Houdini 20.5 What’s new What’s new in Houdini 20.5

What’s new Machine Learning

Comparing the results from linear blend skinning and an ML model against the simulated ground truth.

The new example-based ML nodes allow you to create machine learning setups entirely from within Houdini. A large class of regression problems can be solved using a new, generic ML training node, without needing to write a training script.

The example-based ML toolset consists of two parts:

SOP tools for generating data sets, preprocessing data sets, writing data sets to disk, reading data sets from disk, inferencing, and performing nearest-neighbor queries.
a TOP training node designed to be used with the SOP tools.

The example-based ML nodes consist of:

ML Attribute Generate

Generate a set of packed primitives, each of which contains a geometry with a specific attribute randomized. Each of these can be an input for a data generation step.

ML Example

Create an example by combining an input component, for example generated by ML Attribute Generate with a target component, which may be the output of a procedural network for the input component.

ML Example Decompose

Decompose an example into its input component and its target component. This is useful for pre-processing examples, for example, using Principal Component Analysis.

ML Example Partition

Partition a collection of examples into parts (subcollections). This is useful for optimizing the generation and pre-processing of data, allowing memory use to be balanced against speed.

ML Extract Example

From a collection of examples, extract the single example at a specified index. This is useful when processing examples from within TOPs. This is also a useful node for visualization and troubleshooting.

ML Example Output

Write a collection of examples to a raw data file for training with ML Regression Train.

ML Regression Train

Train a model (feed-forward neural network) on a collection of examples previously written to disk using ML Example Output.

ML Regression Inference

Apply a model trained by ML Regression Train to one or more (previously unseen) inputs.

ML Regression Proximity

Use nearest-neighbor search to find the input component in a collection of examples closest to a given query input and return the corresponding target component. This is an example of ML that is not deep learning.

ML Example Import

Read from disk a collection of examples previously written using ML Example Output.

ML Pose Generate

This adapter node allows the example-based ML tools to work with animation. It generates a randomized set of rig poses, which can be inputs for a rig to apply machine learning to.

ML Pose Serialize

This adapter node allows the example-based ML tools to work with animation. It converts a rig pose into a point float attribute, for inclusion in a data set. This same node would also prepare inputs for ML Regression Inference.

Intermediate preprocessing stage for the ML Deformer: Mapping tet cages to skins.

Final preprocessing stage for the ML Deformer: Serializing poses and compressing displacements using PCA.

Generating synthetic training data using procedural geometry operations.

Training multiple neural networks, each with a different number of hidden layers.