How to reproduce your FLOPs calculation?

[rub2xr6vef]

Hi @feizc! Thanks for your open-sourcing work. Just as the title, I am curious on how can I reproduce your FLOPs calculation (as reported in Table 1)?

[feizc]

Hi, you can see in the test.py here:

DiS/test.py

Line 12 in c3977f4

def test_dismodel():

It enumerate all model variants, and calculate parameters, FLOPs using thop package.
As it return MACs, we *2 to get the FLOPs.

Referring to: https://cvnote.ddlee.cc/2019/09/04/thop-pytorch-mac-flop-counter.html

[rub2xr6vef]

Thanks for your reply! One more question, what is the batch size you used in your training?
In the technical report, you said that conditioned generation on ImageNet-1K was trained with a batch size of 1024. But the training commands given in the README just use the default batch size of 128. Which one is correct?
Looking forward to your reply and thanks in advance.