README.md

---
title: Retina Training
emoji: 🐢
colorFrom: red
colorTo: green
sdk: docker
pinned: false
license: mit
short_description: Training model for some retina dataset
---

# RETFound MAE – Hugging Face Space (FastAPI): Train + Inference

This Space lets you upload a zipped ImageFolder dataset, fine-tune a classifier head on top of RETFound MAE, run predictions, and push the trained model to the Hub.

## Endpoints
- `POST /upload_dataset` → form-data file=`dataset.zip` (contains `train/`, `val/`)
- `POST /train` → form fields: `epochs`, `batch_size`, `lr`, `freeze_backbone`
- `GET /status` → training status & metadata
- `POST /predict` → form-data file=`image.jpg`
- `POST /push` → optional form field: `repo_id` (else uses `HF_PUSH_REPO`)

## Env Vars
- `HF_BASE_MODEL_REPO` (e.g., `username/retfound-model`)
- `HF_BASE_MODEL_FILE` (e.g., `RETFound_mae_meh.pth`)
- `HF_PUSH_REPO` (e.g., `username/retfound-classifier`)

## Dataset format
```

zip-root/
train/
ClassA/*.jpg
ClassB/*.jpg
val/
ClassA/*.jpg
ClassB/*.jpg

```

## Notes
- GPU recommended. If CPU-only, reduce batch size.
- Validation accuracy is reported; best checkpoint saved to `checkpoints/`.