jdc08161063 / CycleMLP

Implementation of "CycleMLP: A MLP-like Architecture for Dense Prediction"

CycleMLP: A MLP-like Architecture for Dense Prediction (arXiv)

    

Updates

• (22/07/2021) Initial release.

Model Zoo

We provide CycleMLP models pretrained on ImageNet 2012.

Model	Parameters	FLOPs	Top 1 Acc.	Download
CycleMLP-B1	15M	2.1G	78.9%	model
CycleMLP-B2	27M	3.9G	81.6%	model
CycleMLP-B3	38M	6.9G	82.4%	model
CycleMLP-B4	52M	10.1G	83.0%	model
CycleMLP-B5	76M	12.3G	83.2%	model

Usage

Install

• PyTorch 1.7.0+ and torchvision 0.8.1+
• timm:

pip install 'git+https://github.com/rwightman/pytorch-image-models@c2ba229d995c33aaaf20e00a5686b4dc857044be'

or

git clone https://github.com/rwightman/pytorch-image-models
cd pytorch-image-models
git checkout c2ba229d995c33aaaf20e00a5686b4dc857044be
pip install -e .

• fvcore (optional, for FLOPs calculation)
• mmcv, mmdetection, mmsegmentation (optional)

Data preparation

Download and extract ImageNet train and val images from http://image-net.org/. The directory structure is:

│path/to/imagenet/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

Evaluation

To evaluate a pre-trained CycleMLP-B5 on ImageNet val with a single GPU run:

python main.py --eval --model CycleMLP_B5 --resume path/to/CycleMLP_B5.pth --data-path /path/to/imagenet

Training

To train CycleMLP-B5 on ImageNet on a single node with 8 gpus for 300 epochs run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --model CycleMLP_B5 --batch-size 128 --data-path /path/to/imagenet --output_dir /path/to/save

Acknowledgement

This code is based on DeiT and pytorch-image-models. Thanks for their wonderful works

Citing

@article{chen2021cyclemlp,
  title={CycleMLP: A MLP-like Architecture for Dense Prediction},
  author={Chen, Shoufa and Xie, Enze and Ge, Chongjian and Liang, Ding and Luo, Ping},
  journal={arXiv preprint arXiv:2107.10224},
  year={2021}
}

License

CycleMLP is released under MIT License.