oonat/ezpc-embeddings

EZPC - Pre-computed CLIP / SigLIP Embeddings

This dataset hosts the pre-computed image embeddings, cached text (classname / concept) embeddings, and class splits used in Explaining CLIP Zero-shot Predictions Through Concepts (CVPR 2026).

Running EZPC from scratch requires extracting CLIP / SigLIP features for every image in five benchmark datasets (CIFAR-100, CUB-200-2011, Places365, ImageNet, ImageNet-100) with multiple backbones. To skip the expensive feature-extraction step, we release the exact tensors we used in the paper - drop them into the EZPC repo's data/ folder and you can train, evaluate, and reproduce quantitative results without touching the raw images.

• 📄 Paper: arXiv:2603.28211
• 💻 Code: github.com/oonat/ezpc
• 🌐 Project page: oonat.github.io/ezpc
• 🤗 Checkpoints: oonat/ezpc-checkpoints

What's Included

For each of the five datasets and each supported backbone, this repository provides:

• Full train / test embeddings + targets: CLIP/SigLIP image features for every training and test image across all classes, along with their class indices. Useful as a general-purpose feature dump and as the input to data/split_dataset.py if you want to regenerate the splits under a different seed or ratio.
• Seen-split train / test embeddings: embeddings restricted to the 80% "seen" classes. The train split is used to fit the EZPC projection matrix A; the test split measures GZSL seen-class accuracy.
• Unseen-split train / test embeddings: embeddings restricted to the 20% "unseen" classes held out from training. Used for GZSL unseen-class evaluation (and for the --target_dataset side of cross-dataset transfer).
• Split-aligned target files: seen_train_ids.pt, seen_test_ids.pt, unseen_train_ids.pt, unseen_test_ids.pt, each giving the class index for every row of the corresponding split.
• Class split file: the exact seen / unseen class partition (seed 42, 80/20) used in the paper.
• Cached text embeddings: L2-normalized "a photo of {x}" text embeddings for class names ({backbone}_classname_embs.pt, shape (C, d)) and concepts ({backbone}_concept_matrix.pt, shape (m, d)). test.py loads these automatically so evaluation reproduces the reported numbers exactly and independent of GPU / CUDA version — without re-running the CLIP/SigLIP text encoder.

Supported backbones: CLIP RN50, CLIP ViT-B/32, CLIP ViT-L/14, SigLIP ViT-SO400M/14.

Repository Layout

Each dataset follows the same two-level structure: top-level embeddings/ holds the full image feature dump (all classes) plus the cached text embeddings, and embeddings/splits/ holds the seen/unseen partition used throughout the paper.

data/
├── CIFAR-100/
│   └── embeddings/
│       ├── RN50_train_embeddings.pt                       # (N_train, d) - all classes
│       ├── RN50_test_embeddings.pt                        # (N_test,  d) - all classes
│       ├── ViT-B-32_train_embeddings.pt
│       ├── ViT-B-32_test_embeddings.pt
│       ├── ViT-L-14_train_embeddings.pt
│       ├── ViT-L-14_test_embeddings.pt
│       ├── siglip-so400m-patch14-384_train_embeddings.pt
│       ├── siglip-so400m-patch14-384_test_embeddings.pt
│       ├── RN50_classname_embs.pt                         # (C, d) - class-name text embeddings
│       ├── RN50_concept_matrix.pt                         # (m, d) - concept text embeddings
│       ├── ViT-B-32_classname_embs.pt
│       ├── ViT-B-32_concept_matrix.pt
│       ├── ViT-L-14_classname_embs.pt
│       ├── ViT-L-14_concept_matrix.pt
│       ├── siglip-so400m-patch14-384_classname_embs.pt
│       ├── siglip-so400m-patch14-384_concept_matrix.pt
│       ├── train_ids.pt                                   # (N_train,)  - class indices
│       ├── test_ids.pt                                    # (N_test,)   - class indices
│       └── splits/
│           ├── class_split.pt                             # {seen_classes, unseen_classes}
│           ├── RN50_seen_train_embs.pt
│           ├── RN50_seen_test_embs.pt
│           ├── RN50_unseen_train_embs.pt
│           ├── RN50_unseen_test_embs.pt
│           ├── ViT-B-32_seen_train_embs.pt
│           ├── ViT-B-32_seen_test_embs.pt
│           ├── ViT-B-32_unseen_train_embs.pt
│           ├── ViT-B-32_unseen_test_embs.pt
│           ├── ViT-L-14_seen_train_embs.pt
│           ├── ViT-L-14_seen_test_embs.pt
│           ├── ViT-L-14_unseen_train_embs.pt
│           ├── ViT-L-14_unseen_test_embs.pt
│           ├── siglip-so400m-patch14-384_seen_train_embs.pt
│           ├── siglip-so400m-patch14-384_seen_test_embs.pt
│           ├── siglip-so400m-patch14-384_unseen_train_embs.pt
│           ├── siglip-so400m-patch14-384_unseen_test_embs.pt
│           ├── seen_train_ids.pt
│           ├── seen_test_ids.pt
│           ├── unseen_train_ids.pt
│           └── unseen_test_ids.pt
├── CUB-200-2011/
│   └── embeddings/ ...
├── ImageNet/
│   └── embeddings/ ...
├── ImageNet-100/
│   └── embeddings/ ...
└── Places365/
    └── embeddings/ ...

Quickstart

1. Download the data

pip install huggingface-hub

# Grab everything
hf download oonat/ezpc-embeddings \
    --repo-type dataset \
    --local-dir data

# Or just one dataset
hf download oonat/ezpc-embeddings \
    --repo-type dataset \
    --local-dir data \
    --include "CIFAR-100/*"

# Or only what EZPC needs to evaluate: the GZSL splits + cached text embeddings
hf download oonat/ezpc-embeddings \
    --repo-type dataset \
    --local-dir data \
    --include "*/embeddings/splits/*" "*/embeddings/*_classname_embs.pt" "*/embeddings/*_concept_matrix.pt"

2. Plug it into EZPC

Clone the EZPC GitHub repo and point --dataset_root at the folder you downloaded:

git clone https://github.com/oonat/ezpc.git
cd ezpc
conda env create -f environment.yml
conda activate ezpc
pip install -e .

# data/ lives next to train.py now - run training:
python train.py \
    --dataset CIFAR-100 \
    --dataset_root ./data \
    --backbone RN50 \
    --lambda_weight 1.0 \
    --lr 0.01 \
    --num_epochs 10000

Citation

If you use this dataset, please cite:

@InProceedings{Ozdemir_2026_CVPR,
    author    = {Ozdemir, Onat and Christensen, Anders and Alaniz, Stephan and Akata, Zeynep and Akbas, Emre},
    title     = {Explaining CLIP Zero-shot Predictions Through Concepts},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {31336-31345}
}

Please also cite the original datasets whose images were used to produce these embeddings (CIFAR-100, CUB-200-2011, Places365, ImageNet, ImageNet-100) and the backbone models (CLIP, SigLIP).

Acknowledgements

• Concept vocabularies and class label mappings are taken from Label-free Concept Bottleneck Models. We thank the authors for open-sourcing them.
• Backbones from OpenAI CLIP, OpenCLIP, and Google SigLIP.

License

Released under the MIT License.

Note that these embeddings are derived from CIFAR-100, CUB-200-2011, Places365, ImageNet, and ImageNet-100. Users are responsible for complying with the original license and terms of use of those datasets, which may restrict commercial use — notably ImageNet and CUB-200-2011, which are released for non-commercial research only.