Split up layers/base.py

[lasagne]

Split up layers/base.py

[benanne]

Looks good to me. I would call the first ? 'dense.py' and the second one 'shape.py' or something. Since these names are not really going to be exposed to the user, they're not really that crucial to get 'right' imo. Short is good :)

On that note, we should probably still make a distinction between the 'base' set of layers, that are available in the nn.layers namespace, and the more specialized stuff that is only accessible through the submodules (like nn.layers.cuda_convnet, nn.layers.corrmm).

I believe all the layers you listed are probably important enough to live in the main namespace, but we'll have to keep this in mind and make a decision when we add in the recurrent layers for example.

[craffel]

I think splitting up base.py is definitely a good idea, but I'd propose that we give access to all of the layer classes straight from nntools.layers (e.g. DropoutLayer would be nntools.layers.DropoutLayer instead of nntools.layers.noise.DropoutLayer) - I don't think there will be any namespace conflicts (e.g. something in some layer submodule which is not in another), and I think it will save a lot of typing. I think this is what you just proposed @benanne; I guess what I'm saying is I think all layers are probably important enough to live in the main namespace.

[benanne]

Yep, that's the idea. But for specialized stuff like the cuda_convnet and cormm layers, I don't think it's necessary. Are you saying you think these should also be available in the main namespace?

[craffel]

I think that as long as there aren't any namespace conflicts (e.g. a layer with the same name in cuda_convnet and conv), we may as well put them all in the same namespace.

[benanne]

I was thinking about doing this quickly, but clearly there are a lot of dependencies between future submodules. How should these be handled?

It seems cleanest to just import those submodules where necessary and use them directly. But this will make the code a bit harder to read, because it is less like 'normal' code that a user would write (i.e. a user would write layers.DenseLayer, or just DenseLayer, but never dense.DenseLayer).

The alternative is to import * the necessary submodules (since that is what happens in __init__.py anyway).

What's the best option? We should probably pick one of these and do that consistently across the library.

[f0k]

I think the best option is yet another way: Importing just the names that are needed, to be explicit about the dependencies. E.g., if the recurrence module depends on dense.DenseLayer, it should do from .dense import DenseLayer. This way the code still reads like 'normal' code, as you called it, and we avoid the namespace mess created by import *.

[benanne]

Yes! Let's do that :)

[benanne]

I'll try to sort this out today.

[benanne]

This was done in #81.