RNN LSTM support about juice HOT 9 CLOSED

@drahnr - We're still waiting on native LSTM support.

(Comment deletion above was spam)

from juice.

I'd be happy to review your contributions! LSTM is still WIP so contributions are very welcome!

from juice.

Looks like no one is working on this right now, but I'm developing an LSTM for my own project with this lib. I'm wondering what other challenges people have faced with this.

from juice.

I tried to implement LSTM layer about 2 years ago, but ran out of time and keeping up with cudnn releases slowed progress down significantly.

I'd be very happy to include your existing layer which can then be improved upon / ported to other backends step by step.

from juice.

#87 #79 #64 should solve this issue as well as #10 for the cuda backend, so there is still space for contributions in the native backend @johnkarasev - the heavy lifting is already done by @lissahyacinth

from juice.

The GBMV stuff (#84 ) is designed to support pointwise vector multiplications, which are an important part of the RNN.

from juice.

My apologies - I should have named you (@paulkass ) in the same sentence - just thought the interface part to be more relevant as an entrypoint for the remaining impl, while they certainly will build their impl on the work you've done already! Again, I am sorry.

from juice.

@paulkass @lissahyacinth is there anything left todo? Or can we close this?

from juice.

I think as long as one backend implements it, we can check this of as covered. Worst case people have a slower experience, but the functionality is there (even if it currently requires nvidia hardware). We should make the feature matrix more public, but that's a different topic.

from juice.