Document intermediate value debugging in FAQ

[ottonemo]

I think this is a topic that is relevant to many users and we can document the multiple output case there as well. See #428 as well.

[thomasjpfan]

There seems to be two big use cases for multiple outputs:

1. The multIple outputs are used to calculate the loss.
2. In #428 where one of the outputs is used to view layers of the model.

For case 1, we explore some for the possible ways to handle this in issue https://github.com/dnouri/skorch/issues/422.

For case 2, I prefer the use of the forward_hooks. In any case, this would require the user to keep track of some state.

[ottonemo]

For case 2 we could provide callbacks that do the bookkeeping for the user, e.g.

net = Net(...,
    callbacks=[
        EpochLogger('fc', lambda mod, in, out: out),
    ]
)

# ... fit/predict ...

dict(net.callbacks_)['EpochLogger'].values

The name is obviously a matter of discussion, as is the callback itself.

[BenjaminBossan]

Regarding said callback: We could also have a more general function signature that takes the callback and the other arguments as input. This would allow things like this:

# function that just appends the values, as suggested above; could be default
def func(callback, mod, in, out):
    callback.values_.append(out)

# function that logs to visdom
def func(callback, mod, in, out):
    do_visdom_things(out)

(btw., in is not a valid variable name ;) )