Implement saving and loading of models

[frostedoyster]

---

📚 Documentation preview 📚: https://metatensor-models--19.org.readthedocs.build/en/19/

[PicoCentauri]

Very nice export. I am not sure where we want the export to be called. Should this be controlled by the train script or the trainer?

[PicoCentauri]

I think we don't have to set hypers to a default value or is there a specific reason?

[frostedoyster]

In this way, it's much easier to use the model from Python. For example, the test I've written for the model_io module would have been much more difficult. In general, I feel like, if we want to support model use from Python as well, we need to make the hypers available from Python as well

[PicoCentauri]

Okay I get your point but then maybe it makes sense to provide a general parser and store the hypers as dictionary in the init of each architecture.

[frostedoyster]

I feel like this is short enough for the moment... Perhaps we can change it once we have more models?

[frostedoyster]

EDIT: I now understand the question

I would say at least in part the trainer (because we need to checkpoint). The last model save (at the end of training) can be done either by the trainer or by train.py. I would say by train.py, so that we can have that line in one file instead of in all the trainers. Moreover, some trainers might not support checkpointing (thinking of linear algebra)

[PicoCentauri]

I would say at least in part the trainer (because we need to checkpoint). The last model save (at the end of training) can be done either by the trainer or by train.py. I would say by train.py, so that we can have that line in one file instead of in all the trainers. Moreover, some trainers might not support checkpointing (thinking of linear algebra)

I am more and more thinking that we should create a class for training with a train method and save method. With this we can do checkpoints within and in the end train.py just calls save again. What do you say?

[PicoCentauri]

I have a minor idea that we should integrate the save_model maybe as a class method. Othwerwise it is very nice!