Matthew

Just my $0.02, but I think the design could be improved a lot if Layer was generalised to multiple inputs and multiple outputs, just like theano's graph. You'd avoid special-case...

> But just because we haven't agreed on multi-output layers yet. It doesn't require get_output() to be moved back into the Layer class. OK -- I struggled to implement a...

@fok OK, yeah I see what you mean re get_output. If you wanted to discourage that use, perhaps you could make it a private method _get_output, then provide a `get_output(*layers)`...

Ah OK, I see you discussed this already :) If you made Layer._get_output private and steer people towards a public API for get_output on multiple layers, seems that would be...

> Layer._get_output() would need to pass around those dictionaries that hold the output expressions of all other layers. This would be harder to understand A matter of taste perhaps --...

If you accept (and document) that as a self-imposed limitation then I guess fair enough, but seems like it hurts the generality of any layers that are parameterised by subgraphs....

Hi all, I just ran into a really nasty-to-track-down bug in xarray (version 2023.08.0, apologies if this is fixed since) where non-nanosecond datetimes are creeping in via expand_dims. Look at...

Hi David That's great, thanks!

I'm afraid we don't provide a training script in this codebase as our training setup is quite tied to internal infrastructure, but I think we give enough to construct one...