uda
uda copied to clipboard
question about full ImageNet experiment
could you provide the methods how to prepare the 1.3M extra JFT images? What's the domain-relevance data filtering method mentioned in paper?