Well that'd be research code anyway :) You'd probably better off rebuilding from scratch based on the research paper. The software shouldn't be that complicated in the end, the main challenge (to me) is gathering that 86M images dataset.