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 MegEngine version YOLOX, there is also a PyTorch implementation.
- 【2021/08/05】 We release MegEngine version YOLOX.
- Faster YOLOX training speed.
- More models of megEngine version.
- AMP training of megEngine.
| Model | size | mAPval 0.5:0.95 |
Params (M) |
FLOPs (G) |
weights |
|---|---|---|---|---|---|
| YOLOX-Tiny | 416 | 32.2 | 5.06 | 6.45 | github |
Comming soon!
Installation
Step1. Install YOLOX.
git clone [email protected]:MegEngine/YOLOX.git
cd YOLOX
pip3 install -U pip && pip3 install -r requirements.txt
pip3 install -v -e . # or python3 setup.py developStep2. Install pycocotools.
pip3 install cython; pip3 install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'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-tiny -c /path/to/your/yolox_tiny.pkl --path assets/dog.jpg --conf 0.25 --nms 0.45 --tsize 416 --save_result --device [cpu/gpu]or
python tools/demo.py image -f exps/default/yolox_tiny.py -c /path/to/your/yolox_tiny.pkl --path assets/dog.jpg --conf 0.25 --nms 0.45 --tsize 416 --save_result --device [cpu/gpu]Demo for video:
python tools/demo.py video -n yolox-s -c /path/to/your/yolox_s.pkl --path /path/to/your/video --conf 0.25 --nms 0.45 --tsize 416 --save_result --device [cpu/gpu]Reproduce our results on COCO
Step1. Prepare COCO dataset
cd <YOLOX_HOME>
ln -s /path/to/your/COCO ./datasets/COCOStep2. Reproduce our results on COCO by specifying -n:
python tools/train.py -n yolox-tiny -d 8 -b 128- -d: number of gpu devices
- -b: total batch size, the recommended number for -b is num-gpu * 8
When using -f, the above commands are equivalent to:
python tools/train.py -f exps/default/yolox-tiny.py -d 8 -b 128Evaluation
We support batch testing for fast evaluation:
python tools/eval.py -n yolox-tiny -c yolox_tiny.pkl -b 64 -d 8 --conf 0.001 [--fuse]- --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 tools/eval.py -n yolox-tiny -c yolox_tiny.pkl -b 1 -d 1 --conf 0.001 --fuseTutorials
Dump mge file
NOTE: result model is dumped with optimize_for_inference and enable_fuse_conv_bias_nonlinearity.
python3 tools/export_mge.py -n yolox-tiny -c yolox_tiny.pkl --dump_path yolox_tiny.mge-
Model Info: yolox-s @ input(1,3,640,640)
-
Testing Devices
x86_64 -- Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHzAArch64 -- Xiaomi phone mi9CUDA -- 1080TI @ cuda-10.1-cudnn-v7.6.3-TensorRT-6.0.1.5.sh @ Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz
| [email protected] +fastrun +weight_preprocess (msec) | 1 thread | 2 thread | 4 thread | 8 thread |
|---|---|---|---|---|
| x86_64(fp32) | 516.245 | 318.29 | 253.273 | 222.534 |
| x86_64(fp32+chw88) | 362.020 | NONE | NONE | NONE |
| aarch64(fp32+chw44) | 555.877 | 351.371 | 242.044 | NONE |
| aarch64(fp16+chw) | 439.606 | 327.356 | 255.531 | NONE |
| CUDA @ CUDA (msec) | 1 batch | 2 batch | 4 batch | 8 batch | 16 batch | 32 batch | 64 batch |
|---|---|---|---|---|---|---|---|
| megengine(fp32+chw) | 8.137 | 13.2893 | 23.6633 | 44.470 | 86.491 | 168.95 | 334.248 |
- 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 ONNXRuntime C++ Demo: lite.ai from DefTruth
- Converting darknet or yolov5 datasets to COCO format for YOLOX: YOLO2COCO from Daniel
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}
}