Closed ixaxaar closed 6 years ago
Are you already diving into this, or do you need some help?
Not looking at this at the moment. Trying to make DNCs work for language modeling & seq2seq.
I've done DNCs for seq2seq before. Does anything about the base implementation need to change, or is it just a matter of structuring the inputs and outputs?
I actually wrote this repo because the other ones were not torch.nn
-like. Making the APIs more like torch.nn
makes it easier to replace nn.LSTM
etc with DNC
otherwise keeping implementations intact, potentially requiring one line of code change.
The ones I tried, I had to also write this piece of code for LMs to detach hidden states: https://gist.github.com/ixaxaar/02b65312eb1a840cc14e74d93b35178f For seq2seq I took the openNMT repo and forked it here.
Other than that it was pretty much a few liners.
Is implemented in #16 and #18 , though code review etc are pending.
Implement sparse access memory for DNCs. Related paper: https://arxiv.org/abs/1610.09027