cameron-d/CNN_Intermediate_Activations_Visualizer

1import keras2from keras import layers3import gradio as gr4import numpy as np5from PIL import Image6import io7import contextlib8 9# Load model10model = keras.models.load_model("dogs_and_cats_CNN.keras")11 12def get_model_summary(model):13    """Return the model summary as a string."""14    stream = io.StringIO()15    with contextlib.redirect_stdout(stream):16        model.summary()17    summary_str = stream.getvalue()18    return summary_str19 20def get_img_array(image, target_size):21    """Resize the image and return it as an array."""22    image = image.resize(target_size)23    array = keras.utils.img_to_array(image)24    array = np.expand_dims(array, axis=0)25    return array26 27def predict(image):28    img_tensor = get_img_array(image, target_size=(180, 180))29 30    # predict class31    predictions = model.predict(img_tensor)32    if predictions[0][0] > 0.5:33        predicted_class = "Dog"34        confidence = predictions[0][0]35    else:36        predicted_class = "Cat"37        confidence = 1 - predictions[0][0]38 39    prediction_text = f"## **Prediction:** {predicted_class}         **Confidence:** {confidence:.2%}"40    41    42    # Collect convolution and pooling layers43    layer_outputs = []44    layer_names = []45    for layer in model.layers:46        if isinstance(layer, (layers.Conv2D, layers.MaxPooling2D)):47            # If a layer of a convolution or max pooling layers, append it's outputs to the visualization48            layer_outputs.append(layer.output)49            layer_names.append(layer.name)50    51    activation_model = keras.Model(inputs=model.input, outputs=layer_outputs)52    activations = activation_model.predict(img_tensor)53 54    # Build visualization grids for each layer55    images = []56    images_per_row = 1657 58    for layer_name, layer_activation in zip(layer_names, activations):59        n_features = layer_activation.shape[-1]60        size = layer_activation.shape[1]61        n_cols = max(1, n_features // images_per_row)62        display_grid = np.zeros(63            ((size + 1) * n_cols - 1, images_per_row * (size + 1) - 1)64        )65 66        for col in range(n_cols):67            for row in range(images_per_row):68                channel_index = col * images_per_row + row69                if channel_index >= n_features:70                    break71                channel_image = layer_activation[0, :, :, channel_index].copy()72                if channel_image.std() > 1e-6:73                    channel_image -= channel_image.mean()74                    channel_image /= channel_image.std()75                    channel_image *= 6476                    channel_image += 12877                channel_image = np.clip(channel_image, 0, 255).astype("uint8")78                display_grid[79                    col * (size + 1):(col + 1) * size + col,80                    row * (size + 1):(row + 1) * size + row,81                ] = channel_image82 83        display_grid = display_grid / 255.084        images.append((display_grid, layer_name))85 86    summary_text = get_model_summary(model)87    return images, summary_text, prediction_text88 89 90 91# Gradio interface with examples92with gr.Blocks() as demo:93    gr.Markdown("# CNN Intermediate Activations Visualizer")94    gr.Markdown("Visualizes activations of all convolutional and pooling layers and displays the model summary.")95    gr.Markdown("Model is trained on a subset of kaggle's dogs vs cats dataset: https://www.kaggle.com/c/dogs-vs-cats/data")96    gr.Markdown("Adapted from: https://deeplearningwithpython.io/chapters/chapter10_interpreting-what-convnets-learn/#visualizing-intermediate-activations")97    98    with gr.Row():99        with gr.Column():100            input_image = gr.Image(type="pil", label="Upload an image")101            submit_btn = gr.Button("Analyze")102 103            # Example images104            gr.Examples(105                examples=[106                    ["images/cat_1.jpg"],107                    ["images/dog.jpg"],108                    ["images/cat_2.jpg"],109                    ["images/cat_and_dog.jpg"]110                ],111                inputs=input_image,112                label="Try an example:"113            )114        115        with gr.Column():116            output_gallery = gr.Gallery(label="Layer Activations", show_label=True, columns=1)117            output_prediction = gr.Markdown(label="Prediction")118 119            gr.Markdown("As you go deeper through the neural network, the activations become more abstract and relate more to the class prediction")120 121            output_summary = gr.Textbox(label="Model Summary", lines=20)122    123    124    submit_btn.click(125        fn=predict,126        inputs=input_image,127        outputs=[output_gallery, output_summary, output_prediction]128    )129 130demo.launch()131