AlkEthOH interaction-typing task

[maxentile]

Add datasets for tasks of classifying atom/bond/angle/torsion types for molecules in AlkEthOH rings set, and provide simple rule-based baselines for each task.

Done:

• Add script to download AlkEthOH rings dataset, label atom- and interaction-types using OpenFF 1.0.0 Parsley forcefield, and save discrete labels. (Tracked using Git LFS.)
• Add PyTorch Dataset interfaces to AlkEthOH{Atom|Bond|Angle|Torsion}TypesDatasets

Todo:

• Update paths from hfgp to espaloma
• Update tests, make sure resources can be found (currently using relative paths, should use pkg_resources.resource_filenames)
• Discuss with @yuanqing-wang whether PyTorch Dataset interface is suitable, make adjustments

[jchodera]

Add script to download AlkEthOH rings dataset

What's the rationale behind just using the rings? Limiting the number of compounds?

[maxentile]

What's the rationale behind just using the rings? Limiting the number of compounds?

No rationale, just a starting point -- intention is still to use also the AlkEthOH chains set and other sets listed in #2 (comment) and #2 (comment) .

[maxentile]

Hmm, although the Pytorch views in https://github.com/choderalab/espaloma/blob/973d5e1de00b60390b93a054c4277db632569b04/espaloma/data/alkethoh/pytorch_datasets.py satisfy the pytorch Dataset interface, they don't yet play nice with DataLoader.

For example,

import torch
dataset = AlkEthOHAtomTypesDataset()
loader = torch.utils.data.DataLoader(dataset, batch_size=10, shuffle=True)

runs into this error:

TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; found <class 'openforcefield.topology.molecule.Molecule'>

Possible workarounds:

• define a non-default collate function that handles OpenFF Molecules
• replace OpenFF Molecule with something that default_collate knows what to do with, such as a dict or a dgl graph containing similar information

[jchodera]

cc @jaimergp @t-kimber on the dataset issue above.