lmjohns3

Interesting, thanks for the report. I'll try to look into it in the next couple weeks. Please feel free to send a PR to fix up the compression and indexing...

I've been absent from Github for some time, but am just starting to revisit these projects. I will make sure this one is transitioned to Py3 ASAP. Thanks!

I tried to implement this several months ago and got what I thought were strange errors; now that I look at it again it's pretty clear that the loss can't...

This can get pretty tricky with text encodings. My preference is to always operate with unicode, because then iterating over a string is guaranteed to iterate over a "letter" instead...

Hm, I've never seen that before. Can you post a snippet of the logging output from the training run?

I'm not sure I follow -- there's not really a way to set the iteration number in these models; the iteration number is only a way of keeping track of...

There is an option to use "true" BPTT with each recurrent layer in theanets (see https://github.com/lmjohns3/theanets/blob/master/theanets/layers/recurrent.py#L61): ``` net = theanets.recurrent.Classifier([ 20, dict(form='rnn', size=20, bptt_limit=5), 10]) ``` Then if you feed...

Yes, I wouldn't be surprised, theanets doesn't try to do any memory management at all, so it's up to Python/Theano to clean up things that have disappeared from the active...