search

kakaobrain / fast-autoaugment

Official Implementation of 'Fast AutoAugment' in PyTorch.
MIT License
1.59k stars 195 forks source link

The result on ResNet18 is frustrating #42

Open iamhankai opened 4 years ago

iamhankai commented 4 years ago

Hi Ildoo, thank you for sharing great code.

I tried ResNet18 on ImageNet, but the result is not good. Have you ever experimented on ResNet18, or do you have any suggestions?

Hyperparameters: SGD with linear lr=0.4, batch=1024, weight decay=1e-4, epochs=120.

method top1
baseline 70.68
fast aa 70.22

The learning curves: image

ildoonet commented 4 years ago

@iamhankai

There is a bug with learning rate scheduler with recent version of pytorch. I fixed it so you can try again.

Also I will look into this problem as soon as I can.

https://github.com/kakaobrain/fast-autoaugment/commit/2424224291616a2158a2265069ee95a78a2d13d9