About FLOPs

[zpc-666]

Hello, I use thop.profile to calculate the FLOPs and Params of FocalNet. The results are 30.53G for FLOPs and 3.74M for Params. The FLOPs is different from that (30.63G) in your paper. Excuse me, how do you calculate FLOPs?

[c-yn]

Hi, we use ptflops to calculate FLOPs.

[zpc-666]

Hi, we use ptflops to calculate FLOPs.

Thanks! But when I use ptflops.get_model_complexity_info(model, (3, 256, 256), as_strings=True, print_per_layer_stat=True), the FLOPs is 30.67GMac. Excuse me, did you overlook some of the modules?

[c-yn]

flops, _ = get_model_complexity_info(model, (3,256,256), as_strings=True, print_per_layer_stat=True, verbose=True) print('{:<30} {:<8}'.format('Computational complexity: ', flops))

Hi, we just directly use the reported FLOPs of ptflops. Can you obtain the same value as ours by using the above snipped?

[zpc-666]

flops, _ = get_model_complexity_info(model, (3,256,256), as_strings=True, print_per_layer_stat=True, verbose=True) print('{:<30} {:<8}'.format('Computational complexity: ', flops))

Hi, we just directly use the reported FLOPs of ptflops. Can you obtain the same value as ours by using the above snipped?

The FLOPs is still 30.67GMac. Maybe it's because I use the latest version （0.7.2.1） of ptflops.