MACs probably indicate FLOPs and FLOPs probably indicate 2x FLOPs

[miyaken1201]

My library environmet is as follows

torch:2.6.0
torchvision:0.14.0+cu116
calflops:0.3.2

I executed the following code.

import torch
from torchvision.models import resnet18, vit_b_16, swin_s
from calflops import calculate_flops

resnet = resnet18(weights=None).cuda()
vit = vit_b_16(weights=None).cuda()
swin = swin_s(weights=None).cuda()
batch_size = 1
input_shape = (batch_size, 3, 224, 224)
resnet_flops, resnet_macs, resnet_params = calculate_flops(model=resnet,
                                      input_shape=input_shape,
                                      output_as_string=False,
                                      output_precision=2,
                                      print_results=False)

vit_flops, vit_macs, vit_params = calculate_flops(model=vit,
                                      input_shape=input_shape,
                                      output_as_string=False,
                                      output_precision=2,
                                      print_results=False)

swin_flops, swin_macs, swin_params = calculate_flops(model=swin,
                                      input_shape=input_shape,
                                      output_as_string=False,
                                      output_precision=2,
                                      print_results=False)

print(f"ResNet18: {resnet_flops} GFLOPS, {resnet_macs} MACs, {resnet_params} params")
print(f"ViT-B/16: {vit_flops} GFLOPS, {vit_macs} MACs, {vit_params} params")
print(f"Swin-S: {swin_flops} GFLOPS, {swin_macs} MACs, {swin_params} params")`

and I got the following results.

ResNet18: 3.64 GFLOPS, 1.81 GMACs, 11.69 M params
ViT-B/16: 33.72 GFLOPS, 16.85 GMACs, 86.57 M params
Swin-S: 17.52 GFLOPS, 8.74 GMACs, 49.61 M params

However, when looking at the MMclassification and torchvision official documents, it appears that MACs are almost equal to FLOPs and FLOPs are about twice as large. Has anyone else had similar results?

[MrYxJ]

谢谢，MrYx已经收到您的邮件!