README.md

---
license: other
library_name: frozen-vitl
tags:
  - pathology
  - feature-extraction
pipeline_tag: image-feature-extraction
---

# Frozen ViT-L/14

A frozen vision encoder for pathology **feature extraction**.  
病理图像**特征提取**模型。

- Input / 输入: RGB `224×224`
- Output / 输出: CLS `[B, 1024]`

## Usage / 使用

```bash
huggingface-cli download minxoy/forzenpath --local-dir ./forzenpath --token $HF_TOKEN
pip install -e ./forzenpath
```

```python
import torch
from PIL import Image
from frozen_vitl import FrozenVitlModel, FrozenVitlImageProcessor

model = FrozenVitlModel.from_pretrained("minxoy/forzenpath", token=True, device="cuda")
processor = FrozenVitlImageProcessor(model.config)
model.eval()

x = processor(Image.open("patch.jpg")).unsqueeze(0).cuda()
with torch.no_grad():
    feat = model(x).pooler_output  # [1, 1024]
```

Local folder / 本地目录（含 `config.json` 和 `model.safetensors`）：

```python
model = FrozenVitlModel.from_pretrained("./forzenpath", device="cuda")
```

Requires `torch`, `torchvision`, `Pillow`.