Depthwise separable convolution explanation with the detailed procedure

1. Depthwise Separable Convolution
Dong-Won Shin

2. • comprise of a convolution performed over each channel of an input layer and
followed by a 1x1 convolution that takes the output channels from the previous
step and then combines them into an output layer.
• different than regular convolutions, mainly because of the reduction in the number
of parameters.
Depthwise Separable Convolution
2 [1] F. Chollet, “Xception: Deep Learning with Depthwise Separable Convolutions.”
16
32
16
16 32
General convolution depthwise separable convolution
depthwise
conv. 1x1 conv.

3. 3
Input feature map
Detailed Procedure

4. 4
Input feature map
Detailed Procedure

5. 5
depthwise conv. filter 1
Detailed Procedure

6. 6
depthwise conv. filter 2
Detailed Procedure

7. 7
depthwise conv. filter 3
Detailed Procedure

8. 8
depthwise conv. filter 4
Detailed Procedure

9. 9
depthwise conv. filter 5
Detailed Procedure

10. 10
pointwise filter 1
(1x1 conv.)
Detailed Procedure

11. 11
pointwise filter 2
(1x1 conv.)
Detailed Procedure

12. 12
pointwise filter 3
(1x1 conv.)
Detailed Procedure

13. 13
pointwise filter 4
(1x1 conv.)
Detailed Procedure

14. 14
pointwise filter 5
(1x1 conv.)
Detailed Procedure

15. 15
pointwise filter 6
(1x1 conv.)
Detailed Procedure

16. 16
pointwise filter 7
(1x1 conv.)
Detailed Procedure

17. 17
pointwise filter 8
(1x1 conv.)
Detailed Procedure

18. 18
pointwise filter 9
(1x1 conv.)
Detailed Procedure

19. 19
pointwise filter 10
(1x1 conv.)
Detailed Procedure

20. Thank you
20