在对输入特征图进行展开的过程中，可以看到对特征图中的元素进行了带冗余的

In the process of unfolding the input feature map, you can see that the elements in the feature map are copied with redundancy, which increases the storage space of the input feature map. When the convolution kernel and step size are relatively small, what is needed The storage space is approximately n×n times the original. In practical applications, the model trained by neural networks with more layers tends to have better results. In order to increase the depth of the network, the actual size of the convolution kernel used is often smaller, so the additional storage overhead required for expansion can be Accepted.

In the process of expanding the input feature map, we can see that the elements in the feature map are redundantly copied, increasing the storage space of the input feature map, when the coil and step length are compared by an hour, the storage space required is about n×n times the original. In practice, the more layers of neural networks trained models tend to work better, and in order to increase the depth of the network, the actual size of the co product cores used is often relatively small, so the additional storage overhead required for unfolding is acceptable.

In the process of expanding the input feature map, we can see that the elements in the feature map are duplicated with redundancy, which increases the storage space of the input feature graph. When the convolution kernel and step size are small, the storage space needed is about n × n times of the original. In practical application, the model trained by the neural network with more layers is often better. In order to increase the depth of the network, the size of convolution kernel used in practice is usually small, so the extra storage cost needed for expansion is acceptable.<br>