Colin Raffel

> To reiterate, I'd propose changing the recurrent layer's get_output_for() to something like: This seems reasonable to me, and arguably a little cleaner than returning multiple things from `get_output` as...

Cool, please assign to me.

A snag: `lasagne.utils.unroll_scan` doesn't support updates. We can either modify it to support updates, (@skaae any ideas?) or we can come up with a way to check whether updates are...

Ah I see, ok. Well, I will add a test for the `unroll_scan` case too to ensure that it works the same when there are updates.

tbh, I'm a bit overwhelmed by the changes here -- @stevenxxiu can you post a few usage examples and/or an overview of the usage of the new classes so I...

Sorry, I'm confused about something fundamental - none of your `InputLayer`s in any of your examples have a "number of time steps" dimension. This makes sense for cells, but not...

> Right, seems he forgot when posting here, the tests have them: Ah, ok, I see that. After spending a bit looking over the code, I think I have a...

> So I'd suggest to have a look at the end-user API first: Does it cover all the use cases, in readable end-user code? I think the best way to...

I think the attention network also requires some additional machinery that might not be covered there, because it would need to use the output of two recurrent processes (`h` and...

Thanks for making that! Some questions (some of these could be cleared up by documentation/comments - if you'd rather go that way that'd be great!) 1. https://github.com/stevenxxiu/lasagne_recurrence_examples/blob/master/primitives.py#L22 So outputs from...