Always read files in double precision #294
Conversation
| devices: List[torch.device], | ||
| train_datasets: List[Union[Dataset, torch.utils.data.Subset]], | ||
| val_datasets: List[Union[Dataset, torch.utils.data.Subset]], | ||
| checkpoint_dir: str, | ||
| ): | ||
| dtype = train_datasets[0][0]["system"].positions.dtype | ||
| assert dtype in AlchemicalModel.__supported_dtypes__ |
There was a problem hiding this comment.
I think this can go outside of the trainers, in our infrastructure code rather than in the model code (same for the other architectures)
There was a problem hiding this comment.
We already do this outside. I just added this is in case one wants to debug the model without going through the whole metatrain infrastructure.
Yes, for me it is also weird. The only reason I see is that it may not work in 64 bit.... |
|
The alchemical tests made no sense in the first place... Not sure why we were parametrizing both float32 and float64 when we wanted to check exact reproducibility with the same result. I'm fixing them |
frostedoyster
force-pushed
the
reader-64bit
branch
from
5abec02
to
660a7de
Compare
frostedoyster
force-pushed
the
reader-64bit
branch
from
660a7de
to
10e80f5
Compare
frostedoyster
requested review from
DavideTisi,
spozdn and
abmazitov
as code owners
As discussed with @frostedoyster it seems to be good idea to read all information from files in double precision and let the model decide when to convert them to the user requested precision. This is in particular useful when subtracting composition energies. Usually, composition energies are very large and one may end up with round-off errors when not doing these operations in the highest precision.
I also extended the
mypylinter to all to be linted files and I also enabled all features of the sphinx-linterContributor (creator of pull-request) checklist
[ ] Issue referenced (for PRs that solve an issue)?📚 Documentation preview 📚: https://metatrain--294.org.readthedocs.build/en/294/