Is there any comparison with lmdb? And is it possible to generate ffcv dataset directly from raw data without going through torch Dataset?