Closed einbandi closed 2 years ago
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.
Issue is fixed, the RuntimeError
is caused by an in-place operation in the StudentTTransform
.
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.