Kliff DNN torch trainer #185
Conversation
| @@ -82,33 +185,22 @@ def collate(self, batch: Any) -> dict: | |||
| """ | |||
| # get fingerprint and consistent properties | |||
| config_0, property_dict_0 = batch[0] | |||
| device = config_0.device | |||
| ptr = torch.tensor([0], dtype=torch.int64, device=device) | |||
| ptr = np.array([0], dtype=np.intc) | |||
There was a problem hiding this comment.
I think the collated data will be provided as the input to a torch NN model. Any reason why use np.array but torch.tensor for all variables?
There was a problem hiding this comment.
Update: NVM, now I understand that the batched data are passed to the descriptor in def _descriptor_eval_batch(self, batch) -> torch.Tensor, which requires numpy array.
| ) | ||
| self.torchscript_file = None | ||
| self.train_dataloader = None | ||
| self.validation_dataloader = None |
There was a problem hiding this comment.
Do we want to add test_dataloader and test the model at the end of training?
There was a problem hiding this comment.
Yes, I will add it uniformly across all trainers after finalizing this release. Idea is to have a base method test() that can run some tests on the test datasets. It will do simple energy and forces test but will not be limited to it, but could also leverage openkim tests if user requests.
|
Looks great! Merged. |
Summary
Added tests and trainer for training generic dense neural networks using libdescriptor and pytorch (as opposed to lightning).
TODO (if any)
Checklist
Before a pull request can be merged, the following items must be checked:
Note that the CI system will run all the above checks. But it will be much more efficient if you already fix most errors prior to submitting the PR.