sparsh007/ExplanableAI

1import os2import keras3from keras.applications import inception_v3 as inc_net4from keras.preprocessing import image5from skimage.segmentation import mark_boundaries6import numpy as np7import matplotlib.pyplot as plt8import gradio as gr9from lime import lime_image10 11# Load the pre-trained InceptionV3 model12inet_model = inc_net.InceptionV3()13 14def transform_img_fn(img_path):15    """Preprocess image for InceptionV3"""16    img = image.load_img(img_path, target_size=(299, 299))17    x = image.img_to_array(img)18    x = np.expand_dims(x, axis=0)19    return inc_net.preprocess_input(x)20 21def explain_image(img_path):22    """Generate LIME explanation and visualization"""23    # Preprocess image24    processed_img = transform_img_fn(img_path)25    26    # Create LIME explainer27    explainer = lime_image.LimeImageExplainer()28    29    # Generate explanation30    explanation = explainer.explain_instance(31        processed_img[0].astype('double'), 32        inet_model.predict, 33        top_labels=5, 34        hide_color=0, 35        num_samples=100036    )37    38    # Get image and mask39    temp, mask = explanation.get_image_and_mask(40        explanation.top_labels[0],41        positive_only=False,42        num_features=10,43        hide_rest=False44    )45    46    # Get top 5 predictions47    predictions = inet_model.predict(processed_img)48    top_5_indices = np.argsort(predictions[0])[-5:][::-1]49    top_5_labels = [inc_net.decode_predictions(predictions, top=5)[0][i][1] for i in range(5)]50    top_5_probs = [inc_net.decode_predictions(predictions, top=5)[0][i][2] for i in range(5)]51    52    # Create visualization53    fig, ax = plt.subplots(figsize=(6, 6))54    55    # Explanation visualization56    ax.imshow(mark_boundaries(temp / 2 + 0.5, mask))57    ax.set_title('Pros (Green) vs Cons (Red)')58    ax.axis('off')59    60    plt.tight_layout()61    62    # Create a string for the top 5 predictions63    predictions_str = "Top 5 Predictions:\n"64    for i, (label, prob) in enumerate(zip(top_5_labels, top_5_probs)):65        predictions_str += f"{i+1}. {label}: {prob:.4f}\n"66    67    # Generate heatmap68    ind = explanation.top_labels[0]69    dict_heatmap = dict(explanation.local_exp[ind])70    heatmap = np.vectorize(dict_heatmap.get)(explanation.segments)71    72    # Plot heatmap73    fig_heatmap, ax_heatmap = plt.subplots(figsize=(6, 6))74    heatmap_plot = ax_heatmap.imshow(heatmap, cmap='RdBu', vmin=-heatmap.max(), vmax=heatmap.max())75    plt.colorbar(heatmap_plot, ax=ax_heatmap)76    ax_heatmap.set_title('Heatmap Explanation')77    ax_heatmap.axis('off')78    79    plt.tight_layout()80    81    return fig, predictions_str, fig_heatmap82 83# Create Gradio interface84demo = gr.Interface(85    fn=explain_image,86    inputs=gr.Image(type="filepath", label="Input Image"),87    outputs=[88        gr.Plot(label="Explanation"),89        gr.Textbox(label="Top 5 Predictions"),90        gr.Plot(label="Heatmap Explanation")91    ],92    title="LIME Image Classifier Explainer",93    description="Upload an image to see which areas positively (green) and negatively (red) influence the classification, the top 5 predictions, and a heatmap explanation."94)95 96# Launch the app97if __name__ == "__main__":98    demo.launch()99