dpressel/rude-carnie

Memory leak and performance issues

Closed this issue · 2 comments

guojia-git commented

Thanks for providing the code and model! It's really a great pointer to start looking at age and gender classification problems.

The current way decoding and cropping are written is causing major performance issues. For one thing, the multi_crop function is constantly creating new Tensorflow resize and crop nodes and adding them to the default TF Graph, leading to ever expanding graph size and thus memory leak.

I tried decoding and cropping with PIL.Image and the speed can be improved by multiple orders of magnitude if running on large amount of images.

dpressel commented

Hi, thanks for your comment. Yes there are some known issues with this, please consider submitting a PR if you have improvements

guojia-git commented

Thanks for your quick response. I realize the problem has been addressed in PR#47. Closing the issue.