Replace sampler balancing by loss balancing

This MR removes balancing of the training sampler and replaces it by loss balancing, if the loss function allows it (= contains a pos_weight parameter). Binary and exclusive, multi-class targets are supported. Using a database with multi-class, non-exclusive targets will raise a NotImplementedError for now.

Models now expect the class and arguments of the loss used instead of an instance, just like it is done with the optimizer. It is instantiated in trainer.py::run() and balanced if possible.

A base Model class has been created that other models should inherit from.

closes #6 (closed)

@dcarron: thanks for this useful set of changes. I left some minor comments, but this looks great already.

Thanks for the feedback, I'll simplify things and use a single loss.

added 1 commit

• 6f04a219 - [model] Use a single type of loss for train and validation

Compare with previous version

added 1 commit

• 9aff0292 - [logging] Prevent PL from adding dataloader index to logged metrics

Compare with previous version

@dcarron: thanks for this useful set of changes. I left some minor comments, but this looks great already.

added 1 commit

• 918a86bf - [trainer] Hardcode name of loss to monitor

Compare with previous version

@dcarron: thanks for the reaction - are we ready?

requested review from @andre.anjos

approved this merge request

assigned to @dcarron

mentioned in issue #6 (closed)

Also, it would be interesting to rebase before the merge, I guess.

resolved all threads

changed the description