Open billbrod opened 1 year ago
Another possibility: get style from Portilla-Simoncelli texture stats, content from deep net. Or content could just be some simple convolutional filters?
I think the torchvision
create_feature_extraction is the way to do this: https://pytorch.org/vision/stable/feature_extraction.html
Would be nice to have an example showing how to use
Metamer
for style transfer, using some standard deep nets. Would also serve as an example of an extension ofMetamer
for other purposes.@jenellefeather pointed out that it's not straightforward to grab the activation of intermediate layers from the standard deep nets in the pytorch model zoo. can see how she did it in her project. Might be worth giving discussion of how to do that a separate tutorial.