fsherry / equivariant_image_recon Goto Github PK

Hi,

Great work, really enjoyed reading it, and thanks for open-sourcing your implementation! I have few questions regarding the implementation of equivariant proximal blocks. Thanks very much in advance.

(1) How did you go about picking the number of regular feature fields as 16 while defining the default feature type? (

) And, what is the intuition around what that multiplicity of 16 represents? Input and output feature types are fairly intuitive as we know the number of input channels, and n_memory.

(2) Did you use _set_equal_to_identity function while initializing the equivariant proximal blocks? If so, why was that necessary?

Thanks a lot again!
Beliz