TensorFlow · Module

Dataset.prefetch

Dataset.prefetch buffers upcoming dataset elements so they can be prepared while current elements are processed. It can improve pipeline latency and throughput, with additional memory use; tf.data.AUTOTUNE can tune the buffer size.

How it works

Calling prefetch adds an internal buffer that prepares upcoming dataset elements before they are requested. This can overlap input preparation with model work, improving pipeline throughput while consuming additional memory; tf.data.AUTOTUNE can select the buffer size dynamically.

Syntax

prefetch(buffer_size, name=None) -> 'DatasetV2'

Parameters

Name	Type	Description
`buffer_size`	`tf.int64 scalar tf.Tensor`	Maximum number of dataset elements to buffer; tf.data.AUTOTUNE enables dynamic tuning.
`name`	`string`	Optional name for the dataset operation.

Returns

A DatasetV2 containing the prefetched elements.

Key points

The method returns a dataset transformation that asynchronously prepares upcoming elements.
Preparation of future input can overlap with processing of the current input.
Use tf.data.AUTOTUNE when TensorFlow should adjust the buffer capacity during execution.
The API treats each emitted dataset element uniformly, regardless of whether that element is an example or a batch.
After batching, a buffer size of two means two complete batches are prefetched.

Common mistakes

The buffer counts dataset elements, not individual examples inside a batch.
Larger buffering can improve throughput but increases memory use.
Place this transformation near the end of the input pipeline, typically after batching and other preparation steps.

Example: Prefetch a dataset

import tensorflow as tf

dataset = tf.data.Dataset.range(3)
dataset = dataset.prefetch(tf.data.AUTOTUNE)
print(list(dataset.as_numpy_iterator()))

Creates a range dataset and lets TensorFlow choose the prefetch buffer size automatically.

When to use it

Overlap input work with model execution

Training input pipelines

Let TensorFlow tune buffering

Variable pipeline workloads

Keep several future batches ready

Batched datasets

Interview questions

QWhat does Dataset.prefetch do?

prefetch prepares future dataset elements in an internal buffer so production can overlap with consumption. This can reduce input latency and improve throughput.

Follow-up: How can the buffer size be selected dynamically? Pass tf.data.AUTOTUNE so the tf.data runtime tunes it during execution.

QWhat does buffer_size mean?

The argument specifies the maximum number of input elements buffered. If it is tf.data.AUTOTUNE, TensorFlow tunes that capacity at runtime.

Follow-up: What does the count represent after batching? It represents batches, because each batch is one dataset element.

QWhat are the main trade-offs and semantics of prefetching?

Prefetching can improve latency and throughput, but buffered elements require additional memory. The trade-off should be considered when choosing a fixed buffer size.

Follow-up: Why might prefetch(2) behave differently before and after batch(20)? Before batching it buffers two examples; after batching it buffers two batches.

tf.data.Dataset.batch
tf.data.Dataset.map
tf.data.AUTOTUNE
tf.data.Dataset

Source: tf.data.Dataset. Checked 2026-10-15.