MaxPool2d preventing learning on BabyAI/MiniGrid about rl-starter-files HOT 6 CLOSED

Thank you Maxime for taking time to fill this issue.

Are you sure that the increase in performance not only due to the fact that removing the maxpooling increases the number of parameters?

Anyway, I will remove it

from rl-starter-files.

Yes, it seems we put too much max pooling. The input is tiny (7x7) so if we do max pooling twice, we're left with almost no signal.

from rl-starter-files.

I have just removed the MaxPool2D!

from rl-starter-files.

Finally, I have just reverted the modification. Removing the MaxPool2D will increase the number of parameters and so improve learning but it also makes learning much slower (for some networks, it multiplies the number of params by 100).

from rl-starter-files.

Could you clarify that on which games it's better to remove maxpool2d and on which games we should keep it?

from rl-starter-files.

I think it is generally better, it just takes much more time. But you are free to change the model as you want!

from rl-starter-files.