Trainable Projected Gradient Method (TPGM)

This repo implements the experiments in the paper Trainable Projected Gradient Method for Robust Fine-Tuning. Specifically, we divde experiments into two categories: DomainNet experiments using ResNet and ImageNet experiments using ViT. The main difference is in how TPGM is applied.

• The environment uses Pytorch 1.7 supported on CUDA 11.x and python 3.8.

cd TPGM
conda env create -f environment.yml
conda activate py38

• The code resides in the folder DomainNet_ResNet_Exp.
• Following the paper, TPGM is used at every iteration of fine-tuning.

• The code resides in the folder ImageNet_ViT_Exp.
• Following the paper, TPGM is only used at the end of fine-tuning.