Colin Raffel

I actually think I like the recurrent layer as merge layer variant more; the multiplex/demultiplex usage pattern is too different from what is used in the rest of Lasagne. If...

I think I like (B) a little more, because in retrospect I think the ideal choice would have been that all layers' `get_output_for`/`get_shape_for`/etc. are required to accept and produce an...

> One option is to try to generaize the GRU and LSTM layers to accept a hidden_to_output layer. See https://github.com/Lasagne/Lasagne/issues/357 One nice thing about only allowing a custom hidden->output connection...

> That code will use a different offset for every batch This shouldn't be an issue - the network should be able to learn the normalization too (right?). Maybe I'm...

> For the target vector, the offset will be about 1.0, since it's the sum of two random variables that each have an expected value of 0.5. Again, it will...

> Hmm, the example does not exactly follow any of these existing variants: Yes; the point of this example is not to show state-of-the-art results on an exactly-specified standard task,...

Hey Patrick, unfortunately I believe that because the mesh tf transformer uses tf.Estimator it is not eager-friendly. In the past when we've needed to do similar things, I'm sad to...

I think we should defer this for now, as it raises a larger question for Lasagne in terms of what the return format of a layer can be.

Thanks, any reason you think 0.6.0 is needed?

Thanks for noticing this. > removing them in `remove_invalid_notes`. I think this is the correct solution, maybe with a warning that a note with duration < resolution has been ignored.