Closed thomasbi1 closed 1 year ago
Good point, they should probably also be changed. There was no specific reason for this, just an oversight. The scaling experiments in the paper happen to be on image-only environments, so they aren't affected.
I see, thanks for clarifying! Do you know if the curves from the DMC Proprio tasks (with model size S) in the paper correspond to the models as in this implementation, or were the mlp_layers scaled for the paper?
It corresponds to this implementation, i.e. the MLPs for reward, continue, actor, and critic were scaled but vector encoder and vector decoder were not.
Hi
As far as I can tell, the different model sizes keep the number of
mlp_layers
andmlp_units
in the encoder/decoder constant and don't change them (as.*\.units
and.*\.layers
do not matchmlp_layers
andmlp_units
). Is there a reason for this? Is it because the MLP layers in the encoder/decoder don't affect the model size much?