I just improved the docs for my Optimal Transport HDA (CPU default but GPU acceleration using CuPy if you have an NVIDIA card), thought I’d share the link since I haven’t posted it here before! github.com/conlen-b/cb-houdini-tools/tree/main/sinkhorn-based-log-domain-optimal-transport
The repo also has the raw python scripts that I heavily documented while creating it so feel free to poke around there to learn how sinkhorn based log domain optimal transport works (or to point out things I can improve!)
I originally made it for my implementation of @jake rice 😮 ‘s awesome Raya Druun paper, when I made it I didn’t realize that he had already posted an optimal transport implementation on [his github](github.com/jakericedesigns/Assorted-Houdini-Demos/tree/main/OpenCL%20Optimal%20Transport) so I reinvented the wheel a little bit 😅 but with CuPy it’s nice and fast (and can support double precision, by default I set it to float precision though) and above all helped me learn a lot!
Attachments in this post:
http://fx-td.com/houdiniandchill/wp-content/uploads/discord/20264107/30/26/OptimalTransportDemo_compressed.mp4