Open kuhanw opened 6 years ago
Hey all,
As a sanity check, I just tried to reproduce the twitter results from http://suriyadeepan.github.io/2016-12-31-practical-seq2seq/, and my training loss does not seem to decrease significantly just using the out of the box code. Do I have to do some specific tuning of the hyperparameters?
My training loss in the final epochs look like :
Model saved to disk at iteration #95000 val loss : 3.397206
Model saved to disk at iteration #96000 val loss : 3.424180
Model saved to disk at iteration #97000 val loss : 3.283931
Model saved to disk at iteration #98000 val loss : 3.356802
Model saved to disk at iteration #99000 val loss : 3.486829
I am using the out of box 03-Twitter-chatbot.py script.
Thank You,
Kuhan
when i at epoch 62401, the loss i got was 2.523555
Hi all,
I have been trying to duplicate the level of quality seen on the example page for the Cornell Movie dialogue and I am wondering how low have you guys driven the min-batch loss to get that type of replies? I am having trouble getting my decoder inference output to that level of comprehension.
I am also wondering for the CornellMovieDialogue corpus. Is the input going into the model as is? When I read the dialogue some of it doesn't even seem to make sense:
('Not the hacking and gagging and spitting part. Please.', "Okay... then how 'bout we try out some French cuisine. Saturday? Night?")
('Can we make this quick? Roxanne Korrine and Andrew Barrett are having an incredibly horrendous public break- up on the quad. Again.', "Well, I thought we'd start with pronunciation, if that's okay with you.")
as in, the answer is not referencing any information in the question.
As an additional note, I am also not exactly using the tutorials shown here, I have been piecing together tutorial scripts on seq2seq from our git for my own understanding so its always possible I have a bug in my code. Nonetheless, I do wonder about my question above.
Cheers,
Kuhan