https://www.slideshare.net/ssuser06e0c5/convolution-77257148
1. Common convolution
input feature map : F x F x N(number of channels)
output feature map : F x F x M(number of channels)
=>
requirement:
numbers of kernel parameters : K x K x N x M(number of kernels)
total calculations : F x F x N x K x K x M
2. Pointwise convolution
input feature map : F x F x N(number of channels)
output feature map : F x F x M(number of channels)
=>
requirement:
number of kernel parameters : 1 x 1 x N x M
total calculations: F x F x 1 x 1 x N x M
3. Depthwise convolution
input feature map : F x F x N(number of channels)
output feature map : F x F x N(number of channels)
=>
requirement:
number of kernel parameters : K x K x N
total calculations: F x F x N x K x K
'Papers > Machine learning' 카테고리의 다른 글
| Loss function for faster rcnn, ssd , yolo (0) | 2019.05.14 |
|---|---|
| Box regression and object classification in RPN (0) | 2019.05.13 |
| Anchor size for yolo (0) | 2019.04.27 |
| Loss function for yolo (0) | 2019.04.27 |
| Faster RCNN vs. SSD (0) | 2019.04.27 |