Question in EX of Art_Generation_with_Neural_Style_Transfer

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Hey there,

generated_image = tf.Variable(tf.image.convert_image_dtype(content_image, tf.float32)) (just before the training loop).

I think this should be

generated_image = tf.Variable(tf.image.convert_image_dtype(generated_image, tf.float32))

Here, the generated_image was created at step 5.3 .

Otherwise, the init-value of training at very begin will be pure content_image based.

Slap me if I am wrong :wink:

2

Hi Chris.X,

You are right that starting with a noisy image may work better. But starting with the content image also works because of changes to the image to match the style.