bionemo-core 0.0.2

BioNeMo core interfaces and PyTorch-related code.

bionemo-core

Common code that all BioNeMo framework packages depend on. Contains highly reusable, battle-tested abstractions and implementations that are valuable across a wide variety of domains and applications.

Crucially, the bionemo-core Python package (namespace bionemo.core) depends on PyTorch and PyTorch Lightning. Other key BioNeMo component libraries, such as bionemo-llm and bionemo-geometric, obtain their PyTorch dependencies via bionemo-core.

Developer Setup

After following the setup specified in the README, you may install this project's code in your environment via executing:

pip install -e .

To run unit tests with code coverage, execute:

pytest -v --cov=bionemo --cov-report=term .

Package Highlights

In bionemo.core.model.config:

• ModelOutput: A Model's forward pass may produce a tensor, multiple tensors, or named tensors.
• LossType: A generic type parameter for a loss function.
• Model: An interface for any ML model that accepts and produces torch.Tensors.
• ModelType: A generic type parameter that is constrained to the Model interface.
• BionemoModelConfig: An abstract class that enables parameterizable model instantiation that is compatible with Megatron.
• BionemoTrainableModelConfig: An extension that includes the loss function to use with the model during training.

In bionemo.core.utils:

• the batching_utils module's pad_token_ids, which pads token ids with padding value & returns a mask.
• the dtype module's get_autocast_dtype, which converts from nemo/nvidia datatypes to their PyTorch equivalents.
• the random_utils module, which includes functions for managing random seeds and performing sampling.

In the bionemo.data package, there is:

• multi_epoch_dataset: contains many dataset implements that are useful for mutli-epoch training.
• resamplers: contains a P-RNG based Dataset implementation.

There's a constant global value, bionemo.core.BIONEMO_CACHE_DIR, which is used as a local on-disk cache for resources.