search

TexasInstruments / edgeai-yolov5

YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite. Forked from https://ultralytics.com/yolov5
https://github.com/TexasInstruments/edgeai
GNU General Public License v3.0
650 stars 120 forks source link

About GMACs reported in the YOLO-Pose paper #26

Closed WongKinYiu closed 2 years ago

WongKinYiu commented 2 years ago

❔Question

Why the GMACs of other methods reported in this paper are doubled when compare with GFLOPs reported in original papers.

Usually, FLOPs may = MACs or = 2 * MACs when use different definition. I have not checked which one is reported in the original paper, but it won't be MACs / 2.

debapriyamaji commented 2 years ago

@WongKinYiu ,

By definition, FLOPs = 2 * MACs. For other bottom-up approaches, all reported results are with flip-test. That's why the complexity is doubled.

Regards, Debapriya

WongKinYiu commented 2 years ago

Thanks.

WongKinYiu commented 2 years ago

@debapriyamaji

add yolor support.

Dataset Model Name Input Size #Params GMACS AP[0.5:0.95]% AP50% AP75%
COCO Yolov5m6_pose_960 960x960 41.4M 66.3 67.4 89.1 73.7
COCO Yolorp6_pose_960 960x960 41.6M 100.6 70.4 90.3 78.1
COCO Yolov5l6_pose_960 960x960 87.0M 145.6 69.4 90.2 76.1

Thanks for your great repo.

debapriyamaji commented 2 years ago

@WongKinYiu That's really awesome. Thanks for training yolor that surpasses yolov5l at lower complexity. Will add these results.

Regards, Debapriya