Definition
Grouped convolution is a variation of the standard convolution operation used in CNN. It was introduced to reduce the computational complexity and number of parameters in deep neural networks while maintaining performance.
If the number of groups is the same as the number of channels, it is called a depthwise convolution.
Algorithm
- Input channels are divided into multiple groups.
- Each group of input channels is convolved with its own set of filters.
- The outputs from each group are concatenated to produce the final output.