Defining a train method will break the model.eval() call. about da-rnn HOT 5 CLOSED

I dont know actually if this will affect it anymore. I am using your base implementation and making it more usable for my software architecture right now. I fixed every bug i found but i dont remember them anymore. Maybe I can just create a merge request when i am done turning this into an OOP approach.

from da-rnn.

Hi @tomsch420 ,

Thank you for your advice, I did not notice this problem at the time I implemented this. Does this affect the repo, since it has been a long time I did not maintain this repo.

Thanks

from da-rnn.

That will be great! I will temporarily close this ticket, but you are always welcome to create a PR.

Thank you!

from da-rnn.

So, im done i guess. The network is converging for large amounts of data. But the predictions differ from run to run after loading the weights for the decoder and encoder. Is there some randomness involved that i missed?

from da-rnn.

How much do the predictions differ?
Will you share your OOP approach?

from da-rnn.