Mukosame / Zooming-Slow-Mo-CVPR-2020

Fast and Accurate One-Stage Space-Time Video Super-Resolution (accepted in CVPR 2020)
GNU General Public License v3.0
915 stars 164 forks source link

OOM in train.py . #71

Closed jiangzhiwei2018 closed 1 year ago

jiangzhiwei2018 commented 1 year ago

My env: OS: Windows10 python version: 3.8 pytorch version: 1.13.1 numpy version: 1.23.5 GPU: RTX3090TI RAM: 32GB

A suspected memory leak occurred when I ran train.py for the training process with a single GPU. The usage memory is 95% after a period of time and keeps rising. 20230403030900

Afterwards, I ran a memory analysis through memory_profiler found that there seemed to be over-occupying memory during the load data phase. memory_profiler Maybe it can provide some suggestions for solutions

I used the same training data (Vimeo90K), and didn't make any major changes to the Vimeo7Dataset class but the following to make it fit my own cache_keys.pkl 1 2 3