Colin Raffel

> Do you mean in the code or examples? Putting all the docs in the code might take some time and it might be better for the api to be...

> Given that l is a MergeLayer, l['output'] returns another layer, not the result of get_output_for. Oops, yes, this is what I meant (I think). > I think DictMergeLayer is...

> a) Add another MergeLayer base class (NamedMergeLayer or DictMergeLayer or something) that has a different interface and adapt lasagne.layers.get_output() accordingly? I think the current `MergeLayer` has semantics which are...

Sorry for missing this issue initially. > @skaae wanted to return the hidden state along with the output for recurrent layers (I don't exactly recall why, please help me out,...

> The most appealing option to me is also to have a single 'object' that represents the output of a layer, and maybe this object could be something other than...

I think it's a safe assumption - doesn't it say somewhere in the docs that layers are meant to operate on and produce Theano expressions? Plus, usually in machine leaning...

> Umm, I think there's no reason to do this, you'd put the DenseLayer inside the second RecurrentLayer then, with a CustomRecurrentLayer (otherwise it would be fully-connected to all inputs...

Yes, I think we're on the same page; I am not married to a particular implementation, but I would propose that the requirements are (at least): 1) Layers which do...

> I.e., each RecurrentLayer would be a MergeLayer that has two incoming layers: The layer before that, and either mask1 or mask2 (which are InputLayer instances). Yes, I see how...

> It's not intuitive for me either, having RecurrentLayer derive from MergeLayer sounds awkward. The sad thing is that it would sound much less awkward to me if it derived...