Exps 1 : Finetune Vit on Image-r

Purpose : Check the difference of SAM and SGD on Finetune Whole Model, Evaluate the flat(weight-loss), and the feature space

Set: SGD vs SAM, Finetune whole model vs Finetune only the calssifier

Base Model : vit_base_patch16_224

Downstream Dataset: imagenet-r

Exp1: Finetune whole model

For SGD:

Exp2: Fix the backbone Finetune only the linear Classifier

python Result_analyse/analyse_total.py

：

• python scripts/prepare_tiny_imagenet_c_split.py –in-root data/tiny-imagenet-c/extracted/Tiny-ImageNet-C –out-root data/tiny-imagenet-c-r –train-ratio 0.8 –seed 42

python scripts/prepare_tiny_imagenet_p_split.py –in-tar-root data/tiny-imagenet-p/extracted/Tiny-ImageNet-P –out-root data/tiny-imagenet-p-r –train-ratio 0.8 –seed 42 –frames-per-video 1