Questions for novel RegNet weights

[Kait0]

Hi,

in one of your latest releases you reported higher numbers for regnets on ImageNet
as well as novel variants (RegNetV and RegNetZ)

https://github.com/rwightman/pytorch-image-models/releases/tag/v0.1-tpu-weights

Could you provide more details on what the changes were made to the architectures / training setup compared to the report in the paper?

Specifically I am interested:
Regnety: What changes were made to the training (I am guessing training on TPUs instead of GPUs does not magically improve performance)
RegNetV: What is a pre-activation model? Can you point me to a paper to read up on that?
RegNetz: What is the difference to regNetY here? What does C/D timm variants refer to?

Thanks a lot!