[WIP] initial pass at convolutional layer visualization

[gallettilance]

    (X_train, y_train), (X_test, y_test) = mnist.load_data()

    X_train = X_train.reshape(X_train.shape[0], 28, 28, 1)
    X_train = X_train.astype('float32')
    X_train /= 255

    number_of_classes = 10
    Y_train = utils.to_categorical(y_train, number_of_classes)

    ACTIVATION = "relu"
    model = keras.models.Sequential()
    model.add(layers.Conv2D(32, 5, input_shape=(28, 28, 1), activation=ACTIVATION))
    model.add(layers.MaxPooling2D())
    model.add(layers.Conv2D(64, 5, activation=ACTIVATION))
    model.add(layers.MaxPooling2D())
    model.add(layers.Flatten())
    model.add(layers.Dense(100, activation=ACTIVATION))
    model.add(layers.Dense(10, activation="softmax"))
    model.compile(loss="categorical_crossentropy", metrics=['accuracy'])

    model.fit(X_train, Y_train, batch_size=100, epochs=5)

    dg = ConvGraph(model)
    X = [np.expand_dims(X_train[np.where(y_train == 0)[0][0]], axis=0)]
    dg.render(X, filename='test_input_mnist')

    X = []
    # get all the digit 0 examples
    for i in range(len(np.where(y_train == 0)[0])):
        X.append(np.expand_dims(X_train[np.where(y_train == 0)[0][i]], axis=0))
    dg.animate(X, filename='test_animate_mnist') 

Render of First Convolutional Layer for digit 0

Render of Second Convolutional Layer for digit 0

Animate of First Convolutional Layer of 50 digit 0 examples

Animate of Second Convolutional Layer of 50 digit 0 examples

Heatmap of First Convolutional Layer of all digit 0 examples

Heatmap of First Convolutional Layer of all digit 1 examples

Heatmap of First Convolutional Layer of all digit 2 examples

Heatmap of First Convolutional Layer of all digit 3 examples

Heatmap of First Convolutional Layer of all digit 4 examples

Heatmap of First Convolutional Layer of all digit 5 examples

Heatmap of First Convolutional Layer of all digit 6 examples

Heatmap of First Convolutional Layer of all digit 7 examples

Heatmap of First Convolutional Layer of all digit 8 examples

Heatmap of First Convolutional Layer of all digit 9 examples

[JamesKunstle]

Illustrations of layers are very interesting.
Consider changing output label to:
"Heatmap of First Convolutional Layer: all instances of '9' input"
for the sake of specificity. When I originally read it I misinterpreted it to mean "all" 9 classes, as opposed to a previous run with only 8 classes.

Also, in architecture for CNN you use 5x5 kernels. 3x3 kernels seems to be a more common choice but that seems up to interpretation.

For the sake of reproducibility on the part of the eventual end-user, I would consider passing input parameters via their local name: e.g.
layers.Conv2D( filters=32, kernel_size=5, strides=(1, 1), padding="same", ... )
This is easier to follow for new users if they're not familiar with the library API.

[JamesKunstle]

The rendering process is difficult to follow so I would suggest adding compartmentalized comments about what is happening at each stage. Additionally, I would write a README of the process of why the activations are rendering the interesting gifs that they are (as I currently understand the gif) and how the gif rendering process works. It would be really useful as a teaching tool to see a scope of the "why" you're rendering each of these gifs and then show the "how."