Houdini 21.0 Nodes Geometry nodes

ML Regression Linear geometry node

Determine an output based on a linear model that optimally fits given labeled examples.

On this page

Overview

This node trains a linear model that minimizes the sum of squared distances between the prediction and the target at each of the provided examples, with an added quadratic regularization term. This linear model is applied to each query input to produce a predicted output. ML Regression Linear provides an alternative to inferencing a trained neural network. In situations where a linear model is a good enough approximation, this node provides a simpler alternative to training and inferencing a neural network.

The first input of ML Regression Linear should consist of labeled examples. For example, created using ML Example inside a SOP for loop or read in from a file using ML Example Import. The second input should be a component with same contributions as each input component of the labeled examples. The input/output contributionsby the multi parameters should match the input/target contributions on the ML nodes from the origin labeled examples.

When ML Train Regression and ML Regression Inference is preferred, ML Regression Linear can be useful for verifying the data set was correctly generated. If you give an input component of a labeled example of the data set, ML Regression Linear should return an exact match for the target. This assumes no two labeled examples in the data set have the same input component.

See Machine Learning documentation for more general information.

Parameters

Batch

In Single mode, a single input is evaluated. In Multiple Packed mode, the model is evaluated on each embedded geometry of the input and the results are stored in corresponding packed primitives.

Weight Decay

The model trained by this node minimizes a quadratic loss. Weight Decay is a coefficient that scales a sum of squared weights (excluding bias parameters) that is a summand of this loss. Increasing the Weight Decay may improve the generalization of the model, while sacrificing how closely it fits the example targets. Increasing the value of this parameter may also help you get a more stable solution in cases that are difficult to train.

Error Threshold

This is a relative error threshold to detect cases where the solution is not accurate enough. If the error of the computed model exceeds this threshold, then the node will display an error. This threshold doesn’t do anything except generating errors when it is exceeded. This is to avoid surprises when an accurate model cannot be computed.

Number of Inputs

Type

Type of input contribution: either a point attribute or a volume.

Point Attribute

Name of a point float attribute.

Volume Name

The name of a volume.

Volume Resolution

Resolution of the volume along the X,Y,Z axis.

Tuple Size

Tuple size of the point attribute or volume.

Number of Outputs

Type

Type of target contribution: either a point attribute or a volume.

Point Attribute

Name of a point float attribute.

Volume Name

The name of a volume.

Volume Resolution

#id targetvolumeresolution Resolution of the volume along the X,Y,Z axis.

Tuple Size

Tuple size of the point attribute or volume.

Inputs

Labeled Examples

The labeled examples whose input components will be compared to the query input component.

Input Component

The query input component.

Outputs

Output Component

The output component that corresponds to the labeled example that matches the query input component.

Geometry nodes