HorNet

Created by Yongming Rao*, Wenliang Zhao*, Yansong Tang, Jie Zhou, Ser-Nam Lim†, Jiwen Lu†

This repository contains PyTorch implementation for HorNet.

HorNet is a family of generic vision backbones that perform explicit high-order spatial interactions based on Recursive Gated Convolution.

[Project Page] [arXiv]

Model Zoo

ImageNet-1K trained models:

name	arch	Params	FLOPs	Top-1	url
HorNet-T (7x7)	`hornet_tiny_7x7`	22M	4.0G	82.8	Tsinghua Cloud
HorNet-T (GF)	`hornet_tiny_gf`	23M	3.9G	83.0	Tsinghua Cloud
HorNet-S (7x7)	`hornet_small_7x7`	50M	8.8G	83.8	Tsinghua Cloud
HorNet-S (GF)	`hornet_small_gf`	50M	8.7G	84.0	Tsinghua Cloud
HorNet-B (7x7)	`hornet_base_7x7`	87M	15.6G	84.2	Tsinghua Cloud
HorNet-B (GF)	`hornet_base_gf`	88M	15.5G	84.3	Tsinghua Cloud

ImageNet-22K trained models:

name	arch	Params	FLOPs	url
HorNet-L (7x7)	`hornet_large_7x7`	209M	34.8G	Tsinghua Cloud
HorNet-L (GF)	`hornet_large_gf`	211M	34.7G	Tsinghua Cloud
HorNet-L (GF)*	`hornet_large_gf_img384`	216M	101.8G	Tsinghua Cloud

*indicate the model is finetuned to 384x384 resolution on ImageNet-22k.

ImageNet Classification

Requirements

torch==1.8.0
torchvision==0.9.0
timm==0.4.12
tensorboardX
six
submitit (multi-node training)

Data preparation: download and extract ImageNet images from http://image-net.org/. The directory structure should be

│ILSVRC2012/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

Evaluation

To evaluate a pre-trained HorNet model on the ImageNet validation set with 8 GPUs, run:

python -m torch.distributed.launch --nproc_per_node=8 main.py \
--model hornet_tiny_7x7 --eval true --input_size 224 \
--resume /path/to/checkpoint \ 
--data_path /path/to/imagenet-1k

Training

To train HorNet models on ImageNet from scratch on a single machine, run:

python -m torch.distributed.launch --nproc_per_node=8 main.py \
--model hornet_tiny_7x7 --drop_path 0.2 --clip_grad 5\
--batch_size 128 --lr 4e-3 --update_freq 4 \
--model_ema true --model_ema_eval true \
--data_path /path/to/imagenet-1k \
--output_dir ./logs/hornet_tiny_7x7

To train HorNet models on ImageNet from scratch with 4 8-GPU nodes, run:

python run_with_submitit.py --nodes 4 --ngpus 8 \
--model hornet_tiny_7x7 --drop_path 0.2 --clip_grad 5\
--batch_size 128 --lr 4e-3 --update_freq 1 \
--model_ema true --model_ema_eval true \
--data_path /path/to/imagenet-1k \
--job_dir ./logs/hornet_tiny_7x7

To finetune a pre-trained model to ImageNet-1K at higher resolution, run:

python run_with_submitit.py --nodes 4 --ngpus 8 \
--model hornet_large_gf_img384 --drop_path 0.4 --clip_grad 1 \
--batch_size 16 --lr 5e-5 --update_freq 1 --weight_decay 1e-8 \
--warmup_epochs 0 --input_size 384 --epochs 30 \
--model_ema true --model_ema_eval true --cutmix 0 --mixup 0 \
--layer_decay 0.7 --finetune ./pretrained/hornet_large_gf_in22k.pth \
--data_path /path/to/imagenet-1k \
--job_dir ./logs/hornet_larget_gf_img384_ft1k

We provide detailed training commands for our models in TRAINING.md.

Downstream Tasks

Please check the object_detection.md and semantic_segmentation.md for training and evaluation instructions on downstream tasks.

License

MIT License

Acknowledgements

Our code is based on pytorch-image-models, DeiT and ConvNeXt. We would like to thank High-Flyer AI Research for their generous support of a part of computational resources used in this project.

Citation

If you find our work useful in your research, please consider citing:

@article{rao2022hornet,
  title={HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions},
  author={Rao, Yongming and Zhao, Wenliang and Tang, Yansong and Zhou, Jie and Lim, Ser-Lam and Lu, Jiwen},
  journal={arXiv preprint arXiv:2207.14284},
  year={2022}
}

sunxingxingtf / HorNet