search

gyyang23 / AFPN

102 stars 12 forks source link

Will there be a latency results or fps results? #8

Open wcyjerry opened 1 year ago

wcyjerry commented 1 year ago

Hey, thanks for your great work! It seems made a great improve than FPN or PaFPN, the parameters in paper is even smaller than those, but what I concerned more is that the latency or fps. In addition, if it can still get equivalent result to use conv 3x3 with s=2 p=1 to do downsample, cause conv 3x3 is more faster in devices than conv 2x2 or conv 4x4.

gyyang23 commented 1 year ago

We tested the FPS of afpn (faster r-cnn) on mmdetction. Unfortunately AFPN's FPS is not as fast as FPN's. We didn't test the speed of PAFPN (faster r-cnn).

The speed of afpn(yolov5-s) is tested on mmyolo, and the gpu is 3090. The FPS of afpn is 36.3img/s, and the FPS of YOLOv5PAFPN is 39.6img/s. The test results may not be very accurate.

You're right. You can try selecting conv 3x3 for downsampling.

wcyjerry commented 1 year ago

It seems accepectable, thx.

------------------ Original ------------------ From: Guoyu Yang @.> Date: Thu,Jul 6,2023 4:31 PM To: gyyang23/AFPN @.> Cc: Chenyi Wang @.>, Author @.> Subject: Re: [gyyang23/AFPN] Will there be a latency results or fps results?(Issue #8)

We tested the FPS of afpn (faster r-cnn) on mmdetction. Unfortunately AFPN's FPS is not as fast as FPN's. We didn't test the speed of PAFPN (faster r-cnn).

The speed of afpn(yolov5-s) is tested on mmyolo, and the gpu is 3090. The FPS of afpn is 36.3img/s, and the FPS of YOLOv5PAFPN is 39.6img/s. The test results may not be very accurate.

You're right. You can try selecting conv 3x3 for downsampling.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>