Tensorflow: visualize convolutional features (conv1) in Cifar10 model

tf_kernel_images.py

def kernel_images(name, x, pad=1, max_images=-1, summarize=True):
    """Create image summaries of 2d convolution kernel weights."""

    def factorization(n):
        """Calculates kernel grid dimensions."""
        for i in range(int(sqrt(float(n))), 0, -1):
            if n % i == 0:
                return i, n // i

    with tf.name_scope(name):
        # Normalize values and pad grid.
        low, high = tf.reduce_min(x), tf.reduce_max(x)
        x = (x - low) / (high - low)
        x = tf.pad(x, ((pad, pad), (pad, pad), (0, 0), (0, 0)))

        # Organize grid.
        r, c, chan_in, chan_out = x.get_shape()
        grid_r, grid_c = factorization(chan_out.value)
        x = tf.transpose(x, (3, 0, 1, 2))  # Move chan_out to front.
        x = tf.reshape(x, tf.stack([grid_c, r * grid_r, c, chan_in]))  # y
        x = tf.transpose(x, (0, 2, 1, 3))  # Switch x and y.
        x = tf.reshape(x, tf.stack([1, c * grid_c, r * grid_r, chan_in]))  # x
        x = tf.transpose(x, (0, 2, 1, 3))  # batch, height, width, channels

        # Last dimension may only contain a maximum of 4 channels.
        while x.get_shape()[3] > 4:
            a, b = tf.split(x, 2, axis=3)
            x = tf.concat([a, b], axis=0)

        # Last dimension may only contain 1, 3 or 4 channels.
        _, a, b, channels = x.get_shape()
        if channels == 2:
            x = tf.concat([x, tf.zeros((1, a, b, 1))], axis=3)

    # Summarize requested amount of images.
    if summarize:
        max_images = x.get_shape()[0] if max_images == -1 else max_images
        tf.summary.image('kernels', x, max_images)

    return x