MobileNetV2 baseline FLOPs

[andreysher]

Hello, i suppose you have some bug in your table. There is 74.29% accuracy, 179.46 MFLOPs and 2.33 MParameters for MobileNetV2 baseline, but according to pytorch-cifar-models repo there is accuracy and parameters for MobileNetV2_1.0, but FLOPs for MobileNetV2_x1_4. And your compressed model have 111.96 FLOPs. This is bigger than original mobilenetv2_x1_0 baseline. Can you comment this situation, or fix your table?

[andreysher]

According to thop profiler your compressed model have 178.64 MFLOPs and 0.816 MParams.

[yjlee0607]

#MAdds in pytorch-cifar-models seems to be meaning multiply-add (MAD (same as MAC)). We used thop profiler and got 94.6M MACs (same as 189MFLOPs). You can see the details in the following codes.

from thop import profile
model = torch.hub.load("chenyaofo/pytorch-cifar-models", "cifar100_mobilenetv2_x1_0", pretrained=True)
device='cpu'
dummy_input = torch.ones((1,3,32,32)).to(device)

model = model.to(device)

original_macs, original_params = profile(model, inputs=(dummy_input, ))
print(original_macs, original_params)

94651392.0 2351972.0

And, we are now in re-check each models' performances.
Thank you very much for your report.