Batch Normalization Parameters

Question

I am reading yolov2 object detection model. So i print the summary of the yolo_model. But when i look the summary line by line, i find that the count of parameters for batch normalization layer is weird, it is the channel * 4, but according to my knowledge, every neuron cell has 4 parameters of batch normalization layer. So the total count for each batch normalization layer should be 4 * cell count. Anybody gets any idea?

Answer 1

Yes, you are right. BatchNorm is applied to every unit in a Dense layer. But, for convolution layer, we apply it for different channels only. Since, the weights are shared in a channel. A more comprehensive answer is given here