Next steps to improve SPADE + AdaIN architecture

[51N84D]

We've converged to a SPADE + MUNIT hybrid, but we need some tricks to improve. Here are the short term ideas to try:

• [x] Gram Matrix as measure of style
• [x] Feature matching loss on global discriminator
• [ ] Give style information directly to spade (via addition, for example)
• [x] Use decaying non-binary masks for smoothing
• [ ] Use image of only water for perceptual loss
• [x] Compute perceptual loss on intersection of ground + flood mask