Number of flops

[Pelups]

Hi !
Does someone knows why the number of flops is about twice the number obtained by the authors ?

Thanks

[MG2033]

Because my implementation of group convolution is based on straightforward tensorflow graph operations. However, I think the authors used a fast cuDNN implementation based on caffe.

[Pelups]

Ok ! Thank you for your answer !

[MG2033]

You're welcome. Also, note that the group convolution operator hasn't been implemented yet officially in Tensorflow.

[MG2033]

@Pelups I have an update to this. I found out that the authors have counted mult-add as 'one' unit not as 'two' units. So, the number I have achieved if roughly divided by two (assuming multiplications and additions have the same count), will give us their reported FLOPS.

[YongKun-Liu]

@MG2033 Would you share the "weights.pkl"