search

sovrasov / flops-counter.pytorch

Flops counter for convolutional networks in pytorch framework
MIT License
2.82k stars 307 forks source link

FLOPS in ResNet50_ImageNet #94

Open miltonmondal opened 2 years ago

miltonmondal commented 2 years ago

Getting 4.12B flops using your code whereas almost all research papers mentioned 4.09B flops for this configuration

(pytorch default 76.15% test accuracy for pretrained model)

Can you please modify the code or mention the reason for getting 0.03B increase in FLOPs?

jkhu29 commented 2 years ago

fvcore will get the output of 4.09G, but it will also print

Skipped operation aten::batch_norm 53 time(s)
Skipped operation aten::max_pool2d 1 time(s)
Skipped operation aten::add_ 16 time(s)
Skipped operation aten::adaptive_avg_pool2d 1 time(s)

Perhaps those papers ignore the computation of some of the operators.

sovrasov commented 1 year ago

@jkhu29 you're right! ptflops also considers batch norms and poolings as non-zero ops, that's why it outputs slightly greater numbers than expected.