search

nengo / keras-lmu

Keras implementation of Legendre Memory Units
https://www.nengo.ai/keras-lmu/
Other
207 stars 35 forks source link

added memory_d>1 functionality #40

Closed NarsimhaChilkuri closed 3 years ago

NarsimhaChilkuri commented 3 years ago

LMUFFT should now be able to handle memory_d > 1, i.e, the input need not be projected down to a scalar before being fed into the delay layer.

arvoelke commented 3 years ago

Awesome! This will be very useful to have. The check inside of LMU that delegates to this implementation can also have the memory_d check removed. Would be good to have a unit test too. Could probably extend one of the ones already in here.

tbekolay commented 3 years ago

Thanks @NarsimhaChilkuri! Can you sign the CAA by adding yourself to this page?

hunse commented 3 years ago

OK, I've addressed your comments @drasmuss.

I also noticed that test_multivariate_lmu was failing on my machine, though it appears unrelated to these changes (it also failed on master for me); upping the tolerances a hair fixed it for me. (I tried explicitly setting all the dtypes to float32, since it seems like it might be some sort of typing issue, but that didn't fix it. :shrug:)