README.md

LIFE

This repo is an official implementation for the Local InFormation Extractor (LIFE) Module framework.

LIFE Module Data Flow

Installation

Requirements

• Python >= 3.6
• PyTorch >= 1.8 and torchvision
• timm:
  • pip install timm

Prepare data

Download and extract ImageNet100 train and val images from https://www.kaggle.com/datasets/ambityga/imagenet100.

The directory structure is the standard layout for the torchvision datasets.ImageFolder, and the training and validation data is expected to be in the train/ folder and val/ folder respectively:

/path/to/ImageNet100/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class/2
      img4.jpeg

Usage

Training

# Running training procedure with specific GPU number
./tools/run_dist_launch.sh <GPU_NUM> python main.py [optional arguments]

Example command for training DeiT Small with LIFE module with batch size of 512 on 2 GPUs

./tools/run_dist_launch.sh 2 python main.py --data-path /path/to/ImageNet100 --data-set IMNET100 --model deit_small_patch16_224_ms --batch-size 256 --input-size 224 --epochs 300 --output_dir /path/to/output/folder

Please refer to main.py for more optional arguments

Inference

# Running inference procedure with specific GPU number and model path
./tools/run_dist_launch.sh <GPU_NUM> <path_to_config> --eval --resume <model_path> [optional arguments]

Please refer to main.py for more optional arguments

Image Classification Results

Object Detection Results on the VOC dataset

Semantic Segmentation Results on the CityScapes dataset

Citation

Please cite the paper in your publications if it helps your research.

@article{akkaya2024enhancing,
  title={Enhancing performance of vision transformers on small datasets through local inductive bias incorporation},
  author={Akkaya, Ibrahim Batuhan and Kathiresan, Senthilkumar S and Arani, Elahe and Zonooz, Bahram},
  journal={Pattern Recognition},
  pages={110510},
  year={2024},
  publisher={Elsevier}
}

Acknowledgements

This code is heavily borrowed from VTPack.