BFlops is not the same as the results reported in the paper.

[tingtingliangvs]

Hi there! Excellent work and thank you for sharing the code!
I have a question about BFlops calculation. I used the efficientdet/model_inspect.py to compute the Efficientdet-D5 flops and got 270.77BFlops, which is about twice the 135 BFlops reported in paper.
Our results,

Reported in EfficientDet

What's more, I noticed that with a unified unit (BFlops), the number of Flops of NAS-FPN is also half of that reported in your paper.
Reported in NAS-FPN,

Reported in EfficientDet,

It would be great if you could solve my confusion. Thanks a lot!

[mingxingtan]

Hi @LiangTinting , you are right the reported FLOPS in the code is about twice larger than the paper.

This is because the EfficientDet paper uses FLOPS to denote "number of multiply-adds" (see page 1 footnote), where each "multiply-add" is counted as 2 ops in the code. However, in the NAS-FPN paper, FLOPs denotes number of operations, same as the code.

I had a hard time to pick which way to go, but somehow we decided to use FLOPs to denote "number of multiple-adds", following the convention of many previous work.

Sorry for the confusion, but I hope it answers your question.

[mingxingtan]

I am going to close this issue, but feel free to reopen it or add more comments. Thanks!