search

yinboc / few-shot-meta-baseline

Meta-Baseline: Exploring Simple Meta-Learning for Few-Shot Learning, in ICCV 2021
MIT License
615 stars 107 forks source link

How to speed up dataloading for Meta-Dataset? #27

Closed LiJunnan1992 closed 3 years ago

LiJunnan1992 commented 3 years ago

Hi, thanks for the amazing code!

I'm running the training script for meta-dataset, but found the GPU utilization rate to be very low. I'm not very familiar with tensorflow, could you help me point out how to speed up the dataloading? Thanks a lot!

yinboc commented 3 years ago

Thanks for your interest in our work!

I am also not quite familiar with tensorflow, thus I remember that I copied their original tensorflow code (in this repo) for data loading. Did you try their original code and see the GPU utilization?

LiJunnan1992 commented 3 years ago

Hi, it seems the speed bottleneck is because of the pytorch augmentations performed in the for loop. I have tried to use multi-processing but cannot get a speed-up due to communication overhead. Please let me know if you have any good solution, appreciate it!

yinboc commented 3 years ago

I see. I also did not find a good way for doing this more efficiently. Ideally, we may want to have these augmentations in the tensorflow code, but I am not very familar with tensorflow and just used the tensorflow loader as a black box.