The flops and parameters of NASFPN is not matched

[lucienne999]

Hi, table.1 in [1] show that NAS-FPN R-50 (7 @ 256) FLops and parameters is 281.3B , 60.3 M. However, I use get_flops.py to get 140.42 GF and 68.02M when using NasFPN config.

Refer:
[1] https://arxiv.org/pdf/1904.07392.pdf

[lucienne999]

The config i use: configs/nas_fpn/retinanet_r50_nasfpn_crop640_50e_coco.py
The code i run python tools/analysis_tools/get_flops.py configs/nas_fpn/retinanet_r50_naspredfpn_crop512_10e_coco_NasFPN.py --shape 640 640

I think it is ok with a slight difference. But it looks unreasonable?

[RangiLyu]

These are two different meanings of FLOPs. In some papers, FLOPs means floating-point operations, and it including mul and add. But in some other papers, a multiply-add counts as one flop.
In mmdet, we use https://github.com/sovrasov/flops-counter.pytorch
And there is a related issue: sovrasov/flops-counter.pytorch#40

[lucienne999]

@RangiLyu Thanks for your reply. But how to explain the difference in parameters?

[lucienne999]

I found that the difference in parameters came from the implementation. The difference in flops because nasfpn uses another calculation mentioned by RangiLyu.