Potential bug in reader.py

[simulkade]

First of all, wonderful tools!
I have difficulty getting chebai-graph (the latest version) to work with the latest version of chebifier. After some investigation manually and with Claude Code, I ended up with a potential bug in the reader. It seems like the issue goes all the way back to this, which means changing the two occurences of _sanitize_molecule(mol) to mol = _sanitize_molecule(mol) must fix the issue. I tested it locally and it worked!

[aditya0by0]

cc: @sfluegel05

[sfluegel05]

Sorry for the late reply. This is indeed a bug, probably stemming from the previous sanitation function which worked inplace while the new one doesn't.

@simulkade Thanks a lot for identifying this issue.

I took the opportunity to also move the molecule processing to a central repository to ensure consistent results for all repositories involved in chebifier.

[simulkade]

Thanks @sfluegel05 . I am still struggling to get the entire ecosystem to work together especially the graph-ai with chebifier. Do you have a workflow when you set things up in you dev environment that I can follow? And do you usually prefer an installation from source or installing the latest releases? I run into issues with both :-)

[aditya0by0]

Thanks @sfluegel05 . I am still struggling to get the entire ecosystem to work together especially the graph-ai with chebifier. Do you have a workflow when you set things up in you dev environment that I can follow? And do you usually prefer an installation from source or installing the latest releases? I run into issues with both :-)

@sfluegel05, I have a suggestion regarding this, we could add an docker file for chebifier to automate the installation of required dependencies and to run the application.

[simulkade]

Thanks @sfluegel05 . I am still struggling to get the entire ecosystem to work together especially the graph-ai with chebifier. Do you have a workflow when you set things up in you dev environment that I can follow? And do you usually prefer an installation from source or installing the latest releases? I run into issues with both :-)

@sfluegel05, I have a suggestion regarding this, we could add an docker file for chebifier to automate the installation of required dependencies and to run the application.

@aditya0by0 Sounds good for using the code. However, I personally would like to look under the hood sometimes, for which an installation workflow (or a development workflow) is invaluable.

[sfluegel05]

Hi @simulkade. The chebifier dependency management is indeed a bit messy right now, sorry for that. Usually, I would recommend installing from the latest release. However, since the last chebifier release, we have made some updates to chebai and chebai-graph which are not compatible with chebifier (I am currently working on a new chebifier version, but that is not yet ready for release).

I have installed a fresh environment with Python 3.12. If you follow these commands, you should get a working chebifier instance:

uv venv --python 3.12
source .venv/bin/activate
uv pip install chebifier==1.2.1 chebai_graph==1.0.0 chebai==1.2.0 chemlog_extra==1.0.1 c3p
uv pip install torch==2.12.0 torch_scatter torch_geometric -f https://data.pyg.org/whl/torch-2.12.0+cpu.html
uv pip install jsonargparse[signatures]==4.27.7 scikit-learn
python -m chebifier predict -e ensemble.yml -s "CCO"

(ensemble.yml refers to this file, c3p does not work on windows machines)

I will also publish a new chebifier release that includes these exact version in the pyproject.toml for an easier installation.

For reference, I have attached the pip freeze for my environment.

chebifier_dependencies.txt

[sfluegel05]

Update: I have released chebifier v1.2.2. This reduces the commands needed to

uv venv --python 3.12
source .venv/bin/activate
uv pip install chebifier[models]
uv pip install torch==2.12.0 torch_scatter torch_geometric -f https://data.pyg.org/whl/torch-2.12.0+cpu.html
python -m chebifier predict -e ensemble.yml -s "CCO"

[simulkade]

It is wonderful, Simon @sfluegel05 . Thanks a lot.