I don't have complete numbers on this (I think it depends a lot on the size of training set), but for one dataset, normalized time for a batch: original

I don't have complete numbers on this (I think it depends a lot on the size of training set), but for one dataset, normalized time for a batch: original : 1.000 precomputed : 0.644 (first improvement) gziplength : 0.428 (+ 2nd improvement)