module

module provides a collection of pre-defined modules for users to implement their own architectures.

MultiMolecule is built upon the ecosystem, embracing a similar design philosophy: Don’t Repeat Yourself. We follow the single model file policy where each model under the models package contains one and only one modeling.py file that describes the network design.

The module package is intended for simple, reusable modules that are consistent across multiple models. This approach minimizes code duplication and promotes clean, maintainable code.

Key Features

• Reusability: The module package includes components that are commonly used across different models, such as the SequencePredictionHead. This reduces redundancy and simplifies the development process.
• Consistency: By centralizing common modules, we ensure that updates and improvements are consistently applied across all models, enhancing reliability and performance.
• Flexibility: While modules such as transformer encoders are widely used, they often vary in implementation details (e.g., pre-norm vs. post-norm, different residual connection strategies). The module package focuses on simpler components, leaving complex, model-specific variations to be defined within each model’s modeling.py.

Modules

• heads: Contains various prediction heads, such as SequencePredictionHead, TokenPredictionHead, and ContactPredictionHead.
• embeddings: Contains various positional embeddings, such as SinusoidalEmbedding and RotaryEmbedding.