Reproducibility on Resnet-110 on cifar-100

[GuoleiSun]

Hi, I run your code on cifar-100 using Resnet-110 because PyramidNet-200 takes long time to finish. I used the default parameters as PyramidNet-200. Here is what I got: try1: 20.23 4.65 try2: 21.3 4.93

However, 20.11, 4.43 are reported in the paper (in Table 6). May I know what settings did you use for Resnet-110?

[hellbell]

@GuoleiSun Thank you for having interests in our work!

We used 2 GPUs with a training setting for ResNet-110 + CutMix as:

--depth 110 
--batch_size 64 
--lr 0.1
--cutmix_prob 1.0
--beta  1.0

Our ResNet-110 experiment logs:

\	top-1	top-5
try1	20.52	4.40
try2	20.10	4.43
try3	19.72	4.46
average	20.11	4.43

Hope this works!

[GuoleiSun]

Oh, you used learning rate of 0.1, instead of 0.25, for Resnet-110+cutmix. I will try 0.1 and let you know the results soon.