Simple ML digit recognition in SideFX Houdini using PyTorch and Numpy

• Houdini 19.0.383 py3
• Python 3.7.4
• PyTorch 1.10.0
• Numpy 1.21.3
• CUDA 11.3.1
• cudNN 11.3

The digit data gets generated in SOPs within Houdini and doesn't rely on the classical MNIST dataset. There is no real advantage using this approach other than experimenting with procedural data generation and the 'Houdini to PyTorch and back' interface. The model isn't trained very well (too few and too similar inputs) since that wasn't the objective in this experiment.

All of the scripts are implemented internally in Houdini's python nodes. The .py files are just for quick code access without having to open Houdini.

There are probably a lot of rookie mistakes and many ways to do things smarter and more efficiently. Keep that in mind when digging into it! =)

Edmond Boulet-Gilly's video helped me figure out lots of things!