nengo / lmu

LMU metapackage
Other
4 stars 0 forks source link

Pytorch implementation #1

Open bratao opened 3 years ago

bratao commented 3 years ago

Are you planning in any near future an pytorch implementation? I'm super curious to try LMU in my torch projects!

arvoelke commented 3 years ago

Thanks for letting us know this is wanted! We are currently tracking pytorch support here: https://github.com/nengo/keras-lmu/issues/8 -- and in particular there is a pytorch version of the LMU here: https://github.com/bjkomer/pytorch-legendre-memory-unit but it does not yet support all of the bells-and-whistles, such as multiple memory vectors operating in parallel. The location and name of this repo may also change, but hope this helps as a start. If you have any specific issues or questions the Nengo forum may also be a helpful resource.

bratao commented 3 years ago

Hi @arvoelke I just tested the https://github.com/AbdouJaouhar/LMU-Legendre-Memory-Unit implementation for my academic research ( ultra long sequence labeling in Legal texts) and the results were fantastic. Slight better than my actual state of art, SRU. The https://github.com/bjkomer/pytorch-legendre-memory-unit unfortunately archived a much worse result.

I will be waiting anxiously for a more featured and maintained implementation 😍