Results question

[sulljohn]

Hi rosinality,

Great project!

1. Were you able to get results for other structures in the paper like buildings or other structures? I was testing training the model on other images than human faces and was having mixed results. For example, on one of the tests I did, it just reproduced the exact same images in the place where it said they would be swapped in generator.py.
2. Do you train on just the two images you are trying to swap the textures of or an entire dataset? I have tried both approaches but was having no luck getting results like in the paper with my dataset.
3. Is there somewhere in the code that you set the weights assigned to the two encoders? I was interested if they could be adjusted to favor either the texture or structure encoder for the image.

Thank you!

[rosinality]

1. I haven't tried it. I will consider the datasets like LSUN church.
2. During training the batch will be splitted into the half and swap occurs between these two. So swap occurs over an entire dataset.
3. I haven't used weight assignments for structure-texture split. This split is driven by structural contraints on the networks and texture swapping during the traing.