Implementation of delated convolution

vision
1

I am interested in time of inference of delated convolution. Is it the same time like usual convolution?
I know that Flops depend on output size, kernel size and amount of in-/out- channels.
I can’t find an implementation of delated convolution. Does kernel size change in delated convolution?

I did a test. It turned out that usual convolution faster, then delated. Why does it happen?

image
image913×280 7.1 KB

image

image
image1376×136 6.57 KB

@ptrblck_de, please, answer the question)

2

The flops will be nominally the same, but

  • The memory accesses will be less “compact”,
  • they are less often used and likely less well-tuned - convolutions are among the crazily well-tuned and computationally very demanding functions.

Best regards

Thomas

3

Thank you, Thomas.

Is there something better then dilated convolution, if I want to have a big receptive field?
I want to replace dilated convolution, because to count it longer then usual convolution.

4

The most standard way is pooling, probably. I’m not sure there are many much better ways.

Best regards

Thomas

5

Thank you very much, Thomas