You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Sample to show how input dim(input_dim for FC layer, for CNN, it's prod(tensor.shape[1:])) reciprocal could be used for initialize NN weights rather than sqrt(input_dim_reciprocal) hinted by Xavier:dragon: