Basic configuration

Scimba package configuration.

At the loading of the scimba_torch module, torch is configured as follows:

>>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
>>> torch.set_default_device(device)
>>> torch.set_default_dtype(torch.double)

This configuration can be printed with:

>>> scimba_torch.print_torch_setting()

To change device and or default floating point arithmetic precision:

>>> torch.set_default_dtype(torch.float32)
>>> torch.set_default_device("mps")

Notice that natural gradient preconditionning used in scimba_torch does not work well with simple floating point precision.

At this time, basic routines of linear algebra required in natural gradient descent preconditioning are not implemented in torch for “mps” (a.k.a. Metal Performance Shaders) and we discourage its use.