Outputs constant: _EOS

[Lldenaurois]

Hi there,

I implemented your code in Tensroflow r1.1 and I am able to train the entire model.

When I then attempt to get a sample, I simply get a constant output [ 2, 2, 2, 2, 2 ]

This means that the trained model outputs _EOS every time.

Any ideas as to why this is happening?

[domerin0]

Hello,

How long did you train it for? This is something I had before. I found it occurs during early training stages. I found the EOS token to disappear (mostly) after about 24 hours on a Titan X. Though I suspect to train a 'good' model requires significantly more training.

[Lldenaurois]

Hey man,

Thanks for the response. I trained for less than 12 hours.

I wonder if maybe there's a tiny bug in your code, where you maybe end each conversation with a newline. This would lead the model to have these examples where outputting the empty sentence was most likely. I can look into it!

Great work on the repo though!

[domerin0]

That's very possible there is a bug like that. If you want to look into it and find a bug, feel free to submit a pr.