Anson Savage

Digipiction
This is the kind of thing I'd like to have built-in, a framework for machine learning. You wouldn't need to know the math behind it, it would come with big building blocks like In PyTorch and other frameworks, but exposed as Houdini nodes.

But maybe this is too far off of what Houdini is, and training your own networks might just not be much of a thing anymore because you can't keep up with the development in this area anyway. It might be more important to have the ability to somehow run existing models instead of building them yourself.

This is a fascinating idea.

Do you guys see a place for a Houdini version of Nuke's CopyCat? In 2D comp work, it makes a ton of sense. I'm trying to imagine how something like that tool might translate to 3D.

Hey!

Say I want to find all the parameters that reference $FSTART. How could I do it?

Thanks!
Anson

Ah, I think I found the answer. From the documentation:

"A handle to the geometry to write to. Currently the only valid value is 0 or geoself, which means the current geometry in a node. (This argument may be used in the future to allow writing to other geometries.)"