COME: Test-time Adaption by Conservatively Minimizing Entropy (ICLR'25)

This is the official implementation of COME: Test-time Adaption by Conservatively Minimizing Entropy on ICLR 2025. We propose Conservatively Minimizating Entropy (COME) as a simple drop-in refinement of Entropy Minimization for test-time adaption.

Installation Requirements

To get started with this repository, you need to follow these installation.

pip install -r requirements.txt

Data preparation

We follow Robustness bench and OpenOOD to prepare the datasets. We provide the links to download each dataset:

• ImageNet-C: download it from here 🔗.

The following datasets are only used in open-world TTA setting:

• iNaturalist: download it from this link
• NINCO, SSB_Hard, Texture, Open-ImageNet: download them from this link

Usage Example

COME can be implemtented by simply replacing the loss function of previous TTA algorithms i.e., Tent, EATA, and SAR from softmax entropy to entropy of opinion.

def entropy_of_opinion(x: torch.Tensor): #key component of COME
    x = x / torch.norm(x, p=2, dim=-1, keepdim=True) * torch.norm(x, p=2, dim=-1, keepdim=True).detach()
    brief = torch.exp(x)/(torch.sum(torch.exp(x), dim=1, keepdim=True) + 1000)
    uncertainty = K / (torch.sum(torch.exp(x), dim=1, keepdim=True) + 1000)
    probability = torch.cat([brief, uncertainty], dim=1) + 1e-7
    entropy = -(probability * torch.log(probability)).sum(1)
    return entropy

 def forward_and_adapt(x, model, optimizer, args):
    """Forward and adapt model on batch of data.
    Measure entropy of the model prediction, take gradients, and update params.
    """
    outputs = model(x)
    # COME: replace softmax_entropy with entropy_of_opinion
    loss = entropy_of_opinion(outputs) 
    loss = loss.mean(0)
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()
    return outputs

Reproduce the results

Baselines

1. no_adapt (source): The original model without any adaptation.
2. Tent & EATA & SAR: Previous TTA methods using Entropy Minimization.
3. Tent_COME & EATA_COME & SAR_COME: The enhanced COME version.

Running the Code

To run the experiments, execute the script start.sh.

Results

Classification Accuracy Comparison on ImageNet-C (Level 5). Substantial (≥ 0.5) improvements compared to the baseline are marked with +.

Methods	COME	Gauss.	Shot	Impul.	Defoc	Glass	Motion	Zoom	Snow	Frost	Fog	Brit.	Contr.	Elast.	Pixel	JPEG	Avg. Acc↑
no_adapt	✗	35.1	32.2	35.9	31.4	25.3	39.4	31.6	24.5	30.1	54.7	64.5	49.0	34.2	53.2	56.5	39.8
Tent	✗	52.4	51.8	53.3	53.0	47.6	56.8	47.6	10.6	28.0	67.5	74.2	67.4	50.2	66.7	64.6	52.8
	✓	53.8	53.7	55.3	55.7	51.7	59.7	52.7	59.0	61.7	71.3	78.2	68.7	57.7	70.5	68.2	61.2
	Improve	+1.4	+1.9	+1.9	+2.7	+4.1	+2.9	+5.0	+48.4	+33.6	+3.9	+4.0	+1.3	+7.5	+3.8	+3.6	+8.4
EATA	✗	55.9	56.5	57.1	54.1	53.3	61.9	58.7	62.1	60.2	71.3	75.4	68.5	62.8	69.3	66.6	62.2
	✓	56.2	56.6	57.2	58.1	57.6	62.5	59.5	65.5	63.9	72.5	78.1	69.7	66.5	72.4	70.7	64.5
	Improve	+0.3	+0.2	+0.1	+4.1	+4.3	+0.6	+0.7	+3.5	+3.7	+1.2	+2.7	+1.2	+3.7	+3.1	+4.0	+2.2
SAR	✗	52.7	52.1	53.6	53.5	48.9	56.7	48.8	22.5	51.9	67.5	73.4	66.8	52.7	66.3	64.5	55.5
	✓	56.2	56.5	57.5	58.3	56.7	62.9	58.2	65.3	64.8	72.6	78.5	69.3	64.4	71.9	69.5	64.2
	Improve	+3.5	+4.4	+3.8	+4.8	+7.7	+6.2	+9.5	+42.9	+12.8	+5.0	+5.1	+2.5	+11.6	+5.6	+5.0	+8.7

Citation

If you find COME helpful in your research, please consider citing our paper:

@inproceedings{zhang2025come,
  title={COME: TEST-TIME ADAPTION BY CONSERVATIVELY
MINIMIZING ENTROPY},
  author={Qingyang Zhang, Yatao Bian, Xinke Kong, Peilin Zhao, Changqing Zhang},
  booktitle = {Internetional Conference on Learning Representations},
  year = {2025}
}

Acknowledgment

This repo is developed upon SAR 🔗.

For any additional questions, feel free to email [email protected].