Data preprocessing should happend on TPU?

From my "knowledge" of TensorFlow 2.0, data preprocessing should happen on TPU. For example, I can process a batch of large images on a micro-size VM (1GB ram) with TPUv3-8. This behavior may be different on GPU.

deeplab2/model/kmax_deeplab.py

Line 177 in 7884ed6

# preprocessing because it is faster on TPUs than on host CPUs. The

Forgive me if my "knowledge" is incorrect

Hi @edwardyehuang,

In our implementation, the data preprocessing happens on the host CPUs.
We do normalization in that line, which will happen on TPUs.

Cheers,

Thanks for the quick reply. Now I understand.