About the test accuracy of cifar 100 on ResNet-18 without cutout

[uoguelph-mlrg]

About the test accuracy of cifar 100 on ResNet-18 without cutout

[TDeVries]

I didn't do any training tricks, aside from standard procedures like learning rate scheduling that are described in the paper. The code I originally used to get those numbers is almost identical to what is in the repo here. The only change is that I made things a bit neater for the release, but all of the actual processes are the same.

[WeihanLikk]

Thank you for your reply! I have tried to run many times, but still fail to achieve this accuracy. I think it is may because of the Pytorch version or hardware environment? I use Pytorch 1.4.0 and my environment is two titan x GPUs

Here is the record of my training process (.csv file) :

cifar100_resnet18.zip

[TDeVries]

Thanks for uploading your log!

There's something weird going on here. I plotted the results from your log and one of my logs (with data augmentation, no cutout), and there are a couple differences. The most obvious one is that yours doesn't have the sudden jumps which the learning rate scheduling causes. It looks like in PyTorch 1.4.0 they deprecated the old scheduler, so maybe that's causing the issue. You can try replacing scheduler.step(epoch) with scheduler.step() in train.py. The other noticeable difference is that your model appears to be learning much slower than mine, but I have no idea what might be causing that issue. Your model starts out fast but then slows down a lot, so maybe it's also related to the scheduler. I would try changing that and see how it does.

Your log

My log

[WeihanLikk]

Thanks for your advice! I manage to achieve the accuracy of 78%. I follow your advice by replacing scheduler.step(epoch) with scheduler.step(), and it works!

And here is my log:
cifar100_resnet18.zip

[TDeVries]

Thanks for the update! I've added some comments to the code in case anyone else runs into the same issue.