Can you put back the train algo?
vedmathai opened this issue · 1 comments
I have been using this library for six months now and must say it is really useful. Good job. However, I noticed you had included the training file in v1 but now removed it. I am really interested in training my own sequences and even the p2g side. I am able to use git history to get the earlier version. But was hoping if you could add it back in case there are any improvements.
Ved
Hi Ved,
Thanks for the sweet feedback.
At this update, I used pytorch instead of TF because it was easier to convert tensor into numpy arrays. See this for the training code: https://github.com/Kyubyong/nlp_made_easy/blob/master/PyTorch%20seq2seq%20template%20based%20on%20the%20g2p%20task.ipynb
Earlier I used a seq2seq framework with attention, but I changed it to a simpler seq2seq, which sets the last state of the encoder as the initial state of the decoder. That's simply for simplicity.