app.py

import os, json, subprocess, shutil, zipfile
from fastapi import BackgroundTasks, FastAPI, UploadFile, File
from transformers import AutoImageProcessor, BeitForImageClassification
from PIL import Image
import torch

MODEL_DIR = os.environ.get("OUTPUT_DIR", "/home/user/outputs/beit-retina")
DATA_DIR = os.environ.get("DATA_DIR", "data2")
CLASSES = ["AMD","DMO","DR","GLC","HR","Normal"]

app = FastAPI(title="Retina Disease Classifier")

processor = None
model = None

# ----------------------------
# MODEL LOADING
# ----------------------------
def load_model():
    global processor, model, CLASSES
    try:
        processor = AutoImageProcessor.from_pretrained(MODEL_DIR)
        model = BeitForImageClassification.from_pretrained(MODEL_DIR)
        labels_path = os.path.join(MODEL_DIR, "labels.json")
        if os.path.exists(labels_path):
            with open(labels_path) as f:
                CLASSES = json.load(f)
        print("✅ Model and processor loaded successfully")
    except Exception as e:
        processor, model = None, None
        print(f"⚠️ Skipping model load: {e}")

# ----------------------------
# BACKGROUND TRAINING
# ----------------------------
def run_training():
    try:
        print("🔹 Starting training subprocess...")
        process = subprocess.Popen(
            ["python", "train2.py"],
            stdout=subprocess.PIPE,
            stderr=subprocess.STDOUT,
            universal_newlines=True
        )
        for line in iter(process.stdout.readline, ""):
            print("TRAIN_LOG:", line.strip())
        process.stdout.close()
        return_code = process.wait()

        if return_code == 0 and os.path.exists(MODEL_DIR):
            load_model()
            print("✅ Training complete and model reloaded")
        else:
            print(f"❌ Training failed with code {return_code}")
    except Exception as e:
        print("⚠️ Training exception:", str(e))

# ----------------------------
# FASTAPI STARTUP
# ----------------------------
@app.on_event("startup")
def startup_event():
    if os.path.exists(MODEL_DIR):
        load_model()
    else:
        print("⚠️ MODEL_DIR not found, skipping model load")

# ----------------------------
# ENDPOINTS
# ----------------------------
@app.post("/load-data")
async def load_data(file: UploadFile = File(...)):
    """
    Upload a ZIP file, extract into `data/` folder for training.
    """
    print("🔹 Received dataset ZIP upload...")
    if os.path.exists(DATA_DIR):
        shutil.rmtree(DATA_DIR)
    os.makedirs(DATA_DIR, exist_ok=True)

    zip_path = "dataset.zip"
    with open(zip_path, "wb") as f:
        f.write(await file.read())
    print(f"   ↪ Saved ZIP to {zip_path}")

    with zipfile.ZipFile(zip_path, "r") as zip_ref:
        zip_ref.extractall(DATA_DIR)
    print(f"✅ Dataset extracted to {DATA_DIR}")

    os.remove(zip_path)
    return {"status": "Dataset uploaded and extracted"}

@app.post("/train")
async def train_endpoint(background_tasks: BackgroundTasks):
    background_tasks.add_task(run_training)
    return {"status": "Training started in background"}

@app.post("/predict")
async def predict(file: UploadFile):
    if model is None:
        return {"error": "Model not trained yet"}
    try:
        img = Image.open(file.file).convert("RGB")
    except Exception as e:
        return {"error": f"Invalid image: {str(e)}"}
    inputs = processor(images=img, return_tensors="pt")
    with torch.no_grad():
        logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=1)[0].tolist()
    pred_id = int(torch.argmax(logits, dim=1).item())
    return {
        "class_id": CLASSES[pred_id],
        "probabilities": {CLASSES[i]: float(p) for i, p in enumerate(probs)}
    }