C4W1, problems with low structural similarity

[adrienne]

Hi All,

In the neural style transfer programming assignment, I’m getting structural similarity of only .47. Even though it asks us not to change the settings, I tried experimenting and can increase ssim only a little by increasing style_weight to 2. Setting content_weight larger than it’s default setting 1e-32 decreases structural similarity. The generated image looks like a Kandinsky dog with some pixel artifacts. This assignment does not add tf.image.total_variation() to the loss, as was done in the colab. I tried adding it and get lower ssim.

Can someone help with this? Thank you!

[adrienne]

The problem was the assignment to variable content_and_style_layers must be in the order style_layers + content_layers, to extract the different features types from the model and compute the style and content losses.