README.md

---
license: mit
library_name: transformers
pipeline_tag: video-classification
tags:
  - video-classification
  - zero-shot
  - vision
  - acaua
datasets:
  - kinetics-400
base_model: microsoft/xclip-base-patch32
---

# X-CLIP (base, patch 32) — acaua mirror

MIT-licensed mirror hosted under `CondadosAI/` for use with the [acaua](https://github.com/CondadosAI/acaua) computer vision library.

This is a **safetensors-only mirror** of the upstream Microsoft weights at the pinned commit shown below. The legacy `pytorch_model.bin` (pickle format) that upstream ships alongside `model.safetensors` has been deliberately removed for security hygiene — pickle loads can execute arbitrary code, and `transformers` auto-prefers safetensors when both are present, so removing it has zero functional impact on downstream users.

X-CLIP is a **zero-shot video classification** model: you provide a list of candidate text labels at inference time and the model ranks them by similarity to the video clip. It is not a closed-set softmax classifier, and it does not appear in `AutoModelForVideoClassification`.

## Provenance

| | |
|---|---|
| Upstream repo | [`microsoft/xclip-base-patch32`](https://huggingface.co/microsoft/xclip-base-patch32) |
| Upstream commit SHA | `a2e27a78a2b5d802e894b8a1ef14f3a8ce490963` |
| Upstream commit date | 2024-02-04 |
| Declared license | MIT |
| Paper | Ni et al., *"Expanding Language-Image Pretrained Models for General Video Recognition"*, ECCV 2022, arXiv:[2208.02816](https://arxiv.org/abs/2208.02816) |
| Official code | [`microsoft/VideoX`](https://github.com/microsoft/VideoX) (MIT) |
| Mirrored on | 2026-04-23 |
| Mirrored by | [CondadosAI/acaua](https://github.com/CondadosAI/acaua) |

## Usage via acaua

```python
import acaua

model = acaua.Model.from_pretrained(
    "CondadosAI/xclip_base_patch32",
    allow_non_apache=True,  # weights are MIT, not Apache-2.0
)
result = model.predict(
    "dance.mp4",
    labels=["dancing", "cooking", "running", "sleeping", "walking"],
    top_k=3,
)
for label, score in zip(result.labels, result.scores.tolist()):
    print(f"{label}: {score:.3f}")
```

## Usage via 🤗 Transformers

This mirror is drop-in compatible with the upstream repo.

```python
from transformers import XCLIPModel, XCLIPProcessor

processor = XCLIPProcessor.from_pretrained("CondadosAI/xclip_base_patch32")
model = XCLIPModel.from_pretrained("CondadosAI/xclip_base_patch32")
```

## Expected input

- **Frames:** 8 uniformly-sampled frames per clip (`vision_config.num_frames=8`).
- **Resolution:** 224 × 224 after resize + center-crop.
- **Normalization:** ImageNet mean/std (handled by `XCLIPProcessor`).
- **Text prompts:** supplied at inference time — any natural-language strings.

## License and attribution

Redistributed under MIT, consistent with the upstream declaration. See [`NOTICE`](./NOTICE) for required attribution.

## Citation

```bibtex
@inproceedings{ni2022expanding,
  title={Expanding language-image pretrained models for general video recognition},
  author={Ni, Bolin and Peng, Houwen and Chen, Minghao and Zhang, Songyang and Meng, Gaofeng and Fu, Jianlong and Xiang, Shiming and Ling, Haibin},
  booktitle={European Conference on Computer Vision (ECCV)},
  pages={1--18},
  year={2022},
  publisher={Springer}
}
```