Reproducibility/setup issues: FCD install breaks RDKit/Torch/CUDA, missing QM9 data/configs, unclear experiment steps #1
Description
Hi,
I'm trying to reproduce your code and ran into major trouble with the fcd dependency and some missing files:
- There is no mention of how to install or set up fcd in the repo README or requirements.
- When I try pip install fcd, it tries to install its own versions of RDKit and Torch, which then breaks my CUDA setup and makes the environment incompatible with the rest of the requirements (torch, torch_geometric, etc).
- I tried various combinations: installing fcd before/after torch/rdkit, using --no-deps, but it always ends up messing up versions or missing CUDA support.
- Eventually, I found fcd_torch (link), but the codebase expects import fcd, not import fcd_torch, so this also breaks without symlinks or hacking site-packages.
Questions:
- What is the recommended way to install the correct, CUDA-compatible fcd version for this repo, without breaking RDKit/Torch versions?
- Should we use fcd_torch? If so, is there a reason the code still imports fcd?
- Could you clarify the intended FCD setup or update the documentation to help avoid these environment problems?
Related questions:
- I also noticed the QM9 dataset/config files are missing from the repo (or at least not directly available). Is there an official source, or a script to process QM9 as expected by your code?
- Are the configuration files meant to match the DiGress paper, or is there anything special I should know about running your provided experiments?
- Is there any other “gotcha” or undocumented step I need to know to fully reproduce your results?
Sorry if this is obvious, but I couldn’t find clear instructions on this and I’m sure others will run into it as well.
This would really help, since I’ve lost a lot of time on environment rebuilds due to these conflicts.
Thanks!