search

kjunelee / MetaOptNet

Meta-Learning with Differentiable Convex Optimization (CVPR 2019 Oral)
Apache License 2.0
524 stars 97 forks source link

Accuracy of CIFAR-FS #24

Closed SivanDoveh closed 5 years ago

SivanDoveh commented 5 years ago

Hi, I downloaded your code and run it on CIFAR-FS with this line: python train.py --gpu 0 --save-path "./experiments/CIFAR_FS_MetaOptNet_RR" --train-shot 5 \ --head Ridge --network ResNet --dataset CIFAR_FS

but I got 'best accuracy' of 78%. What can I do in order to get 84%?

Thank you! Sivan

kjunelee commented 5 years ago

Hello. I have just finished a new experiment on a new server and the accuracies on CIFAR_FS are: 77.96 ± 0.38 % (meta-validation set) 84.34 ± 0.49 % (meta-test set). You might want to check if you ran your test code on meta-test set as opposed to meta-validation set.

SivanDoveh commented 5 years ago

Thanks, it worked. Do you know why there is such a difference between the test and the validation set?

kjunelee commented 5 years ago

My intuition is that each split (train, test and validation) contains different categories and the result depends on the similarity between training categories and test categories.