Scaling Local Self-Attention for Parameter Efficient Visual Backbones, arxiv

PaddlePaddle training/validation code and pretrained models for HaloNet.

The official pytorch implementation is N/A.

This implementation is developed by PaddleViT.

HaloNet local self-attention architecture

Update

• Update (2021-12-09): Initial code and ported weights are released.

Models Zoo

Model	Acc@1	Acc@5	#Params	FLOPs	Image Size	Crop_pct	Interpolation	Link
halonet26t	79.10	94.31	12.5M	3.2G	256	0.95	bicubic	google/baidu(ednv)
halonet50ts	81.65	95.61	22.8M	5.1G	256	0.94	bicubic	google/baidu(3j9e)

*The results are evaluated on ImageNet2012 validation set.

Notebooks

We provide a few notebooks in aistudio to help you get started:

*(coming soon)*

Requirements

• Python>=3.6
• yaml>=0.2.5
• PaddlePaddle>=2.1.0
• yacs>=0.1.8

Data

ImageNet2012 dataset is used in the following folder structure:

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

Usage

To use the model with pretrained weights, download the .pdparam weight file and change related file paths in the following python scripts. The model config files are located in ./configs/.

For example, assume the downloaded weight file is stored in ./halonet_50ts_256.pdparams, to use the halonet_50ts_256 model in python:

from config import get_config
from halonet import build_halonet 
# config files in ./configs/
config = get_config('./configs/halonet_50ts_256.yaml')
# build model
model = build_model(config)
# load pretrained weights, .pdparams is NOT needed
model_state_dict = paddle.load('./halonet_50ts_256')
model.set_dict(model_state_dict)

Evaluation

To evaluate HaloNet model performance on ImageNet2012 with a single GPU, run the following script using command line:

sh run_eval.sh

or

CUDA_VISIBLE_DEVICES=0 \
python main_single_gpu.py \
    -cfg='./configs/halonet_50ts_256.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \
    -eval \
    -pretrained='./halonet_50ts_256'

Run evaluation using multi-GPUs:

sh run_eval_multi.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python main_multi_gpu.py \
    -cfg='./configs/halonet_50ts_256.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \
    -eval \
    -pretrained='./halonet_50ts_256'

Training

To train the MobileVit XXS model on ImageNet2012 with single GPU, run the following script using command line:

sh run_train.sh

or

CUDA_VISIBLE_DEVICES=0 \
python main_singel_gpu.py \
  -cfg='./configs/halonet_50ts_256.yaml' \
  -dataset='imagenet2012' \
  -batch_size=32 \
  -data_path='/dataset/imagenet' \

Run training using multi-GPUs:

sh run_train_multi.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python main_multi_gpu.py \
    -cfg='./configs/halonet_50ts_256.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \

Visualization Attention Map

(coming soon)

Reference

@inproceedings{vaswani2021scaling,
  title={Scaling local self-attention for parameter efficient visual backbones},
  author={Vaswani, Ashish and Ramachandran, Prajit and Srinivas, Aravind and Parmar, Niki and Hechtman, Blake and Shlens, Jonathon},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={12894--12904},
  year={2021}
}