jiaqixuac / MAP-Net

[CVPR'23] Video Dehazing via a Multi-Range Temporal Alignment Network with Physical Prior
MIT License
107 stars 10 forks source link

Too much time for test #11

Closed HaoyouDeng closed 6 months ago

HaoyouDeng commented 9 months ago

Hi, when I test the trained ckp on the test dataset of HazeWorld, it takes too much time, about 20s/task using V100. Is this a normal time cost? Thanks!

jiaqixuac commented 9 months ago

It takes 10+ hours in my case.

HaoyouDeng commented 9 months ago

Is this a issue of mmcv or mmedit? I find that gpu-util is under 10% at most of the time.

YuZheng9 commented 1 month ago

Is this a issue of mmcv or mmedit? I find that gpu-util is under 10% at most of the time.

Hi, I had the same problem and wanted to ask if you solved it yet

HaoyouDeng commented 1 month ago

Is this a issue of mmcv or mmedit? I find that gpu-util is under 10% at most of the time.

Hi, I had the same problem and wanted to ask if you solved it yet

@YuZheng9 I made two changes, and it can be faster.

  1. Do not save images to disks during evaluation.
  2. Use gpus to calculate psnr or ssim. I use the package Kornia, and the result is similar to the code in this repository.

Hope this is helpful to you.

YuZheng9 commented 1 month ago

Is this a issue of mmcv or mmedit? I find that gpu-util is under 10% at most of the time.

Hi, I had the same problem and wanted to ask if you solved it yet

@YuZheng9 I made two changes, and it can be faster.

  1. Do not save images to disks during evaluation.
  2. Use gpus to calculate psnr or ssim. I use the package Kornia, and the result is similar to the code in this repository.

Hope this is helpful to you.

Thank you very much for your reply and good luck with your studies and papers!