saft 1.1.0

SAFT: Self-Attention Factor-Tuning for Parameter-Efficient Fine-Tuning

SAD

Implementation of Self-Attention Tensor-Decomposition, an original technique for Parameter-Efficient Fine-Tuning

Requirements

- Python == 3.8
- torch == 1.10.0
- torchvision == 0.11.1
- timm == 0.4.12
- avalanche-lib == 0.1.0

Data

To get started, head over to the SSF data preparation page and download the VTAB-1K dataset: SSF Data Preparation. Once you've obtained the data, place the dataset folders into <YOUR PATH>/SAD/data/.

Pretrained Model

Download the pretrained ViT-B/16 to <YOUR PATH>/SAD/ViT-B_16.npz

Results (.055 Million Trainable Backbone Parameters) (ViT-B/16)

Coming soon...

Caltech101: 90.6%

Run

cd <YOUR PATH>/SAD
python3 Test.py

Modular Quick-Start

from SAD import SAD
from Utils import *

if __name__ == "__main__":
    sad = SAD(
        model='vit_base_patch16_224',
        num_classes=get_classes_num('oxford_flowers102'),
        validation_interval=1,
        rank=3,
        scale=10
    )
    train_dl, test_dl = get_data('oxford_flowers102')
    sad.upload_data(train_dl, test_dl)
    sad.train(10)