plot_shap_top_explanations.py

def plot_shap_top_explanations(
    model: Model, 
    image: np.array, 
    class_names_mapping: Dict[int, str],
    top_preds_count: int = 3,
    fig_name: Optional[str] = None
) -> None:
    
    image_columns = 3
    image_rows = math.ceil(top_preds_count / image_columns)
    
    segments_slic = slic(image, n_segments=100, compactness=30, sigma=3)
    
    def _h(z):
        return model.predict(preprocess_input(mask_image(z, segments_slic, image, 255)))
    
    explainer = shap.KernelExplainer(_h, np.zeros((1,100)))
    shap_values = explainer.shap_values(np.ones((1,100)), nsamples=1000)
    
    preds = model.predict(np.expand_dims(image, axis=0))
    top_preds_indexes = np.flip(np.argsort(preds))[0,:top_preds_count]
    top_preds_values = preds.take(top_preds_indexes)
    top_preds_names = np.vectorize(lambda x: class_names[x])(top_preds_indexes)
    
    plt.style.use('dark_background')
    fig, axes = plt.subplots(image_rows, image_columns, figsize=(image_columns * 5, image_rows * 5))
    [ax.set_axis_off() for ax in axes.flat]
    
    max_val = np.max([np.max(np.abs(shap_values[i][:,:-1])) for i in range(len(shap_values))])
    color_map = get_colormap()
    
    for i, (index, value, name, ax) in \
        enumerate(zip(top_preds_indexes, top_preds_values, top_preds_names, axes.flat)):
    
        m = fill_segmentation(shap_values[index][0], segments_slic)
        subplot_title = "{}. class: {} pred: {:.3f}".format(i + 1, name, value)
        ax.imshow(image / 255)
        ax.imshow(m, cmap=color_map, vmin=-max_val, vmax=max_val)
        ax.set_title(subplot_title, pad=20)
        
    if fig_name:
        plt.savefig(fig_name)
    plt.show()