cliptta

KO | EN

gitlite — search

#hacktoberfest #react #ai #llm #go #golang #machine-learning #linux #nodejs #macos #java #security #mcp #open-source

(★ 12)

Official implementation of the NeurIPS 2025 paper: CLIPTTA: Robust Contrastive Vision-Language Test-Time Adaptation

cliptta Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# CLIPTTA [NeurIPS 2025]
Official implementation of the NeurIPS 2025 paper: CLIPTTA: Robust Contrastive Vision-Language Test-Time Adaptation. Please followe the following instructions to execute the main experiments found in the paper.

## Install the library

The source code works as a library for Test-Time Adaptation of VLMs such as CLIP. Several methods from the state-of-the-art are ready to use off-the-shelf. First, install the library in your environment

```
pip install -U pip
pip install -e .
```

Once installed, all the sub-packages will be available to use.

## Running scripts

The hyperparameters of each method are described in detail in the file ```configuration.py```, and the main file to execute any of these 
adaptation strategies is ```main.py```. We also include scripts for closed-set and open-set adaptation inside the folder ```scripts```, 
with the necessary configurations to run CLIPTTA and competitor baselines.

Key changes to be made on the scripts include adding your specific Python environment, as well as defining the parameters ```--root``` 
for the path where you store the project, ```--dataroot``` for the path where you store the datasets, and ```--save_root``` as the path 
where some results can be stored.

## Reproducing the open-set results (Tables 3, 8 and 9)

The OCE loss weight `--beta_ood` (λ<sub>oce</sub> in the paper) is a **per-dataset** hyperparameter.
The paper's implementation details state λ<sub>oce</sub> = 1, which only holds for ImageNet/ImageNet-C; the
values used for all benchmarks are listed below and encoded in `ttavlm/scripts/open_set/*/cliptta.sh`.
The weight differences compensate for the scale of the `max_log_prob` score distribution, which grows with the
number of classes and shrinks under corruption.

| ID dataset | OOD dataset | `--beta_ood` | `--num_shots` / `--sample_size` | `--update_alpha` |
|---|---|---|---|---|
| CIFAR-10 | SVHN | 20.0 | 16 / 128 | yes |
| CIFAR-10-C | SVHN-C | 125.0 | 16 / 128 | yes |
| CIFAR-100 | SVHN | 5.0 | 8 / 128 | yes |
| CIFAR-100-C | SVHN-C | 200.0 | 8 / 128 | yes |
| ImageNet | Places | 1.0 | 8 / 256 | no |
| ImageNet-C | Places-C | 1.0 | 8 / 256 | no |

All other settings are shared: ViT-B/16, Adam, `--lr 1e-4`, `--steps 10`, `--batch_size 128`,
`--ood_batch_size 128`, `--beta_tta 1.0`, `--beta_reg 1.0`, `--loss_ood_type avg_contrastive`,
`--score_type max_log_prob`, `--id_score_type max_prob`, `--use_softmax_entropy`, `--use_memory`.

The Places OOD set is the curated 10,000-image subset of Places365 released with the MOS benchmark
(categories de-duplicated against ImageNet-1k, see
[deeplearning-wisc/large_scale_ood](https://github.com/deeplearning-wisc/large_scale_ood#out-of-distribution-dataset));
place it under `<dataroot>/ood_data/Places`. The corrupted OOD sets SVHN-C and Places-C are not
downloadable; generate them with `ttavlm/scripts/other/generate_corruptions/make_svhn_c.py` and
`make_places_c.py` (severity 5 is the one used in the benchmark, but SVHN-C `.npy` files must contain
all 5 severities).

With these configurations we reproduce (seed 42): CIFAR-10 94.5 ACC / 99.8 AUC / 0.6 FPR95
(paper: 94.6 / 99.8 / 0.4) and CIFAR-10-C 79.0 / 98.7 / 6.1 (paper: 79.1 / 98.7 / 6.1).

cliptta

cliptta Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

cliptta

cliptta Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README