Building Machine Learning Pipelines | Houdini 21 | Michiel Hagedoorn

Learn how to build supervised ML pipelines in Houdini 21. Explore the ML nodes and features that have been added with Houdini 21. These ML nodes are designed to both leverage and enhance Houdini's proceduralism with fully automated end-to-end ML pipelines that include data generation, training and inference.

Houdini's ML nodes can be used to perform sub-tasks that are common to a variety of machine learning applications. This can save you a lot of work when you're creating your own machine learning pipeline. Among other things, the ML nodes support example generation, preprocessing, saving & loading data sets, neural-network training, and inference. We show how the ML nodes were used to build ML applications that are being released as part of Houdini 21. These applications include a volume upresser that can be applied to pyro sims and an improved character deformer.

Michiel Hagedoorn is a Senior 3D Scientist at SideFX Software. He currently works on ML tools and various ML-related projects. His other work at SideFX includes computational geometry, soft-body dynamics and character FX. Michiel has a Ph.D. in Computer Science from Utrecht University and was a post-doc researcher at the Max Planck Institute for Informatics. His research areas included computer vision and proximity search. In addition to that, Michiel was a game-engine developer at Digital Extremes, where he worked on real-time physics and game AI. Michiel has been passionate about math, programming and computer graphics from an early age.

00:00:00 Introduction
00:03:32 Regression & Toy Examples
00:07:05 Labeled Examples Overview
00:10:32 Data Preprocessing
00:14:02 Working with Unlabeled Data
00:17:26 Traditional vs Neural Techniques
00:20:48 Neural Network Training
00:24:03 Code Snippet Highlights
00:27:27 Correction Examples
00:31:10 Volume Upscaling