I recently checked the PyTorch official website(Models and pre-trained weights — Torchvision 0.18 documentation) and noticed that the recorded GFLOPS for MobileNet V2 in torchvision is 0.3. However, according to the original MobileNet V2 paper, the reported MAdds is 300M (million multiply-add operations). As I understand it, FLOPS should typically be twice the number of MAdds, which would mean it should be 0.6GFLOPS.
Could anyone explain how the 0.3 GFLOPS is calculated? Is there a specific interpretation or context that relates this value to the MAdds reported in the original paper? I would appreciate any insights or clarifications on this discrepancy.