Would you point out where the double-buffered weight mechanism mentioned in the paper is implemented in the source code? Thanks for your help!