Implementing the Convolution and Pooling Layers
So far, we have seen convolution and pooling layers in detail. In this section, we will implement these two layers in Python. As described in Chapter 5, Backpropagation, the class that will be implemented here also provides forward and backward methods so that it can be used as a module.
You may feel that implementing convolution and pooling layers is complicated, but you can implement them easily if you use a certain "trick." This section describes this trick and makes the task at hand easy. Then, we will implement a convolution layer.
Four-Dimensional Arrays
As described earlier, four-dimensional data flows in each layer in a CNN. For example, when the shape of the data is (10, 1, 28, 28), it indicates that ten pieces of data with a height of 28, width of 28, and 1 channel exist. You can implement this in Python as follows:
>>> x = np.random.rand(10, 1, 28, 28) # Generate data randomly >>>...
