Adapt losses to automatically use correct device

Question

Adapt losses to automatically use correct device

Closed this issue 2 years ago · 2 comments

Device-related errors are fixed in most places, but they remain a problem with the current loss implementation. The losses should know whether the model of a ParametricDR lives on CPU or GPU and move the .sub output accordingly.

Answer 1 · 2022-07-21T20:55:15.000Z

This is mostly done. There remains a RuntimeError when a model parameter is used inside a transform, but I couldn't figure out yet if this issue is connected with the loss modifications.

Answer 2 · 2022-07-21T22:44:26.000Z

Issue is fixed, the RuntimeError is caused by an in-place operation in the StudentTTransform.