Adaptive Debiasing Tsallis Entropy for Test-Time Adaptation.

Introduction

This repo officially implements Adaptive Debiasing Tsallis Entropy for Test-Time Adaptation.[ICLR2026]

Fig.1: (a) VLM bias, showing higher confidence and accuracy for head classes and lower confidence and accuracy for tail classes. (b) The standard Shannon Entropy~(SE)-based method is widely used in TTA. (c) and (d) Our proposed method, which uses Tsallis Entropy~(TE) and Adaptive Debiasing Tsallis Entropy~(ADTE) for selecting high-confidence views.

Dataset dir:

dataset
	|-imagenet-1k
	|-imagenet-a
	|-imagenet-r
	|-imagenet-s
	|-imagenet-v2
	|-cross_domain_dataset...

Dataset download and setup can be referred at CoOp

Run:

ImageNet-A, Zero-shot CLIP, hand-crafted templates, ViT-B/16:

python run.py --set_id=A --gpu=0 --is_zero --arch=ViT-B/16 --templates

ImageNet-A, Shannon Entropy, hand-crafted templates, ViT-B/16:

python run.py --set_id=A --gpu=0 --arch=ViT-B/16 --templates

ImageNet-A, Tsallis Entropy (q=0.1), hand-crafted templates, ViT-B/16:

python run.py --set_id=A --gpu=0 --same_q --scale_w=0.1 --arch=ViT-B/16 --templates

ImageNet-A, Adaptive Debiasing Tsallis Entropy, hand-crafted templates, ViT-B/16:

python run.py --set_id=A --gpu=0 --use_adte --arch=ViT-B/16 --templates --q_path=bias_pt/imagenet_A_q_1_10_ensemble_B16.pt

Code:

Shannon Entropy

# Shannon Entropy
def shannon_entropy_minimization(logits, probs, bias):
    entropy = -torch.sum(probs * torch.log(probs + 1e-8), dim=1)
    return entropy

Tsallis Entropy and Adaptive Debiasing Tsallis Entropy

def tsallis_entropy_minimization(probs, q=1):
    if torch.is_tensor(q):
        # Adaptive Debiasing Tsallis Entropy
        entropy = torch.sum((torch.pow(probs,q))/(1 - q), dim=1)
    else:
        # Tsallis Entropy
        entropy = (1 - torch.sum(probs ** q, dim=1)) / (q - 1)
    return entropy