Generator architecture

[vict0rsch]

• Encoder: Resnet-based Content Encoder from MUNIT
  • image => 64 (=encoder.dim) channels with 1 conv layer, same size
  • conv-based downsamplings (encoder.n_downsample times)
  • resblocks (encoder.n_res blocks)
• Decoders: Resnet-based Decoders from MUNIT for all tasks but the translation
  • resblocks projections (decoder.n_res blocks)
  • Sequence of nn.Upsampling > Conv2dBlock (decoder.n_upsample times)
  • should match encoder.n_downsample
  • final conv to get a feature map with 1 (h, d, w), 3 (a) or 19 (s) channels
• Translation decoder: SPADEResnet-based Decoder inspired by MUNIT and SPADE
  • Conditioning the translation by SPADE([h, d, s, w])

[vict0rsch]

@adrienju there's a lot to review here, I'm not expecting you to go into details but if you could simply go over the updates to the code and tell me if anything seems weird to your eyes / what tests might be missing that'd be great