TensorFlow · Class

layers.MaxPooling2D

MaxPooling2D downsamples 2D spatial data by taking the maximum in sliding windows independently for each channel. Configure its window, stride, padding mode, and tensor data layout.

How it works

MaxPooling2D slides a window across an image's height and width independently for every channel, retaining the largest value in each window. The window movement is controlled by strides; padding determines whether border areas are excluded or padded so spatial dimensions can be preserved.

Syntax

tf.keras.layers.MaxPool2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None, name=None, **kwargs)

Parameters

Name	Type	Description
`pool_size`	`int or tuple of 2 integers`	Pooling-window size along the two spatial dimensions; an integer uses the same size in both dimensions.
`strides`	`int, tuple of 2 integers, or None`	Distance the window moves along each spatial dimension; defaults to pool_size when None.
`padding`	`string`	Padding mode: valid adds no padding; same pads spatial borders toward equal output dimensions.
`data_format`	`string or None`	Input dimension ordering, either channels_last or channels_first.
`name`	`string or None`	Optional layer name.
`**kwargs`	`additional keyword arguments`	Additional layer keyword arguments.

Returns

A tensor containing the max-pooled 2D feature maps, with spatial dimensions determined by pool_size, strides, padding, and data_format.

Key points

Use this layer to reduce two-dimensional spatial feature maps, such as image representations.
Each output value is the largest activation found within the corresponding pooling window.
Strides control how far the pooling window moves across height and width.
pool_size accepts one integer or two integers describing the window dimensions.
padding accepts valid or same without case sensitivity.
The channels-last layout places the channel dimension after the two spatial dimensions.

Common mistakes

With valid padding, border positions are not padded and the output may be smaller than expected.
Same padding preserves height and width only according to the documented padding behavior; it does not mean every stride preserves dimensions.
Set data_format consistently with the tensor layout; channels_last and channels_first expect different dimension orders.

Example: Downsample image features

import tensorflow as tf

layer = tf.keras.layers.MaxPooling2D(
    pool_size=(2, 2), strides=2, padding="same"
)
x = tf.random.normal((1, 28, 28, 3))
y = layer(x)
print(y.shape)  # (1, 14, 14, 3)

Creates a 2×2 max-pooling layer, applies it to a channels-last image batch, and prints the downsampled shape.

When to use it

CNN feature-map downsampling

Reduce height and width in convolutional image models while retaining the strongest local activations.

Computation reduction

Lower the number of spatial positions before later convolutional or dense processing.

Controlled spatial dimensions

Use same padding when border coverage and a target spatial sizing policy matter.

Interview questions

QWhat does MaxPooling2D do?

It performs spatial downsampling by taking the maximum value inside each pool_size window independently for every channel. The window moves over height and width using strides.

Follow-up: How does padding affect the result? With valid padding, no border padding is added; same padding adds balanced padding around spatial edges to target the documented output sizing.

QExplain pool_size and strides.

pool_size specifies the spatial window, and it may be one integer or a pair of integers. If strides is None, TensorFlow uses pool_size as the stride.

Follow-up: Why might explicit strides be useful? A smaller stride creates overlapping pooling windows, while a larger one advances farther between windows and can reduce spatial resolution more aggressively.

QHow does data_format change the input shape?

The layer accepts channels_last or channels_first layouts through data_format. For channels_last, inputs use batch, height, width, channels ordering; channels_first places channels before height and width.

Follow-up: What should be checked when changing data_format? The input layout and the expected output layout must match the selected setting, or dimensions may be interpreted incorrectly.

tf.keras.layers.MaxPool2D
tf.keras.layers.GlobalMaxPool2D
tf.keras.layers.AveragePooling2D
tf.keras.layers.UpSampling2D

Source: tf.keras.layers.MaxPool2D | TensorFlow v2.16.1. Checked 2026-10-15.