DarkRank icon indicating copy to clipboard operation
DarkRank copied to clipboard

Published 20 hours ago • verified publisher icon TuSimple

Data sampling

Open mnikitin opened this issue 4 years ago • 0 comments

Hello! There is an example of running darkrank distillation you provided in Readme, and there you use --even-iter option. This option means that for each example in mini-batch there always would be another example of the same class. Is such way of sampling a strong requirement for the darkrank loss? Can it successfully converge with totally random data sampling?

mnikitin avatar Jan 27 '20 09:01 mnikitin