app.py

import pickle

import cv2
import joblib
import numpy as np
import streamlit as st
import torch
from streamlit_drawable_canvas import st_canvas

BINARY = joblib.load("model.joblib")
ML_MODEL = pickle.loads(BINARY)
ML_MODEL.load_state_dict(
    torch.load("model_weights.pth", map_location=torch.device("cpu"))
)
ML_MODEL.eval()


def predict_number(img):
    if img is None:
        return None, None

    inp = torch.tensor(img, dtype=torch.float32).unsqueeze(0).unsqueeze(0) / 255.0

    with torch.no_grad():
        output = ML_MODEL(inp)

    probabilities = torch.nn.functional.softmax(output[0], dim=0)
    values, indices = torch.topk(probabilities, 5)
    confidences = {f"is number {i.item()}": v.item() for i, v in zip(indices, values)}

    return confidences


canvas_result = st_canvas(
    fill_color="rgba(255, 165, 0, 0.3)",
    stroke_width=17,
    stroke_color="#000000",
    background_color="#ffffff",
    background_image=None,
    update_streamlit=True,
    height=200,
    width=200,
    drawing_mode="freedraw",
    key="canvas",
)

if canvas_result.image_data is not None:
    image_data = canvas_result.image_data[:, :, 0]
    image_data = np.squeeze(image_data)
    image_data = cv2.blur(image_data, (10, 10))
    image_data = cv2.resize(image_data, (28, 28))
    image_data = cv2.bitwise_not(image_data)
    st.image(image_data)
    confidences = predict_number(image_data)
    st.write(confidences)