Available Model Architectures
If you would like your architecture added, please submit a
pull request. Standard
Fourier Neural Operator implementation with support for deeper architectures (8 and 4 layers)
ResNet architectures using 8 residual blocks, no downscaling
ResNet where each block individually consists of 7 dilated CNN layers with dilation rates of [1, 2, 4, 8, 4, 2, 1], no downscaling
Original U-Net implementation
Our interpretation of original U-Net implementation without bottleneck layer and using
Modern U-Nets with Wide ResNet blocks, as used in various diffusion modeling applications
Modern U-Nets with [*] Fourier blocks in the downsampling path
Original U-shaped Neural Operator Implementation
Currently only scalar parameter conditioning is available.
Addition based conditioning in both spatial and spectral domain.
AdaGN style conditioning in the spatial domain.
AdaGN style conditioning in the spatial domain,
Addition in the spectral domain.
Model Architecture Registry Philosophy
While in principle we can make every architecture fully configurable via configuration files, we find it can affect the readability of the code quite a bit. Feel free to open issues or pull-requests for further configuration ability or any other suggestions for managing the configurability-readability tradeoffs.
Other Projects of Interest