The accuracy of resnet family

[swift1988]

Hi, thanks for your excellent work~
Training a baseline model(train_scratch.py) on cifar100, resnet18 achieved a higher accuracy than resnet34 and resnrt50. I have nothing changed.
Do you know what is wrong? Looking for your reply!

[HowieMa]

Hi, thanks for asking. Could you please show the specific results you get, rather than just say "higher"?

I didn't try Resnet34 on CIFAR-100, but I did try ResNet18 and 50 on it as shown in my paper.
With the hyperparameters in the folder (https://github.com/VITA-Group/Nasty-Teacher/tree/main/experiments/CIFAR100),
we have 77.44% for resnet18 and 78.12% for resnet50 on CIFAR-100, where ResNet50 performs slightly better than ResNet18. As you can see, the difference is just 0.68%.

You may expect to see slightly different results due to different seeds, different versions of PyTorch, NumPy, etc. Based on my experience, the STD on CIFAR can be ~0.5%. To this end, it is possible that you obtain a slightly better result on Resnet18.

[swift1988]

Thx for your reply.
I have 78.085% for resnet18 and 76.681% for resnet50 on CIFAR-100. The version of PyTorch and Numpy is 1.4.0 and 1.17.2, which is identical with your envs as reported in requirments.txt.

Here are training logs for resnet18 and resnet50.

training-50.txt
training-18.txt

[HowieMa]

I guess that it is because of the random seed ...

My training log at that time and trained resnet50 are here: https://drive.google.com/file/d/1YrdQOLubB9i5JJb9bezrgAJV2MDr0ILC/view?usp=sharing.
Hope it is helpful.

[swift1988]

Ok then, I will try that.
thx for sharing.