extension of distributed training to sparse linear models (L1 regularizers, using primal CoCoA, instead of dual CoCoA for the L2 case in default tensorflow), by @LiamHe - View it on GitHub