Default optimizer in Scholar.Linear.LogisticRegression

[krstopro]

It seems there is an issue with the training of Scholar.Linear.LogisticRegression; see here. I believe this is because SGD is used as a default optimizer. Setting optimizer: adam in fit/3 solves the problem. Perhaps we should change the default optimizer?

[josevalim]

Sounds good to me.

[krstopro]

The issue happens to be more complex than the default optimizer. Using Adam instead of SGD seems to have worked with the binary backend, not with EXLA. However, I am not able to investigate further as cannot install EXLA on my M1 Mac for some reason. Is this expected?

[krstopro]

Closing this issue as the problem doesn't seem to be in optimizer. See #301.