Tim Cooijmans

@rizar's solution works in this case but in general I would prefer @lukemetz's proposal. If you didn't have control over MultiBrick's constructor you would still want to be able to...

Sure, that works, but I'm talking in general. Initialization should be configurable e.g. from a YAML file with hyperparameters. It shouldn't require special cases in code.

This keeps coming back to bite me. I think not overriding children's initialization configuration as suggested by @lukemetz is the obviously right thing to do. It would silently change the...

I suppose the `update` attribute might be useful here, but a complicating factor is that at each timestep a different shared variable or a different subtensor of the shared variable...

I confirm this happens on my laptop as well, also 32-bit Ubuntu, but with libhdf5 from Ubuntu. The segfault seems to occur when control leaves H5PYDataset.create_split_array.

This is a bit of a random comment, but the biases introduced here will be redundant; LSTM takes its biases from the inputs. However, the improvement reported in the paper...

I use it and it seems to work, but I don't have a simple experiment to back it up and I may be confounding it with the effect of identity...

Yes, this is something we've noticed as well. We're experimenting with two possibilities: 1. Pad with repetitions of the data rather than zeros 2. Normalize the input `x` sequence-wise as...

> Padding with repetitions of data is an interesting idea. However, I do feel that would throw off the network's performance as it would have to learn to "throw out"...

> You said that you started it with this idea, but you don't think its most optimal. Why do you feel this way? It seems like it should work. The...