mtenn.combination
Implementations for the Combination block in a GroupedModel.
The Combination block is responsible for combining multiple single-pose model
predictions into a single multi-pose prediction. For more details on the implementation
of these classes, see the Combination docs page and the guide on
Implementing a new Combination.
All equations referenced here correspond to those in Math for Implemented Combinations.
Classes
Abstract base class for the |
|
Approximate max/min of the predictions using the LogSumExp function for smoothness. |
|
Custom autograd function that will handle the gradient math for us for taking the max/min of the \(\mathrm{\Delta G}\) predictions. |
|
Combine a list of predictions by taking the mean. |
|
Custom autograd function that will handle the gradient math for us for combining \(\mathrm{\Delta G}\) predictions to their mean. |