TensorFlow · Function

tf.reduce_mean

tf.reduce_mean computes averages across selected tensor dimensions. It reduces all dimensions by default and can preserve reduced axes as length-one dimensions.

How it works

The function averages values over the dimensions selected by axis. With no axis, it reduces every dimension to one scalar; with keepdims enabled, reduced dimensions remain as size-one axes, which preserves a broadcast-friendly shape.

Syntax

tf.math.reduce_mean(input_tensor, axis=None, keepdims=False, name=None)

Parameters

Name	Type	Description
`input_tensor`	`Tensor`	Tensor to average; it should use a numeric type.
`axis`	`int, list of int, or None`	Dimensions to reduce. None reduces all dimensions; values must be in [-rank(input_tensor), rank(input_tensor)).
`keepdims`	`bool`	If true, retains reduced dimensions with length 1.
`name`	`string`	Optional operation name.

Returns

The reduced tensor.

Key points

Averages tensor elements along selected dimensions.
The axis argument controls which dimensions participate in the reduction.
Leaving axis unspecified produces a one-element result from the whole tensor.
Use keepdims=True when the output should preserve reduced axes as size-one dimensions.
axis accepts the dimensions selected for averaging.
The input tensor is required to use a numeric data type.
name can optionally label the TensorFlow operation.

Common mistakes

Each axis index must fall within the valid range for the input tensor's rank.
Do not provide duplicate dimensions in axis.

Example: Reduce tensor means

import tensorflow as tf

x = tf.constant([[1., 1.], [2., 2.]])
print(tf.reduce_mean(x).numpy())
print(tf.reduce_mean(x, axis=0).numpy())
print(tf.reduce_mean(x, axis=1, keepdims=True).numpy())

Computes the overall mean, column means, and row means while retaining the reduced row dimension.

When to use it

Global average

Compute a single average from all values in a tensor.

Axis-wise aggregation

Average features or samples independently along one or more axes.

Shape-preserving reduction

Keep output dimensions compatible with the original tensor for broadcasting.

Interview questions

QWhat does tf.reduce_mean compute, and what happens when axis is None?

It computes the mean of input_tensor across the dimensions listed in axis. If axis is omitted or None, every dimension is reduced and the result contains one element.

Follow-up: How can the result retain reduced dimensions? Set keepdims=True; each reduced dimension remains with length one.

QWhat is the effect of keepdims?

keepdims=False is the default, so reduced dimensions are removed from the result shape. With keepdims=True, those dimensions remain but have length one.

Follow-up: Why might retaining them help? The resulting shape is often easier to broadcast against the original tensor.

QWhat input and output requirements does the function document?

input_tensor must have a numeric type. axis identifies dimensions to reduce, and the function returns the reduced tensor.

Follow-up: What naming option is available? name optionally assigns a name to the operation.

tf.math.reduce_mean
tf.reduce_sum
tf.math.reduce_sum
tf.keras.layers.GlobalAveragePooling3D

Source: tf.math.reduce_mean | TensorFlow v2.16.1. Checked 2026-10-15.