app.py

"""
IM 417 Classification
"""
import cv2
import numpy as np
import os
import streamlit as st
import tempfile
from streamlit_drawable_canvas import st_canvas
from tensorflow import keras
from huggingface_hub import from_pretrained_keras
from transformers import AutoModel

# Load the model (using a Streamlit caching mechanism for efficiency)
#@st.cache(allow_output_mutation=True)  # Allow model mutation for prediction
def load_model():    
    try:
        model = from_pretrained_keras("tomeheya/IM-417-128x128-classification", force_download=True)
        #model = AutoModel.from_pretrained("tomeheya/IM-417-128x128-classification")
        print("Model loaded successfully!")
        return model
    except Exception as e:
        print(f"Error loading model: {e}")
        return None

# Function to preprocess the input image
def preprocess_image(image):
    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    image = cv2.bitwise_not(image)
    image = np.expand_dims(image, axis=-1)  # Add channel dimension
    image = image / 255.0  # Normalize pixel values
    return image

# Function to perform inference
def predict(image):
    model = load_model()
    if not model:
        return "Model couldn't be loaded !!!"
    preprocessed_image = preprocess_image(image)
    preprocessed_image = np.expand_dims(preprocessed_image, axis=0)  # Add batch dimension
    prediction = model.predict(preprocessed_image)
    predicted_label = decode_prediction(prediction)
    return predicted_label

# Function to decode the model prediction
def decode_prediction(prediction):
    # Decode the prediction based on your model's output format
    # For example, if your model outputs class probabilities:
    class_idx = np.argmax(prediction)
    predicted_label = "IM-417 sign number :: " + str(class_idx + 1)
    return predicted_label

# Initialize the canvas
canvas_result = st_canvas(
    fill_color="rgba(255, 165, 0, 0.3)",
    stroke_width=3,
    stroke_color="rgba(0, 0, 0, 1)",
    background_color="rgba(255, 255, 255, 1)",
    height=128,
    width=128,
    update_streamlit=True,
    key="canvas"
)

# Add a submit button
if st.button("Submit"):
    if canvas_result.json_data:
        image_from_canvas = canvas_result.image_data
        st.text(f"Original data: {len(image_from_canvas)}")
        preprocessed_image = preprocess_image(image_from_canvas)
        predicted_label = predict(preprocessed_image)
        st.text(f"Predicted label: {predicted_label}")
    else:
        st.text("Please draw an image on the canvas before submitting.")

# Streamlit app layout
st.title("IM 417 Classification")