search

likyoo / Siam-NestedUNet

The pytorch implementation for "SNUNet-CD: A Densely Connected Siamese Network for Change Detection of VHR Images"
MIT License
259 stars 60 forks source link

Calculate FLOPs and Params #36

Open yiyi-today opened 2 years ago

yiyi-today commented 2 years ago

You recommend using thop (https://github.com/Lyken17/pytorch-OpCounter) to calculate FLOPs, but the official example of pytorch-OpCounter calculates MACs, does MACs(G)=FLOPs？In addition, why is the Params I calculated slightly different from yours? FC-EF: Params(M)-1.1, MACs=2.02; FC-Siam-conc: Params(M)-1.55, MACs=5.33; FC-Siam-diff: Params(M)-1.35, MACs=4.73; SNUNet(32):Params(M)-12.03, MACs=54.83. Hope to get your answer, thank you！

likyoo commented 2 years ago

here

yiyi-today commented 2 years ago

Thanks a lot for your answer. I still have the following doubts: 1) Following the link you gave, FLOPs = 2 * MACs, right? 2) In addition, I read some papers, why there is some gap between the FLOPs and Params of the same method of calculation, and the gap between Params is even large. Why is this? Is it related to the configuration of different computers? 3) Are FLOPs and Params used to evaluate model complexity? If I want to consider time cost, what indicator should I use? Sorry for asking you a lot of questions at one time, I hope to get your answer within your acceptable range, thank you!