CutMix-PyTorch
CutMix-PyTorch copied to clipboard
About the result of PyramidNet-110 baseline in the paper
In the paper (table 6), PyramidNet-110 without cutmix is reported to achieve 19.85% top1 error on CIFAR100. Since in the paper of PyramidNet, the error rate with the same setting (except that lr equals 0.5) only achieves 23.12%, I wonder how the result of 19.85% is achieved? Could you provide any example training code to get this result? Thanks a lot!
After some experiments, it is found out that this basline here should use bottleneck as the block type, so that the model can reach the result. Right?