Hi,
I download the newest official gmac caculation script and has some question. I found that there are some op which not be included.
For example,
(1) function LDR2HDR: img ** gamma / expo
(2) tensor pixel-wise multiply or add: x1_lin * x1_lin or x1_lin + x1_lin
(3) tensor matrix multiply: x1_lin @ torch.rand(1,3,1900,1900)
I wonder if the above op should be included at last. Or the the output from function get_gmacs_and_params is the final criterion. Thank you.
Hello Liu Cen,
Indeed, we are aware that some operations might not be counted by the current operations-measuring example. To the best of our knowledge the operations currently not included generally account for a very low number of operations when compared to e.g. convolutional layers. We will use several other methods to verify the total number of operations from participants, including but not limited to e.g. fvcore library, manual count, runtime checking. Therefore, as a rule of thumb for all those cases where there are only small variations in our verifications, we will accept the get_gmacs_and_params number. However, if upon verification we find a large divergence on the real number of operations vs the reported number of operations we will investigate carefully and contact the authors for the opportunity to clarify their solution. Potentially this could lead to disqualification if the number of operations is substantially larger than the limit (200GMACs) or the self-reporting example has been used with ill-intend.
Best regards,
Edu.