app.py

from fastapi import FastAPI, File, UploadFile
from fastapi.responses import JSONResponse
import pickle
import numpy as np
from PIL import Image

app = FastAPI()
model = pickle.load(open("mnist_model.pkl", "rb"))

def preprocess_image(file_bytes) -> np.ndarray:
    # 1) Load into PIL, convert to grayscale 'L'
    img = Image.open(file_bytes).convert("L")
    # 2) Resize to 28×28 (use ANTIALIAS for quality)
    img = img.resize((28,28), Image.ANTIALIAS)
    # 3) Convert to numpy array (uint8), flatten to length-784
    arr = np.array(img).astype("uint8").reshape(1, -1)
    # 4) Optionally invert colors if your MNIST is white-on-black:
    #    arr = 255 - arr
    return arr

@app.post("/predict-image/")
async def predict_image(file: UploadFile = File(...)):
    # read the incoming UploadFile into BytesIO
    arr = preprocess_image(file.file)
    pred = model.predict(arr)[0]
    return JSONResponse({"prediction": int(pred)})