search

csxmli2016 / DFDNet

Blind Face Restoration via Deep Multi-scale Component Dictionaries (ECCV 2020)
915 stars 215 forks source link

About training #43

Closed xjw00654 closed 4 years ago

xjw00654 commented 4 years ago

I have tried to use official pix2pixHD repo as the template to reproduce your code on 2080Ti but found OOM ERROR.

Noticed that you used Tesla V100, so i am questioning about how much memory will costed (on 512x512) when training when i use --fp16 or not use --fp16 respectively？

csxmli2016 commented 4 years ago

I have tried to use official pix2pixHD repo as the template to reproduce your code on 2080Ti but found OOM ERROR.

Noticed that you used Tesla V100, so i am questioning about how much memory will costed (on 512x512) when training when i use --fp16 or not use --fp16 respectively？

Hi, the Tesla V100 server that I used was released. You can use a LightCNN to replace Vgg and only use scale-1 and 2 features. This can reduce the memory and achieve comparable performance.