Rework how models are constructed and add support for models that have different architectures for handling complex, protein, and ligand.

What's Changed

• Fix torchdata pins for dgl by @kaminow in #82
• try empty conftest for cov by @hmacdope in #81
• Add split handling for ligand and protein representations by @kaminow in #78
• Initialize Concat network automatically by @kaminow in #77

Full Changelog: 0.6.3...0.7.0

What's Changed

• Adjust PKiReadout math for numerical stability by @kaminow in #71
• Add 3.12 to test matrix by @hmacdope in #75
• Migrate to pydantic 2.0+ by @hmacdope in #73

Full Changelog: 0.6.1...0.6.3

What's Changed

• Fix version and add test by @hmacdope in #69

Full Changelog: 0.6.0...0.6.2

What's Changed

• Fix version and add test by @hmacdope in #69

Full Changelog: 0.6.0...0.6.1

• Main (breaking) change of this release is the change of the Combination classes to return the per-pose predictions as well as the final combined prediction. This was done in order to connect these per-pose predictions back into the computation graph, so that they can be included in a loss function
• Fleshed out docs and comments a bit more

• More tests
• Fix small bug when passing Irreps to e3nn (#55)
• Add missing package to conda env (#57)

Some additions to the package, which should be fully backwards-compatible with v0.5.0:

• Docs (including RTD)
• Wrapper and config for the ViSNet model from torch_geometric
• New pKi Readout class

This release modifies the way that mtenn.conversion_utils models are built so that all can be built by directly passing args for the underlying model. Additionally, we introduce mtenn.config, which implements Pydantic schema that can be used to reproducibly mtenn models.

This version implements a major change in the way the GroupedModel and Combination classes are written so that they now work natively in pytorch while also not gobbling all GPU memory. Additionally, all the different types of classes were split into their own files, which will break some API calls.

Update to the PIC50Readout class that will invalidate models trained in prior releases. NOTE THAT ANY MODELS TRAINED WITH PREVIOUS RELEASES WILL NEED TO BE RETRAINED.

Also add some Combination classes.