The calculation process of convolution in Caffe

The recent experiment is to see that the code can increase its own understanding of the recent convolution of the operation of the summary

Convenient for future inspection, but also corrects the previous understanding deviation.

Convolution is an essential operation in convolution neural networks,

The following diagram is a simple relationship between each layer.

You can see a good extension of the relationship, the following is the entire convolution of the approximate process

The upper part of the graph is the traditional convolution operation, and the following figure is the multiplication of a matrix.

The following figure is a detailed operation of the convolution operation in a convolution layer, in him, according to the size of the volume of kernel data and then expand, in the same picture of the different convolution kernel selected row by line, different n words, on the same line after the stitching, different can be multiple channels, Note, however, that each paragraph in the same line should correspond to the convolution window of a position in the original image.

From the above diagram, we can see clearly that the convolution calculation process is

Split a three-channel RGB image into three single channel images, and then have a K-convolution kernel for each channel, the convolution operation, from the diagram can be clearly seen each channel corresponding K convolution and are different, but after the completion of the calculation, the corresponding convolution of the results of the superposition, can be, This will get the feature image of the entire convolution layer. The next step is other actions.

The following figure is a detailed operation of the convolution operation in a convolution layer, in him, according to the size of the volume of kernel data and then expand, in the same picture of the different convolution kernel selected row by line, different n words, on the same line after the stitching, different can be multiple channels, Note, however, that each paragraph in the same line should correspond to the convolution window of a position in the original image.

For convolution operations in the convolution layer, there is also a group concept to illustrate that groups is the number that represents the filter group. The introduction of GRUOP is mainly for the selective connection of the input and output terminals of the channels, otherwise there will be too many parameters. A convolution operation is performed for each group and 1/group input channel and 1/group output channel. For example, there are 4 input, 8 output, then 1-4 belong to the first group, 5-8 belong to the second Gruop