GFLOPs & FPS Precision

[igm503]

Hi, thank you for your work!

I have two questions about your paper.

First, you report GMACs for each model throughout the paper, but I want to confirm this is true, and that the reported numbers are not in fact GFLOPs. For example, the GMACs your report for the ConvNeXt models in the paper is what the ConvNeXt paper itself reports as GFLOPs, but my understanding is that GFLOPs tend to be about 2x that of GMACs. Could you provide any clarification about that?

Second, you report FPS measurements for different Cascade MaskRCNN models in Table 9, but I don't think you report what precision the models were running at. Do you happen to have that recorded somewhere? I'm wondering whether it was fp16, fp32, or AMP.