---
library_name: transformers
tags:
  - vision
  - image-feature-extraction
  - dinov3
  - vit
license: other
---

# EUPE ViT-B/16

This repository contains the EUPE ViT-B/16 backbone converted to the Hugging Face `transformers` DINOv3 ViT format.

- Hugging Face repo: `kittn/eupe_vitb16`
- Original EUPE checkpoint: [`facebook/EUPE-ViT-B`](https://huggingface.co/facebook/EUPE-ViT-B)
- Architecture: ViT-B/16
- Hidden size: `768`
- Layers: `12`
- Attention heads: `12`
- Register tokens: `4`
- Patch size: `16`

This conversion is meant for direct use with `AutoModel` / `DINOv3ViTModel`. It uses the stock Hugging Face DINOv3 implementation, so small numerical differences versus the original EUPE implementation are expected.

## RoPE note

The original EUPE checkpoints use persisted RoPE values derived from bf16-rounded periods, while the stock Hugging Face DINOv3 implementation reconstructs RoPE analytically from `rope_theta`. Because of that, outputs are very close but not exactly identical to the original EUPE implementation.

## Minimal-loss inference

If you want to minimize the discrepancy versus the original EUPE inference path, prefer running the Hugging Face model on CUDA under `torch.autocast("cuda", dtype=torch.bfloat16)` rather than hard-casting the full model to `bfloat16`.

## Usage

```python
import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModel

repo_id = "kittn/eupe_vitb16"

image = Image.open("image.jpg").convert("RGB")

processor = AutoImageProcessor.from_pretrained(repo_id)
model = AutoModel.from_pretrained(repo_id).eval()

inputs = processor(images=image, return_tensors="pt")

with torch.inference_mode():
    outputs = model(**inputs)

last_hidden_state = outputs.last_hidden_state
pooler_output = outputs.pooler_output

print("last_hidden_state:", last_hidden_state.shape)
print("pooler_output:", pooler_output.shape)
```

On CUDA, a low-loss inference pattern is:

```python
with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
    outputs = model(**inputs)
```

`last_hidden_state` contains:

- token `0`: CLS token
- tokens `1:5`: 4 register tokens
- remaining tokens: patch tokens

## Optional masking

For masked inference, pass `bool_masked_pos` with shape `batch_size x num_patches`. For a `224 x 224` image with patch size `16`, `num_patches = 14 x 14 = 196`.

```python
import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModel

repo_id = "kittn/eupe_vitb16"

image = Image.open("image.jpg").convert("RGB")
processor = AutoImageProcessor.from_pretrained(repo_id)
model = AutoModel.from_pretrained(repo_id).eval()
inputs = processor(images=image, return_tensors="pt")

num_patches = (inputs["pixel_values"].shape[-2] // 16) * (inputs["pixel_values"].shape[-1] // 16)
bool_masked_pos = torch.zeros(1, num_patches, dtype=torch.bool)
bool_masked_pos[:, 0] = True

with torch.inference_mode():
    outputs = model(pixel_values=inputs["pixel_values"], bool_masked_pos=bool_masked_pos)

print(outputs.last_hidden_state.shape)
```