-
Notifications
You must be signed in to change notification settings - Fork 70
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Torchdrug integration #49
Conversation
Open for review and discussion about pytorch-scatter/torchdrug installation workflow @drugilsberg! |
Final updates on this PR @drugilsberg:
I'm glad I managed to find fixes for all these issues, but the mere wall of issues I encountered for doing this PR should make it evident that the torchdrug integration raises some concerns regarding the overall code reliability 😐 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Great job some general comments we need to address before merging:
- I would not have configuration classes for specific datasets. Since the classes are the same we should simply use versions for the different datatsets (see comment in PR).
sentencepiece
treatment. Instead of doing the trick every time I would do it once insrc/gt4sd/__init__.py
. After the definition of__version__
and__name__
we can have the following:
import sentencepiece as _sentencepiece # noqa: F401
import pytorch_lightning as _pl # noqa: F401
In this way we only have it there and inside the module we can import pytorch_lightning without any issue.
- I left a note on the installation.
Once those are addressed I would say we are good to go.
I addressed all comments @drugilsberg, but unfortunately your solution with importing |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks good great job, just few comments remained.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks good great job, just few comments remained.
src/gt4sd/training_pipelines/pytorch_lightning/language_modeling/core.py
Outdated
Show resolved
Hide resolved
src/gt4sd/training_pipelines/pytorch_lightning/language_modeling/models.py
Outdated
Show resolved
Hide resolved
updating comment.
updating comment.
updating comment.
updating comment.
…g/models.py updating comment.
…g/core.py updating comment.
updating comment.
updating comment.
Signed-off-by: Matteo Manica <drugilsberg@gmail.com>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks great well done
TorchDrug integration (inference-only) as discussed @drugilsberg!
2 models implemented
GCPN
(Graph Convolutional Policy Network, NeurIPS 2018) andGAF
(Graph AutoregressiveFlow, ICLR 2020).For each models, 3 pretrained models are made available:
TorchDrugZincGCPN
andTorchDrugZincGAF
TorchDrugQedGAF
andTorchDrugQedGCPN
.TorchDrugPlogpGCPN
andTorchDrugPlogpGAF
unittests implemented for all models
ToDo:
Regarding CI:
I knew this would backfire, the tests are still failing at installation. Locally, I had to install
pytorch-scatter
via conda, I did the following:and this worked fine.
When I reproduce the CI pipeline locally, I also do not get a functioning installation. The env is created successfully but importing
pytorch-scatter
results in segmentation faults.Possible solution: Adapt the
conda.yml
like so: