IMSE

Official repository of IMSE: Intrinsic Mixture of Spectral Experts Fine-tuning for Test-Time Adaptation (ICLR 2026 Poster) Supports multiple adaptation methods including Source, TENT, CoTTA, ViDA, and IMSE.

Requirements

• Python 3.9+
• CUDA GPU
• Conda

Installation

conda env create -f environment.yaml
conda activate vida

Data Preparation

ImageNet-C (Required)

Download ImageNet-C and organize it as shown below. The ImageNet-C folder must be placed inside the path specified by --data_dir. The default path can be changed in conf.py Line 269.

<data_dir>/
└── ImageNet-C/
├── gaussian_noise/
│ ├── 1/
│ ├── 2/
│ ├── 3/
│ ├── 4/
│ └── 5/
├── shot_noise/
├── impulse_noise/
├── defocus_blur/
├── ...
└── jpeg_compression/

ImageNet (Optional, for IMSE)

IMSE requires the ImageNet validation set for source domain feature extraction. Update the data_dir path in the prepare_test_data function at main.py Line 251.

ViDA Source Model (Required for ViDA)

Download the ViDA source model from Google Drive and place it in the checkpoint/ folder.

Usage

Basic

python main.py --cfg ./cfgs/vit/source.yaml --data_dir <path_to_data> --exp_name my_exp

IMSE-Retrieval

We provide two domain selection strategies for continual TTA:

Select most similar domain, IMSE-Retrieval

CUDA_VISIBLE_DEVICES=0 python main.py --cfg ./cfgs/vit/imse.yaml --data_dir <path_to_data> --exp_name my_exp \
SETTING continual

Domain-distance-based spectral code generation, extension of IMSE-Retrieval

CUDA_VISIBLE_DEVICES=0 python main.py --cfg ./cfgs/vit/imse.yaml --data_dir <path_to_data> --exp_name my_exp \
SETTING continual SVD.DYNAMIC_MODE mix_adapt SVD.TEMP 1.0

Scripts

Pre-defined scripts for each setting are available in the bash/ directory.

Script	Setting	Methods
bash/0_others_ctta.sh	C-TTA	Source, CoTTA, TENT, ViDA
bash/1_imse_ctta.sh	C-TTA	IMSE
bash/2_others_recurring.sh	Recurring	CoTTA, ViDA, TENT
bash/3_imse_recurring.sh	Recurring	IMSE

Project Structure

.
├── main.py # Entry point
├── conf.py # Configuration
├── environment.yaml # Conda environment
├── bash/ # Shell scripts
├── cfgs/ # Config files
├── method/ # TTA methods (tent, cotta, vida, svd)
├── robustbench/ # Data loading, model zoo
├── timm/ # ViT models
└── ckpt/ # ViT Checkpoint