---
license: apache-2.0
library_name: libreyolo
pipeline_tag: image-classification
tags:
- image-classification
- vision-transformer
- vit
- imagenet
- libreyolo
datasets:
- imagenet-1k
---

# LibreViTb-cls

Classic ViT-Base/16 image classifier (224px, ImageNet-1k, 1000 classes),
repackaged for [LibreYOLO](https://github.com/LibreYOLO/libreyolo). The model
has 86.6 million parameters and is inference-only in LibreYOLO.

## Source

Derived from [timm](https://github.com/huggingface/pytorch-image-models) model
`vit_base_patch16_224.augreg2_in21k_ft_in1k` at timm v1.0.28 (commit
`8ef73809f622e0031bd7f4940265734aef8b9978`). The exact source checkpoint is
[`timm/vit_base_patch16_224.augreg2_in21k_ft_in1k`](https://huggingface.co/timm/vit_base_patch16_224.augreg2_in21k_ft_in1k/tree/063c6c38a5d8510b2e57df480445e94b231dad2c)
at revision `063c6c38a5d8510b2e57df480445e94b231dad2c`. The source and weights are
Apache-2.0. Copyright (c) Ross Wightman and the timm contributors.

The architecture originates from Google Research's Apache-2.0
[Vision Transformer](https://github.com/google-research/vision_transformer).
These AugReg2 weights use ImageNet-21k pretraining followed by ImageNet-1k
fine-tuning.

## Modifications

Learned parameters are unchanged. Conversion adds LibreYOLO checkpoint
metadata (`model_family`, `size`, `task`, `nc`, `names`, and `imgsz`) only.
LibreYOLO's native graph loads the state dict strictly and produces exactly
the same pretrained logits as timm (`max_abs_diff == 0`). See
`weights/convert_vit_weights.py` and `docs/provenance/vit.md` in the
[LibreYOLO source repository](https://github.com/LibreYOLO/libreyolo).

Source `model.safetensors` SHA-256:
`32aa17d6e17b43500f531d5f6dc9bc93e56ed8841b8a75682e1bb295d722405b`.

## Usage

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreViTb-cls.pt")
result = model.predict("image.jpg")
print(result.probs.top1, result.probs.top5)
```

## License

Apache License 2.0. See [`LICENSE`](./LICENSE) and [`NOTICE`](./NOTICE).