What is the definition of MACs ？ about pytorch-opcounter HOT 8 OPEN

I made clarification here https://github.com/Lyken17/pytorch-OpCounter/tree/master/benchmark#macs-flops-what-is-the-difference

from pytorch-opcounter.

I am confused by the FLOPs results in original ResNet paper. Let's take resnet-50 as an example. In ResNet paper the FLOPs value is 3.8G while in this repo we get 4.14G for MACs which means 4.14*2G FLOPs. So can we infer that FLOPs definition is the same here with MACs?

OpCounter
ResNet

The original resnet paper reports MACs - see the bottom-right paragraph on page 3: https://arxiv.org/pdf/1512.03385.pdf . Technically a multiply-add is a single floating point operation since 2008 (https://en.wikipedia.org/wiki/IEEE_754-2008_revision), which is probably why they still call it flops. I suppose one could make the distinction as IEEE754-1985 FLOPs vs IEEE754-2008 FLOPs.

from pytorch-opcounter.

As far as I can understand, the direct output of the "profile" function is MACs instead of FLOPs. However, the demo code in the readme file uses flops as the name of the output variable. Is this a typo? Thanks.

from pytorch-opcounter.

I am confused by the FLOPs results in original ResNet paper. Let's take resnet-50 as an example. In ResNet paper the FLOPs value is 3.8G while in this repo we get 4.14G for MACs which means 4.14*2G FLOPs. So can we infer that FLOPs definition is the same here with MACs?

OpCounter
ResNet

from pytorch-opcounter.

This might help! sovrasov/flops-counter.pytorch#16

from pytorch-opcounter.

Thanks for your supply! It is really helpful.

from pytorch-opcounter.

As far as I can understand, the direct output of the "profile" function is MACs instead of FLOPs. However, the demo code in the readme file uses flops as the name of the output variable. Is this a typo? Thanks.

the demo code in benchmark/readme is wrong. Macs returned by the "profile" function . If you want get the Flops, you should multiply 2.

from pytorch-opcounter.

MACs can also defined as "Memory Access Cost" which is strongly relative to inference time, but I think in this repo is Multi-Add Calculation

from pytorch-opcounter.