when the network is trained, the effective feature grid can be obtained by multiplying the feature grid * (mask>0). Why do you need to restore the mask during the compression process and then compress the mask? I think the mask is no longer needed. Could you explain in detail the function of the recovery mask?
when the network is trained, the effective feature grid can be obtained by multiplying the
feature grid * (mask>0)
. Why do you need to restore the mask during the compression process and then compress the mask? I think the mask is no longer needed. Could you explain in detail the function of the recovery mask?