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!!

• 【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.

Comming soon

• YOLOX-P6 and larger model.
• Objects365 pretrain.
• Transformer modules.
• More features in need.

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

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
• Manipulating training image size
• 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

Third-party resources

• 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}
}