Hi, I'm having a bit of trouble reproducing the claimed accuracy on CIFAR-100. I tried both training the provided architecture in genotypes.py and searching an architecture from scratch, but I can only achieve a final validation accuracy of ~82.4%-82.8%. In both instances I'm using the default hyperparameters with cutout and aux towers.
I'm wondering if there are any changes made to the hyperparameters used in both architecture search and final architecture evaluation for CIFAR-100? Could you also provide the genotype of the architecture reported in the paper?
Hi, I'm having a bit of trouble reproducing the claimed accuracy on CIFAR-100. I tried both training the provided architecture in
genotypes.pyand searching an architecture from scratch, but I can only achieve a final validation accuracy of ~82.4%-82.8%. In both instances I'm using the default hyperparameters with cutout and aux towers.I'm wondering if there are any changes made to the hyperparameters used in both architecture search and final architecture evaluation for CIFAR-100? Could you also provide the genotype of the architecture reported in the paper?
Thank you!
Jeffrey