Add an interface to PET #68
Conversation
frostedoyster
requested review from
frostedoyster,
Luthaf,
PicoCentauri and
abmazitov
| EPOCH_NUM_ATOMIC: 1000000000000000000 | ||
| SCHEDULER_STEP_SIZE_ATOMIC: 500000000 | ||
| EPOCHS_WARMUP_ATOMIC: 250000000 |
There was a problem hiding this comment.
Do we really need these hypers if in practice you never change them?
Also, we need a page explaining them as for the other architectures.
There was a problem hiding this comment.
We could do this, depending on how keen @serfg is. However, this should not be required for experimental models
There was a problem hiding this comment.
No it is not required but maybe makes sense to have this is bit cleaned up.
src/metatensor/models/experimental/pet/utils/systems_to_batch_dict.py
Outdated
Show resolved
Hide resolved
| if len(validation_datasets) != 1: | ||
| raise ValueError("PET only supports a single validation dataset") | ||
|
|
||
| if device_str == "gpu": |
There was a problem hiding this comment.
this should never arrive to a model. IMO we should convert the device to a torch device before passing it to the train script
There was a problem hiding this comment.
You're probably right... or maybe a list of devices. We can open an issue for this, because it concerns all the models
src/metatensor/models/experimental/pet/tests/test_torchscript.py
Outdated
Show resolved
Hide resolved
Integration of PET into metatensor-models. This PR is intended to have a proper PET wrapper that accepts the unified data structure of metatensor-models and predicts energies and forces. So that one can run molecular dynamics with pre-trained PET models when this PR is ready.
📚 Documentation preview 📚: https://metatensor-models--68.org.readthedocs.build/en/68/