add PassthroughScoring

[BenjaminBossan]

As mentioned here, I factored out the addition of PassthroughScoring from the PR #593

Here is a copy of the reasoning:

---

Currently, when adding new losses on a batch level, it's very fiddly to log them on an epoch level. You have to define a custom function that picks the right value from history and passes it through without modification, then add a BatchScoring callback that uses this mock scoring function, defines the name of the score again, and pass noop.

Here is the code before and after:

# somewhere in the net code
    ...
    self.history.record_batch('my_loss', my_loss)
    ...

# before
from skorch.callbacks import BatchScoring
from skorch.utils import noop

def my_loss_score(net, X=None, y=None):
    return net.history[-1, 'batches', -1, 'my_loss']

callbacks = [
    BatchScoring(
        my_loss_score,
        name='my_loss',
        target_extractor=noop,
    ),
]

# after
from skorch.callbacks import PassthroughScoring

callbacks = [
    PassthroughScoring(name='my_loss'),
]

---

I replaced BatchScoring with PassthroughScoring in the existing code for train and valid loss. Even though skorch.utils.train_loss_score and skorch.utils.valid_loss_score would not be needed anymore with this change, I still left them in place. This is for 2 reasons:

• someone might use them directly (after all, they're public functions)
• if they are removed, old pickled nets cannot be unpickled anymore

The latter could be fixed by introducing a special case in __getstate__ and replacing the old BatchScoring with the new PassthroughScoring which doesn't require these functions, but this seems to be far too much work for very little gain. Instead, I propose to leave those functions for now, maybe deprecate them at some future time.

I'm not quite certain what to do with test_net_initialized_with_initalized_dataset. I'll think of something, but if anyone has an idea, I'll be glad to hear it.

[ottonemo]

LGTM

[ottonemo]

Suggested change

	for each batch and determines the average of this score, which is
	for each batch (the train loss, for example) and determines the average of this score, which is

Maybe this makes it easier to relate how and when this callback is useful?

[BenjaminBossan]

fixed

[ottonemo]

Do we need an issue to track this?

[BenjaminBossan]

tbh I don't remember why I put this here but the test xfails with the expected error, so I removed the comment and skip