search

CyFeng16 / MVIMP

Mixed Video and Image Manipulation Program
GNU General Public License v3.0
71 stars 20 forks source link

waifu2x don't work on tpu #36

Closed Mimocro closed 3 years ago

Mimocro commented 3 years ago

I try to run MVIMP_Waifu2x-ncnn-Vulkan_Demo.ipynb on Google Colab with TPU and it's so much faster than GPU (5.00s/it vs 15.00it/s) but after that Output directory is empty. Is there a way to run it normally on TPU?

CyFeng16 commented 3 years ago

Hi, welcome!

As early as 2017, I participated in the open test of TPU, when TPU was only used as a quick verification of the feasibility of the model. Now with the support of ONNX, XLA, and PyTorch-lightning, cross-device training and inference become more convenient.

But I have to admit that the entire pipeline support of this project on TPU equipment is not complete, considering that using TPU to inference small batch-size can not gain its real performance, we did not imagine the scenario of running the project on the TPU hardware initially.

BTW, the TPU unit is a self-consistent CPU/GPU/TPU core combination. Perhaps the result we want is not saved on colabVM.

Thanks for your feedback :)