A to(device) method on the high-level interface

[sfriedowitz]

Describe the workflow you want to enable

Hello,

I've been using this package to power a personal algo-trading project, thanks for the great work!

I would find it incredibly helpful to add a to(device) on the top-level TabPFNRegressor or TabPFNClassifier classes. I see that it is possible to move the architecture and associated components across device by a save/load from disk, but this is cumbersome and raises exceptions for non-fitted models. Adding this high-level method would be incredibly useful and aid in usage of the package.

Thanks!

Describe your proposed solution

On both TabPFNRegressor and TabPFNClassifier, add a to(device) or a to_torch_device method or something similar that handles moving all weights/torch-components of the model to device.

Describe alternatives you've considered, if relevant

No response

Additional context

No response

Impact

None

[brendan-priorlabs]

Hi @sfriedowitz, this isn't exactly what you're asking for, but I'm curious, does this setting help at all?

TabPFN/src/tabpfn/classifier.py

Line 233 in ebc1346

device:

If not, what are the limitations you're hitting?

[sfriedowitz]

Hi @brendan-priorlabs , as far as I understand this setting is only passed when instantiating the estimator for the first time.

But I am looking to instantiate the estimator on one device and then move it to another. From what I can tell this is only possible by saving/loading the model from disk using the utility methods (which raise an exception if the model has not yet been fitted).

[oscarkey]

Hi @sfriedowitz, this is a great feature suggestion! Unfortunately we don't have capacity internally to work on it currently, but I will leave this issue open and we will get to it in the future.

Alternatively, we would welcome a PR from the community that adds it!