The fvcore's FlopCounterMode reports the MACs and not FLOPs right ?
I tried out the tool using Resnet50 and the output came out to be around 4, which has to be the GMACs acc to the original paper but the tool reports this as FLOPs.
The FLOPs would be 2 * GMACs that is 8 in this case.
I also verified the same using an alternate tool that reports GFLOPs and the forward pass GFLOPs are reported as 8, which is correct.
PFA result of running Resnet50 on the alternate tool :
Hello @ppwwyyxx , @nikhilaravi
The
fvcore's FlopCounterMode
reports the MACs and not FLOPs right ? I tried out the tool using Resnet50 and the output came out to be around 4, which has to be the GMACs acc to the original paper but the tool reports this as FLOPs. The FLOPs would be2 * GMACs
that is 8 in this case.I also verified the same using an alternate tool that reports GFLOPs and the forward pass GFLOPs are reported as 8, which is correct. PFA result of running Resnet50 on the alternate tool :
PFA result of running Resnet50 on FVCORE tool :