TensorFlow · Class

layers.Resizing

A Keras preprocessing layer that resizes individual images or batches to a chosen height and width. It supports several interpolation methods and optional aspect-ratio-preserving cropping or padding.

How it works

Resizing transforms each 3D image or 4D image batch to the requested height and width. By default it uses bilinear interpolation and channels-last ordering; optional cropping or padding can preserve the original aspect ratio instead of distorting the image.

Syntax

tf.keras.layers.Resizing(height, width, interpolation='bilinear', crop_to_aspect_ratio=False, pad_to_aspect_ratio=False, fill_mode='constant', fill_value=0.0, data_format=None, **kwargs)

Parameters

Name	Type	Description
`height`	`int`	Integer output height.
`width`	`int`	Integer output width.
`interpolation`	`str`	Interpolation method; defaults to bilinear.
`crop_to_aspect_ratio`	`bool`	Whether to crop while preserving aspect ratio.
`pad_to_aspect_ratio`	`bool`	Whether to pad while preserving aspect ratio.
`fill_mode`	`str`	Padding fill mode.
`fill_value`	`float`	Padding value.
`data_format`	`str`	Input dimension ordering: channels_last or channels_first.
`**kwargs`	`dict`	Additional layer keyword arguments.

Returns

A 3D unbatched or 4D batched tensor resized to the target dimensions, with channel placement determined by data_format.

Key points

Resizing is a Keras preprocessing layer for standardizing image spatial dimensions.
Set height and width to define the output spatial resolution.
Interpolation can be selected from five documented methods.
The layer changes spatial size without requiring a particular pixel-value range.
Use crop_to_aspect_ratio when cropping is acceptable and preserving proportions matters.
Use pad_to_aspect_ratio when padding is preferable to cropping.

Common mistakes

Inputs must represent either one image or a batch; other tensor ranks are unsupported.
Match data_format to the tensor layout, especially when using channels_first.

Example: Resize a batch of images

import tensorflow as tf

images = tf.random.uniform((2, 32, 48, 3))
resize = tf.keras.layers.Resizing(64, 64)
output = resize(images)
print(output.shape)  # (2, 64, 64, 3)

Creates a resizing layer and applies it to a batch of RGB images.

When to use it

Normalize varying image sizes to a fixed resolution

Model input preprocessing

Crop or pad images without distortion

Aspect-ratio-sensitive vision tasks

Choose nearest interpolation for discrete-valued image data

Segmentation and other label-image pipelines

Interview questions

QWhat input and output shapes does Resizing support?

The layer accepts a 3D unbatched image or a 4D batch and produces the requested spatial dimensions. It supports channels_last by default and can also use channels_first.

Follow-up: How would you avoid geometric distortion? Enable crop_to_aspect_ratio or pad_to_aspect_ratio, choosing the option that matches the preprocessing requirement.

QWhich interpolation methods can the layer use?

Bilinear interpolation is the default. The documented alternatives are nearest, bicubic, lanczos3, and lanczos5.

Follow-up: Which method is often preferable for discrete masks? Nearest-neighbor interpolation avoids blending class IDs.

QHow can Resizing preserve aspect ratio?

crop_to_aspect_ratio preserves proportions by cropping the largest matching window before producing the target size. pad_to_aspect_ratio preserves proportions by adding padding instead.

Follow-up: What constraint applies when selecting these options? They cannot both be enabled simultaneously.

tf.keras.layers.Rescaling
tf.image.resize_with_pad
tf.keras.layers.RandomZoom

Source: tf.keras.layers.Resizing. Checked 2026-10-15.