Hi organizers, i have some confusion when read this line "Only one DL model can be used, we can not accept the fusion results from many DL models. The computational cost of a single DL model should be less than 100G FLOPs.". How about some method for training like cross-validation (Just only use one model but load different trained weights) can be used for this challenge ? And if i train 2 model than concatenate 2 embedding and train another FC , so i just have only 1 model for inference , is it ok ? Hope to see your answer soon !
Posted by: fruitai @ Feb. 10, 2023, 10:29 a.m.Supplementally, can we use the TTA(Test time augmentation) for inference?
Posted by: BokingChen @ Feb. 10, 2023, 11:17 a.m.Where is the license agreement for the application data usage?
I can't find it on the homepage (https://codalab.lisn.upsaclay.fr/competitions/10080#participate). Can someone tell me?
Thanks!
Posted by: eempty @ Feb. 11, 2023, 11:59 a.m.The computational cost of a single DL model should be less than 100G FLOPs.
Posted by: AjianLiu @ Feb. 26, 2023, 2:10 a.m.I don't see an answer regarding "train 2 model then concatenate 2 embedding". Is this acceptable as long as the combined model is under 100GFLOPS?
Posted by: Mail @ March 1, 2023, 2:01 p.m.