Training time per epoch for us was roughly 10k ... Assuming you have downloaded these to a directory named ditm-data (driving in the matrix data): $ ls -1 ditm-data repro_10k_annotations.tgz repro_10k ...