Doubts about the networks parameters and FLOPs

[kravrolens]

Hi, Fancy! Thanks for your excellent work.

StyleSwin synthesizes a 1024x1024 image with 40.86M params and 50.90B FLOPs, as shown in the paper of Table 6.

But I reproduced the results by running:

from thop import profile
flops, params = profile(generator, (noise,))             # noise: torch.Size([1, 512])
print('flops: ', flops / 1000000000, 'params: ', params / 1000000)
flops, params = profile(discriminator, (real_img,))  # real_img: torch.Size([1, 3, 1024, 1024])
print('flops: ', flops / 1000000000, 'params: ', params / 1000000)

The generator params are 28.28M with 47.36B FLOPs.
The discriminator params are 27.73M with 50.19B FLOPs.

I don't know where the problem is. Looking forward to your reply!

[ForeverFancy]

It is because thop does not recognize costome operates like EqualConv, so it will not take these parameters into account. Therefore, we recommand to calculate the params and FLOPs using provided code. Thanks.