Bag of Tricks for Fully Test-Time Adaptation

This repository contains the code used for Bag of Tricks for Fully Test-Time Adaptation 🔗 by Saypraseuth Mounsaveng, Florent Chiaroni, Malik Boudiaf, Marco Pedersoli and Ismail Ben Ayed (WACV 2024).

Code

The code is mainly inspired from SAR 🔗 and adapted with additional methods and tricks.

Links to the source code of the methods mentioned in the article: The code used in this paper was mostly inspired by the

Method	Code Link
Tent	https://github.com/DequanWang/tent
SAR	https://github.com/mr-eggplant/SAR
Delta	https://github.com/bwbwzhao/DELTA
DUA	https://github.com/jmiemirza/DUA
Hebbian	n/a

Preparation

Links to the weights of the pretrained models mentioned in the paper:

Architecture	Code Link
ResNet50-BN	https://download.pytorch.org/models/resnet50--9c8e357.pth
ResNet50-GN	timm
ResNet-101	https://github.com/Albert0147/NRC\_SFDA
VitBase-LN	timm
WRN28-10	RobustBench
WRN40-2	RobustBench
SVHN model	Pytorch-Playground

Installation:

Packages to install:

Data preparation:

This repository contains code for evaluation on different datasets. Here are the links to download them:

ImageNet-C 🔗. ImageNet-Sketch 🔗. ImageNet-Rendition 🔗. VisDA-C 2017 🔗. CIFAR10 🔗. CIFAR100 🔗. MNIST 🔗. MNIST-M 🔗. USPS 🔗.

Example: Adapting a pre-trained model on ImageNet-C (Corruption).

Usage:

python3 main.py --data_corruption /path/to/imagenet-c --exp_type [normal/bs1/mix_shifts/label_shifts] --method [no_adapt/tent/delta/sar] --model [resnet50_bn_torch/resnet50_gn_timm/vitbase_timm] --test_batch_size 16 --output /output/dir

'--exp_type' is choosen from:

'normal' means the same test setting to prior mild data stream in Tent and EATA
'bs1' means single sample adaptation, only one sample comes each time-step
'mix_shifts' conducts exps over the mixture of 15 corruption types in ImageNet-C
'label_shifts' means exps under online imbalanced label distribution shifts. Moreover, imbalance_ratio indicates the imbalance extent

Experimental results

Please check our PAPER 🔗 for experimental results.

Correspondence

For any questions or comments, please contact Saypraseuth Mounsaveng by [saypraseuth.mounsaveng.1 at etsmtl.net] . 📬

Citation

If you think our work is helpful for the community, please consider citing it:

@InProceedings{Mounsaveng_2024_WACV,
    author    = {Mounsaveng, Saypraseuth and Chiaroni, Florent and Boudiaf, Malik and Pedersoli, Marco and Ben Ayed, Ismail},
    title     = {Bag of Tricks for Fully Test-Time Adaptation},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024}
}

smounsav/tta_bot