app.py

import gradio as gr
import numpy as np
import tensorflow as tf
from tensorflow.keras.models import load_model
from PIL import Image

# -------------------------------
# Load Model and Data
# -------------------------------
MODEL_PATH = "satellite_segmentation_unet.h5"
X_TEST_PATH = "X_test.npy"
Y_TEST_PATH = "y_test.npy"
TOTAL_CLASSES_PATH = "total_classes.npy"

print("Loading model and test data...")
model = load_model(MODEL_PATH, compile=False)
X_test = np.load(X_TEST_PATH)
y_test = np.load(Y_TEST_PATH)
total_classes = int(np.load(TOTAL_CLASSES_PATH)[0])

# -------------------------------
# Class color map
# -------------------------------
CLASS_COLORS = {
    0: [226, 169, 41],   # water
    1: [132, 41, 246],   # land
    2: [110, 193, 228],  # road
    3: [60, 16, 152],    # building
    4: [254, 221, 58],   # vegetation
    5: [155, 155, 155],  # unlabeled
}

# -------------------------------
# Helper Functions
# -------------------------------
def decode_segmentation(mask):
    """Convert model output or label mask to RGB."""
    if mask.ndim == 4:  # Model output shape (1, H, W, num_classes)
        mask = np.argmax(mask[0], axis=-1)
    elif mask.ndim == 3 and mask.shape[-1] == total_classes:
        mask = np.argmax(mask, axis=-1)
    elif mask.ndim == 3:
        mask = mask.squeeze()
    color_mask = np.zeros((mask.shape[0], mask.shape[1], 3), dtype=np.uint8)
    for class_idx, color in CLASS_COLORS.items():
        color_mask[mask == class_idx] = color
    return color_mask

def predict_segmentation(index):
    """Run model inference and return all visualizations."""
    try:
        index = int(index)
        image = (X_test[index] * 255).astype(np.uint8)

        # Model prediction
        pred = model.predict(np.expand_dims(X_test[index], axis=0))
        pred_mask = decode_segmentation(pred)

        # Ground truth
        gt_mask = decode_segmentation(y_test[index])

        return image, gt_mask, pred_mask

    except Exception as e:
        print("Error during prediction:", e)
        return None, None, None

# -------------------------------
# Gradio Interface
# -------------------------------
title = "🛰️ Satellite Segmentation (U-Net)"
description = """
Select one of the **test images** from the dataset to visualize:
- The original input image  
- The ground truth mask  
- The predicted segmentation mask from the trained **U-Net**
"""

indices = list(range(min(50, len(X_test))))

interface = gr.Interface(
    fn=predict_segmentation,
    inputs=gr.Dropdown(
        choices=[str(i) for i in indices],
        label="Select Test Image Index",
        value="0",
        interactive=True,
    ),
    outputs=[
        gr.Image(label="Original Image"),
        gr.Image(label="Ground Truth Mask"),
        gr.Image(label="Predicted Mask"),
    ],
    title=title,
    description=description,
    allow_flagging="never",
    theme="gradio/soft",
)

# -------------------------------
# Launch
# -------------------------------
if __name__ == "__main__":
    interface.launch(server_name="0.0.0.0", server_port=7860)