calculation of AvgPooling

[MarkusAmann]

Could you please give a closer explanation for the calculation of the AvgPool FLOPs? Despite I know that the impact of pooling layers is minimal, I'd like to understand the calculation. As I get from the code the formular for 2dAvgPooling you use is: (kernel_size + 1) * H_out * W_out * in_channels * out_channels
But if I consider the formular on the pytorch website I'd calculate them as follows:
For each kernel beeing applied on your input you have kernel_size * kernel_size - 1 additions plus one division plus one multiplication of your kernel_size inside the denominator. You apply out_channels kernels on in_channels input channels which result in a H_out * W_out output. This is the computation for all multiplications and additions. To be compliant with the regular understanding of MACs and FLOPs you devide by 2 as two MACs result in one FLOP. So finally I get the formular: FLOPs = (kernel_size * kernel_size - 1) * H_out * W_out * in_channels * out_channels / 2

[Lyken17]

I think you suggestion is correct. In total there are H_out * W_out * out_channels pixels and each pixel takes ** in_channels * (kernel_size * kernel_size - 1) ** operations.

For MACs and FLOPs, it is not simple 2 times calculation. In short, it depends how you count the multiplication for floating numbers. I will refactor the codebase to fix the issue.

[MarkusAmann]

Thank you! Will you please let me know as soon as you fixed the code, so that I can update the python package?

[Lyken17]

After talking to an expert in hardware, I made a clarification of MACs and FLOPs. For AVGPooling, the updated MAC should be H_out * W_out * out_channels. If you agree with it, I will update the corresponding code recently.

[Lyken17]

@fortunex3000 any comments?

[llbin-oppo]

After talking to an expert in hardware, I made a clarification of MACs and FLOPs. For AVGPooling, the updated MAC should be H_out * W_out * out_channels. If you agree with it, I will update the corresponding code recently.

But if you do this, it is different from the statistical method of TensorFlow and the user’s conventional perception