Implementation Detail

Open QZJ-2003 opened this issue 2 years ago • 1 comments

I‘m a beginner in the mixup field. After I read the paper, I'm very confused about the implementation details, e.g. transformations of tensors in $\mathcal{M}_\phi$. But, the mix_up function in your code depends on the timm.data.Mixup.

However, in the timm.data.Mixup, there is a annotation saying that the Mixup class uses Mixup or Cutmix ( I maybe misunderstand ?)

Therefore, I want to ask that is there any code for me to understand the detail about AutoMix method easily?

Dec 18 '23 11:12 QZJ-2003

Hi, @QZJ-2003, sorry for making you confused. The implementation of AutoMix was provided in OpenMixup, referring to automix_V1plus.py for the training pipeline and pmix_block.py for the MixBlock architecture. In our previous plan, this repo will provide the timm-style implementation of AutoMix, but haven't finished it yet. To easily utilize AutoMix, please refer to cifar100/automix for CIFAR-100 experiments of AutoMix. Feel free to ask me if there are more questions about AutoMix implementations.

Dec 18 '23 12:12 Lupin1998