Refactor model architecture: extract backbone, activations, and configs into separate modules

[nimanzik]

[opencode-agent]

This issue requested refactoring the PyTorch MDN model architecture to improve modularity by extracting core components into separate, dedicated modules. The refactoring has been completed successfully.

What was done:

1. backbone.py - Extracted the MLPBackbone class, a configurable multi-layer perceptron for feature extraction with support for various normalisation types (batch/layer), dropout, and activation functions

2. activations.py - Created a dedicated module containing the ACTIVATION_FUNCTIONS registry mapping activation names to PyTorch classes, plus the custom ElevatedELU activation used for ensuring positive sigma outputs

3. configs.py - Built a Pydantic-based configuration system with MDNConfig for model architecture settings and mixup configurations (RandomMixupConfig, KDEMixupConfig) for type-safe setup

Benefits of this refactoring:

• Improved code organisation and maintainability
• Clear separation of concerns (model architecture vs. activation functions vs. configuration)
• Better reusability - MLPBackbone can now be used independently in other models
• Easier to extend or modify individual components without affecting others
• The main MixtureDensityNetwork in model.py now cleanly imports and composes these modules

opencode session  |  github run

[opencode-agent]

Not Found - https://docs.github.com/rest/reactions/reactions#list-reactions-for-an-issue-comment

opencode session  |  github run

[opencode-agent]

Not Found - https://docs.github.com/rest/reactions/reactions#create-reaction-for-an-issue-comment

github run