Visualize convolutional kernels with Tensorflow 2.0

kernel_visualization.py

import numpy as np
import tensorflow as tf

# Layer name to inspect
layer_name = 'block3_conv1'

epochs = 100
step_size = 1.
filter_index = 0

# Create a connection between the input and the target layer
model = tf.keras.applications.vgg16.VGG16(weights='imagenet', include_top=True)
submodel = tf.keras.models.Model([model.inputs[0]], [model.get_layer(layer_name).output])

# Initiate random noise
input_img_data = np.random.random((1, 224, 224, 3))
input_img_data = (input_img_data - 0.5) * 20 + 128.

# Cast random noise from np.float64 to tf.float32 Variable
input_img_data = tf.Variable(tf.cast(input_img_data, tf.float32))

# Iterate gradient ascents
for _ in range(epochs):
    with tf.GradientTape() as tape:
        outputs = submodel(input_img_data)
        loss_value = tf.reduce_mean(outputs[:, :, :, filter_index])
    grads = tape.gradient(loss_value, input_img_data)
    normalized_grads = grads / (tf.sqrt(tf.reduce_mean(tf.square(grads))) + 1e-5)
    input_img_data.assign_add(normalized_grads * step_size)

Excuse my stupid question, but how do we actually see/save the image(s)?

At the end of the loop, input_img_data is a 4D tensor holding the generated image. What you want to do is convert it to a numpy with .numpy() and visualizing it with matplotlib for example

Is this the expected outcome in this case?

For matplotlib to perform well, you want either to normalize values between 0 and 1, or convert the image to int. As the warnings says, it's clipping values from 0-255 range to 0-1 which makes the image so poor

@konradsemsch I solved it by converting input_img_data this way:

input_img_data = input_img_data.numpy().astype(np.uint8)