app.py

import gradio as gr
import tensorflow as tf
import numpy as np
from PIL import Image

# Load the trained model
# Note: Ensure "rice_stage_model.keras" is in the exact same directory as this script
model = tf.keras.models.load_model("rice_stage_model.keras")

# Define class labels mapping to your model's outputs
classes = ["flowering", "germination", "noise", "tillering"]

# Set the confidence threshold to prevent false positives
CONFIDENCE_THRESHOLD = 0.6

def predict_stage(image):
    if image is None:
        return "⚠️ Please upload an image first."
        
    # 1. Convert to RGB to drop alpha channels (crucial for PNG compatibility)
    image = image.convert("RGB")
    
    # 2. Resize to match the model's expected input shape
    image = image.resize((224, 224))

    # 3. Preprocess the image array (normalize pixel values)
    img = np.array(image) / 255.0
    img = np.expand_dims(img, axis=0)

    # 4. Run prediction
    predictions = model.predict(img)[0]

    # 5. Extract results
    predicted_index = np.argmax(predictions)
    confidence = float(predictions[predicted_index])
    predicted_label = classes[predicted_index]

    # 6. Handle edge cases (noise class or low confidence)
    if predicted_label == "noise":
        return "⚠️ Invalid image. Please upload a clear rice crop image."

    if confidence < CONFIDENCE_THRESHOLD:
        return "⚠️ Model uncertain. Please upload a clearer crop photo."

    # 7. Return the final formatted string
    return f"🌾 Predicted Stage: {predicted_label.capitalize()} | Confidence: {confidence:.2f}"

# Set up the Gradio interface
# Set up the Gradio interface
interface = gr.Interface(
    fn=predict_stage,
    inputs=gr.Image(type="pil"),
    outputs="text",
    title="Rice Crop Growth Stage Classifier",
    description="Upload a picture of the rice crop from the field to instantly detect its growth stage.",
    flagging_mode="never" # Updated from allow_flagging for Gradio 4 compatibility
)

if __name__ == "__main__":
    # Launching directly prevents Gradio's buggy auto-reloader from crashing the app
    interface.launch(
        server_name="0.0.0.0",
        server_port=7860,
        ssr_mode=False,
        show_error=True
    )