Retina Disease Classification – FastAPI Inference Server

This repository provides a minimal FastAPI server to deploy your trained ViT model for retina OCT disease classification.

Prepare the model

After training in your notebook/script, save artifacts to a directory, for example my-trained-vit-model/:

model.save_pretrained('my-trained-vit-model')
processor.save_pretrained('my-trained-vit-model')

Ensure the directory contains files like config.json, pytorch_model.bin (or safetensors), preprocessor_config.json, etc.

Install

python -m venv .venv
. .venv/Scripts/activate  # Windows PowerShell: .venv\Scripts\Activate.ps1
pip install -r requirements.txt

Run the API

set MODEL_DIR=my-trained-vit-model  # Windows
uvicorn app.main:app --host 0.0.0.0 --port 8000

Open http://127.0.0.1:8000 to use the simple HTML upload form.

Health check:

curl http://127.0.0.1:8000/health

Predict with cURL:

curl -X POST "http://127.0.0.1:8000/predict?top_k=3" \
  -H "Content-Type: multipart/form-data" \
  -F "file=@sample.jpg"

Sample JSON response:

{
  "predictions": [
    {"label": "CNV", "score": 0.91},
    {"label": "DME", "score": 0.06},
    {"label": "DRUSEN", "score": 0.02}
  ]
}

Notes

• The server reads MODEL_DIR env var; defaults to my-trained-vit-model in the repo root.
• CPU inference by default. If you want GPU, move tensors to CUDA and ensure PyTorch with CUDA is installed.