---
library_name: transformers
tags:
- image-classification
- vision
- roadwork-detection
- bittensor
- natix-subnet
---

# roadwork_v3_focal

Vision Transformer (ViT) model for binary roadwork detection, trained for Natix Subnet 72.

## Model Details

- **Base Model**: google/vit-base-patch16-224-in21k
- **Architecture**: ViT-Base (86M parameters)
- **Input Size**: 224×224
- **Output Classes**: 2 (None, Roadwork)

## Performance

| Metric | Value |
|---|---|
| Accuracy | 0.9839 |
| MCC | 0.9443 |
| F1 Score | 0.9903 |
| AUC | 0.9950 |
| Subnet Reward | 0.8962 |

## Training Details

- **Training Samples**: 34,968
- **Validation Samples**: 4,418
- **Validator Augmentations**: True
- **Label Smoothing**: 0.05
- **Confidence Margin**: 0.1

## Usage

```python
from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image

processor = AutoImageProcessor.from_pretrained("infinite000/in-20001")
model = AutoModelForImageClassification.from_pretrained("infinite000/in-20001")

image = Image.open("your_image.jpg")
inputs = processor(image, return_tensors="pt")
outputs = model(**inputs)
probs = outputs.logits.softmax(dim=1)
roadwork_prob = probs[0][1].item()
```

## Model Card

See `model_card.json` for detailed metadata.