s.py -c /path/to/your/yolox_s.pth --path assets/dog.jpg --conf 0.25 --nms 0.45 --tsize 640 --save_result --device [cpu/gpu]

Demo for video:

python tools/demo.py video -n yolox-s -c /path/to/your/yolox_s.pth --path /path/to/your/video --conf 0.25 --nms 0.45 --tsize 640 --save_result --device [cpu/gpu]

Reproduce our results on COCO

Step1. Prepare COCO dataset

cd <YOLOX_HOME>
ln -s /path/to/your/COCO ./datasets/COCO

Step2. Reproduce our results on COCO by specifying -n:

python -m yolox.tools.train -n yolox-s -d 8 -b 64 --fp16 -o [--cache]
                               yolox-m
                               yolox-l
                               yolox-x

• -d: number of gpu devices
• -b: total batch size, the recommended number for -b is num-gpu * 8
• --fp16: mixed precision training
• --cache: caching imgs into RAM to accelarate training, which need large system RAM.

When using -f, the above commands are equivalent to:

python -m yolox.tools.train -f exps/default/yolox_s.py -d 8 -b 64 --fp16 -o [--cache]
                               exps/default/yolox_m.py
                               exps/default/yolox_l.py
                               exps/default/yolox_x.py

Multi Machine Training

We also support multi-nodes training. Just add the following args:

• --num_machines: num of your total training nodes
• --machine_rank: specify the rank of each node

Suppose you want to train YOLOX on 2 machines, and your master machines's IP is 123.123.123.123, use port 12312 and TCP.

On master machine, run

python tools/train.py -n yolox-s -b 128 --dist-url tcp://123.123.123.123:12312 --num_machines 2 --machine_rank 0

On the second machine, run

python tools/train.py -n yolox-s -b 128 --dist-url tcp://123.123.123.123:12312 --num_machines 2 --machine_rank 1

Logging to Weights & Biases

To log metrics, predictions and model checkpoints to W&B use the command line argument --logger wandb and use the prefix "wandb-" to specify arguments for initializing the wandb run.

python tools/train.py -n yolox-s -d 8 -b 64 --fp16 -o [--cache] --logger wandb wandb-project <project name>
                         yolox-m
                         yolox-l
                         yolox-x

An example wandb dashboard is available here

Others

See more information with the following command:

python -m yolox.tools.train --help

Evaluation

We support batch testing for fast evaluation:

python -m yolox.tools.eval -n  yolox-s -c yolox_s.pth -b 64 -d 8 --conf 0.001 [--fp16] [--fuse]
                               yolox-m
                               yolox-l
                               yolox-x

• --fuse: fuse conv and bn
• -d: number of GPUs used for evaluation. DEFAULT: All GPUs available will be used.
• -b: total batch size across on all GPUs

To reproduce speed test, we use the following command:

python -m yolox.tools.eval -n  yolox-s -c yolox_s.pth -b 1 -d 1 --conf 0.001 --fp16 --fuse
                               yolox-m
                               yolox-l
                               yolox-x

Tutorials

• Training on custom data
• Caching for custom data
• Manipulating training image size
• Assignment visualization
• Freezing model

Deployment

1. MegEngine in C++ and Python
2. ONNX export and an ONNXRuntime
3. TensorRT in C++ and Python
4. ncnn in C++ and Java
5. OpenVINO in C++ and Python
6. Accelerate YOLOX inference with nebullvm in Python

Third-party resources

• YOLOX for streaming perception: StreamYOLO (CVPR 2022 Oral)
• The YOLOX-s and YOLOX-nano are Integrated into ModelScope. Try out the Online Demo at YOLOX-s and YOLOX-Nano respectively 🚀.
• Integrated into Huggingface Spaces 🤗 using Gradio. Try out the Web Demo:
• The ncnn android app with video support: ncnn-android-yolox from FeiGeChuanShu
• YOLOX with Tengine support: Tengine from BUG1989
• YOLOX + ROS2 Foxy: YOLOX-ROS from Ar-Ray
• YOLOX Deploy DeepStream: YOLOX-deepstream from nanmi
• YOLOX MNN/TNN/ONNXRuntime: YOLOX-MNN、YOLOX-TNN and YOLOX-ONNXRuntime C++ from DefTruth
• Converting darknet or yolov5 datasets to COCO format for YOLOX: YOLO2COCO from Daniel

Cite YOLOX

If you use YOLOX in your research, please cite our work by using the following BibTeX entry:

 @article{yolox2021,
  title={YOLOX: Exceeding YOLO Series in 2021},
  author={Ge, Zheng and Liu, Songtao and Wang, Feng and Li, Zeming and Sun, Jian},
  journal={arXiv preprint arXiv:2107.08430},
  year={2021}
}

In memory of Dr. Jian Sun

Without the guidance of Dr. Jian Sun, YOLOX would not have been released and open sourced to the community. The passing away of Dr. Sun is a huge loss to the Computer Vision field. We add this section here to express our remembrance and condolences to our captain Dr. Sun. It is hoped that every AI practitioner in the world will stick to the belief of "continuous innovation to expand cognitive boundaries, and extraordinary technology to achieve product value" and move forward all the way.

没有孙剑博士的指导，YOLOX也不会问世并开源给社区使用。 孙剑博士的离去是CV领域的一大损失，我们在此特别添加了这个部分来表达对我们的“船长”孙老师的纪念和哀思。 希望世界上的每个AI从业者秉持着“持续创新拓展认知边界，非凡科技成就产品价值”的观念，一路向前。