---
license: apache-2.0
library_name: transformers
pipeline_tag: image-classification
base_model: facebook/deit-tiny-patch16-224
datasets:
- AI-Lab-Makerere/beans
metrics:
- accuracy
- f1
tags:
- image-classification
- mlflow
---

# beans-deit-tiny

Fine-tuned `facebook/deit-tiny-patch16-224` for image classification (3 classes: angular_leaf_spot, bean_rust, healthy).
Trained in Google Colab (GPU), experiments tracked with MLflow, winner selected by validation accuracy.

## Training
- Dataset: `AI-Lab-Makerere/beans`
- Train / val / test: 1034 / 133 / 128
- Learning rate: 0.0002
- Epochs: 3
- Batch size: 32
- MLflow run: `run1_lr0.0002_ep3` (a4050afef1c545cd8e87cc47cba1bd02)

## Results
| split | accuracy | macro F1 |
|---|---|---|
| validation | 0.9624 | 0.9628 |
| test | 0.9453 | 0.9462 |

### Per-class metrics (test)
| class             |   precision |   recall |     f1 |   support |
|:------------------|------------:|---------:|-------:|----------:|
| angular_leaf_spot |        1    |   0.8837 | 0.9383 |        43 |
| bean_rust         |        0.86 |   1      | 0.9247 |        43 |
| healthy           |        1    |   0.9524 | 0.9756 |        42 |

![Confusion matrix](eval/confusion_matrix.png)

## Usage
```python
from transformers import pipeline
clf = pipeline("image-classification", model="Konradowsky/beans-deit-tiny")
clf("image.jpg", top_k=3)
```