mmpretrain icon indicating copy to clipboard operation
mmpretrain copied to clipboard

Published 20 hours ago • verified publisher icon open-mmlab

Efficientnet Flops

Open 3huo opened this issue 2 years ago • 9 comments

请问，mmclassification 统计的efficientnet flops 为什么和官方论文统计的不一样呢

3huo avatar Mar 08 '22 09:03 3huo

Hi, the Efficientnet Flops computed in mmclassification are quite different from the official paper. Why is it

3huo avatar Mar 08 '22 09:03 3huo

Thank you for your report. We use the tool in mmcv to get_flops, it can not calculate flops of Conv2dAdaptivePadding. image

We will fix later.

Ezra-Yu avatar Mar 08 '22 10:03 Ezra-Yu

yea, I also found it. I added the Conv2dAdaptivePadding to the 'get_modules_mapping()' function.

3huo avatar Mar 09 '22 01:03 3huo

Have you gotten the relatively correct flops of Efficientnet?

Ezra-Yu avatar Mar 09 '22 02:03 Ezra-Yu

Have you gotten the relatively correct flops of Efficientnet?

yes, I got the relatively correct flops, but I'm not very sure if it's precise enough.

3huo avatar Mar 09 '22 06:03 3huo

Yep, Can you create a PR and modify the flops num in readme.md in Efficientnet?

Ezra-Yu avatar Mar 09 '22 07:03 Ezra-Yu

Hi, I just modified the efficientnet flops results in README.md as well as the model_zoo.md on my repo.

3huo avatar Mar 09 '22 11:03 3huo

you can create a PR refer to this doc. welcome your contribution to MMClassification.

Ezra-Yu avatar Mar 09 '22 11:03 Ezra-Yu

you can create a PR refer to this doc. welcome your contribution to MMClassification.

I created it just now

3huo avatar Mar 09 '22 12:03 3huo

This issue will be closed as it is inactive, feel free to re-open it if necessary.

tonysy avatar Dec 12 '22 15:12 tonysy