performances of rnn2.py

[elasryidriss]

Hi, I realy like your post on creating level caracter models on your blog. I tried to run doc-rnn2.py without the chekpoint callback, but the accuarcy on the validation in the 10th epoch still 49.74%. have you any suggesions to achieve the same performances as mentionned on your blog. Thanx

[maryamhonari]

Hi elasryidriss, The reported accuracy seemed to be from doc-cnn4.py model. I'm still curious why the the rnn2.py doesn't improve and keeps around 49%.

[offbit]

Hi all, I don't maintain the repository any more. It was just an instructional post.

The drop is a result of the version switch - keras from 1.0 -> 2.0 and tf 0.x -> 1.x . Probably some of the the graph building is being done differently, and the internal implementation of layers changed.

[fast-reflexes]

I experience that there is a sensitivity to weight initialization. I can run the model (rnn2) after the version switch but only occasionally does it actually learn something and then, the result is as indicated in the tutorial (+80% acc.). Other times, it just hovers around 50% epoch in and epoch out...