SoongE/TTC-features

TTC features

Pre-extracted features for Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026). Code

Every config of the code reads its features from this tree:

Folder	Features	Config
zeroshot/<backbone>/<dataset>/	CLIP RN50 / ViT-B16, single view, 11 datasets	zeroshot
domain_gen/<backbone>/<dataset>/{0aug,10aug}/	CLIP RN50 / ViT-B16 on ImageNet-A, -R, -Sketch, -V2	domain_generalization with --data.aug 0 or 10 (default)
fewshot/<dataset>/<shots>shots/seed<s>/	CoOp-tuned ViT-B/16, 1–16 shots, CoOp seeds 1–3	fewshot
base2new/<dataset>/{base,new}/seed<s>/	CoOp-tuned ViT-B/16 trained on the base classes	base_to_novel
cross/<dataset>/seed<s>/	16-shot ImageNet CoOp prompts on ImageNet (the 16-shot run) and the other 10 datasets	cross_dataset

Each file is <name>.safetensors holding a single tensor under the key <name>, e.g. load_file("test_f.safetensors")["test_f"]. The test stream, test_f (samples × dim, or samples × views × dim) and test_l, sits in zeroshot/<backbone>/<dataset>/, domain_gen/<backbone>/<dataset>/<views>aug/, fewshot/<dataset>/, base2new/<dataset>/{base,new}/ and cross/<dataset>/. CoOp only tunes the text prompt, so it is shared by every shot count and CoOp seed. The text classifier (classes × dim) sits next to it in zeroshot, one level up in domain_gen (text_weights_cupl, the CuPL classifier from CLIP templates and CuPL descriptions; domain_gen uses the ImageNet one), and in each seed<s>/ folder in fewshot, base2new and cross (text_weights, the classifier of that CoOp prompt), together with the few-shot cache keys_<shots>shots and values_<shots>shots where the setting has one.

Usage

From the root of the code repository:

hf download SoongE/TTC-features --repo-type dataset --local-dir features
python -m scripts.run --config zeroshot

Each config needs only its own folder, so a single setting can be downloaded with --include, e.g. --include "fewshot/*" or --include "zeroshot/ViT-B16/*", or from Python:

from huggingface_hub import snapshot_download
snapshot_download("SoongE/TTC-features", repo_type="dataset", allow_patterns="cross/*", local_dir="features")

The files are tensors read by the TTC code, not tables, so they are not loaded with datasets.load_dataset.

License

The features are derived from the evaluation datasets (ImageNet and its variants, Caltech101, DTD, EuroSAT, FGVC-Aircraft, Oxford Flowers, Food-101, Oxford-IIIT Pet, Stanford Cars, SUN397, UCF101) and keep their terms of use, which are research-only for several of them. CLIP and CoOp are MIT-licensed. See LICENSE.md.

Citation

@inproceedings{oh2026ttc,
  title     = {Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts},
  author    = {Oh, Seungmin and Kang, Seunghun and Ryu, Jongbin},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}