Issue:

Currently a pl.Trainer object can be initialized via from lightning import pytorch as pl, but also via import pytorch_lightning as pl. We use methods to get and set params in Chemprop. Unfortunately a Trainer Object is newly initialized upon calling set or update params. However, at this time an Accelerator is already instatiated, which leads to an issue as lightning requires a string. The "get device" function detects via isinstance wether a CPU or GPU accelerator was chosen and will transform this back into a string. The isinstance Method however only works for the lighning import but fails if pytorch_lighning is used.

Solution:

To prevent adding pytorch_lighning as a dependency we now enhanced the validation. If a user does not use lightning (but used pytorch_lighning instead) a corresponding value Error will now be raised.