Compatibility with the latest release of rdkit (2023.03.1) has been fixed (also fixes #11) . This problem was caused by to the way that descriptors were being computed - importing all available ones instead of fixing a set of them. A new descriptorAvgIpc was introduced in the last release breaking all expected model shapes with the default model.
Other fixes:
The default MolSkill model (trained on GPU) could not be used in machines without a GPU.
the models/ folder has been removed from the repo - we now solely rely on the default models available on FigShare.
pinned pytorch_lightning to 1.6.3 due to breaking changes in seeding. I plan to relax these requirements in a future PR.
Cleaning of several unused functions
A release will follow after this PR to fix the conda package, which will also be broken for new installations that include the newer rdkit version.
EDIT: For some reason, conda is currently unable to correctly suggest the py39 package for Python 3.9 and instead falls back to py310 unless explicily pointed to a build number. I am fixing this in the README.md for the time being while I look for a solution.
Compatibility with the latest release of rdkit (
2023.03.1) has been fixed (also fixes #11) . This problem was caused by to the way that descriptors were being computed - importing all available ones instead of fixing a set of them. A new descriptorAvgIpcwas introduced in the last release breaking all expected model shapes with the default model.Other fixes:
models/folder has been removed from the repo - we now solely rely on the default models available on FigShare.pytorch_lightningto 1.6.3 due to breaking changes in seeding. I plan to relax these requirements in a future PR.A release will follow after this PR to fix the conda package, which will also be broken for new installations that include the newer rdkit version.
EDIT: For some reason, conda is currently unable to correctly suggest the
py39package for Python 3.9 and instead falls back topy310unless explicily pointed to a build number. I am fixing this in theREADME.mdfor the time being while I look for a solution.