search

Joshua-Ren / Neural_Iterated_Learning

Pytorch implementation of the paper 'Compositional language emerge in a neural iterated learning' (ICLR 2020).
MIT License
13 stars 2 forks source link

Data is updated once every 4 iterations during PhaseB #3

Open lavoiems opened 3 years ago

lavoiems commented 3 years ago

See: https://github.com/Joshua-Ren/Neural_Iterated_Learning/blob/master/train.py#L222.

It is common to update the data every iterations. Since I don't remember seeing it being discussed in the paper, I am wondering if this is intended. If so, what is the rational behind not updating the data every iteration?

Thank you.

Joshua-Ren commented 3 years ago

Hi Sam,

Thanks very much for your question. I am sorry for not mentioning this small trick in the appendix (oh actually I delete this in the final version). This trick seems can stabilize the training during the ablation study, i.e., only Alice (or Bob) is updated when the learning rate is small. (It can let the agent learn the mapping better.) But I think you can safely delete this line without harming the performance. Because later I found that the training can still converge as long as the learning rate is properly chosen.

Cheers,

Joshua

Sam @.***> 于2021年4月6日周二 上午3:56写道：

See: https://github.com/Joshua-Ren/Neural_Iterated_Learning/blob/master/train.py#L222 .

It is common to update the data every iterations. Since I don't remember seeing it being discussed in the paper, I am wondering if this is intended. If so, what is the rational behind not updating the data every iteration?

Thank you.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Joshua-Ren/Neural_Iterated_Learning/issues/3, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGGLYAMCCZRVMKNYFHUH543THIIYXANCNFSM42NLUOOQ .