google-research/deeplab2

How long it takes to train kMaXDeepLab on Cityscapes with batch size = 32?

edwardyehuang opened this issue · 1 comments

I found TPUv4 Pod runs very slow when using batch size = 32 on Cityscapes. Reading and augmenting 32 high-resolution images seems to reach the bottleneck of TPU. How long it takes to train kMaXDeepLab on Cityscapes with batch size = 32?