Pooling layers support non-square (rectangular) pooling windows where the height and width are specified independently. When using a rectangular window, the padding and stride can also be set asymmetrically to match each spatial dimension. For example, a max-pooling layer with a $$(2, 3)$$ window, a stride of $$(2, 3)$$, and padding of $$(0, 1)$$ applies a $$2$$-row by $$3$$-column window, adds $$1$$ column of padding on each side of the width dimension, and moves the window $$2$$ rows vertically and $$3$$ columns horizontally per step. Applied to a $$4 	imes 4$$ input with values $$0$$ through $$15$$, this configuration produces a $$2 	imes 2$$ output with elements $$5, 7, 13, 15$$.

Claude

Although deep learning frameworks assign default stride and padding values for pooling layers, both can be manually overridden to control the output dimensions. Padding adds extra border elements (typically zeros) to the input before pooling, allowing the pooling window to cover boundary regions. Setting the stride smaller than the window size introduces overlap between adjacent windows, whereas a larger stride increases the downsampling factor. These controls mirror the padding and stride mechanisms available in convolutional layers and are essential for achieving a desired output shape in network architectures.

Overriding Padding and Stride in Pooling Layers

Dive into Deep Learning

Consider a $$4 	imes 4$$ input tensor populated with values $$0$$ through $$15$$. Applying a $$3 	imes 3$$ max-pooling layer with padding $$1$$ and stride $$2$$ yields a $$2 	imes 2$$ output tensor with elements $$5, 7, 13, 15$$. The padding of $$1$$ extends the input borders so the pooling window can be centered near the edges, and the stride of $$2$$ moves the window two positions per step. This demonstrates how adjusting padding and stride together controls the spatial resolution of the pooling output.

Learn Before

Related