Is it possible to use peephole connections for LSTMs in PyTorch?
I found this discussion: https://github.com/pytorch/pytorch/issues/630 so it seems like it hasn’t been done yet.
Someone on there said they made a CPU version, but said that it was 4 times slower than a GPU version?
There was also this pull request that was never merged: https://github.com/pytorch/pytorch/pull/1665
Is it worth for me to use their code if I wanted to use peephole connections or just re-implement it myself? Or should I not even bother with trying to use peephole LSTMs at all and use some other kind of LSTMs? I wanted to use peepholes in the first place because I wanted to replicate some old Theano code.