Open wzn0828 opened 8 months ago
Hello,
The observation dropout is only used to drop the input observation fed to the posterior. There is no need to additionally dropout when feeding the state to the policy.
Hello,
The observation dropout is only used to drop the input observation fed to the posterior. There is no need to additionally dropout when feeding the state to the policy.
However, In the screenshot, you always use the posterior sample to decoder action and BEV. That's to say, you do not use the dropout when action decoding and BEV decoding, you just use it for the input of GRU.
That's correct, dropout is only used for the input of the GRU. During training, the action and BEV output are predicted from the posterior.
Through the paper, I learned the working mechanism of observation dropout. But, during training, when decoding the h and s to the output, why not use the observation dropout ?