Release single-step model environments #15
Conversation
There was a problem hiding this comment.
LGTM!
Just one question, l noticed that there is an environment.yml file in the root directory of Syntheseus. When should users install this file, and when should users install the environment_shared.yml file located in reaction_prediction/environments directory ?
Some clarifications in the Readme might be helpful?
When trying to set up just the core of We should think how/where to best explain this to the users though (perhaps easier when we also add the model checkpoints). |
There seem to be two issues with `setup_megan.sh` that evaded the testing done in #15: - Two necessary dependencies (`gitpython` and `scipy`) are missing - The version of `torchtext` used forces `pip` to switch to a different version of `torch` than the one used in the shared environment This PR addresses both problems.
Continuing from previous PRs (most notably #15), this PR releases trained model checkpoints for all currently integrated baselines.
This PR releases instructions for building an appropriate environment for each of the supported single-step models. I've managed to unify all models apart from GLN into a shared base (containing mostly just
python,torchandrdkit) which each model type extends with a handful of model-specific dependencies. One last piece still missing are checkpoint files for each model type.For now, the setup instructions live under
reaction_prediction/environments. Longer term we could consider forking the external single-step model repositories and adding simplepip-installability to them, which could allow us to streamline the setup steps further through installation extras, e.g. be able to do something likepip install syntheseus[retro-knn]. Currently the dependencies in the environments are mostly pinned for reproducibility, but if we go with the installation extras approach then we can unpin them wherever possible.