TensorFlow · Module

model.summary

tf.summary provides TensorFlow’s public API for writing training and evaluation summaries. It includes writers and functions for recording values such as scalars, images, histograms, and text.

How it works

The module exposes TensorFlow’s summary-writing API for producing data that tools such as TensorBoard can visualize. Create or select a SummaryWriter, then call type-specific functions inside its context and provide a monotonic step when needed.

Syntax

tf.summary

Parameters

Name	Type	Description
`logdir`	`str`	Directory in which the event file is written.
`max_queue`	`int \| None`	Maximum number of queued summaries before flushing; defaults to 10.
`flush_millis`	`int \| None`	Maximum interval between automatic flushes.
`filename_suffix`	`str \| None`	Optional suffix appended to the event filename.
`name`	`str \| None`	Optional name for the writer operation.
`experimental_trackable`	`bool`	Whether to make the writer trackable.
`experimental_mesh`	`Any`	Optional experimental mesh configuration.

Returns

A SummaryWriter object.

Key points

The module is TensorFlow’s public v2 summary namespace.
create_file_writer produces a writer that stores event data in the specified log directory.
create_noop_writer supplies a context-manager-compatible writer for paths where recording should be inert.
scalar records numeric values for monitoring during training or evaluation.
image records image tensors for later inspection in TensorBoard.
flush makes buffered summary data get sent to the underlying storage.

Example: Write a scalar summary

import tensorflow as tf

writer = tf.summary.create_file_writer("logs")
with writer.as_default(), tf.summary.record_if(True):
    tf.summary.scalar("loss", 0.42, step=1)
writer.flush()

Creates an event writer, records a scalar at step 1, and flushes buffered data to storage.

When to use it

Monitor scalar metrics

Track loss, accuracy, learning rates, or other numeric signals across training steps.

Review image outputs

Inspect predictions, generated samples, or intermediate activations visually.

Record text information

Capture textual metadata, notes, or model-related annotations for later analysis.

Interview questions

QWhat is tf.summary used for?

tf.summary is the public API namespace for writing summaries such as scalars, images, histograms, audio, and text. These summaries are associated with steps and are intended for later visualization or analysis.

Follow-up: How would you prevent summary work during selected iterations? Use tf.summary.record_if with a boolean, scalar boolean Tensor, or callable condition.

QHow do you write summaries to an event directory?

Create a writer with tf.summary.create_file_writer, make it current with its as_default context, and invoke summary functions inside that context. The writer can then be flushed to send buffered event data to storage.

Follow-up: What happens without a current writer? A summary write can be skipped; use a file writer context when output is required.

QHow is the summary step determined?

Summary-writing functions use an explicit int64-castable monotonic step, or the current experimental default step when step is omitted. The default step must not be None.

Follow-up: How can a writer supply the step conveniently? SummaryWriter.as_default(step=...) sets the default step used by summary-writing functions in that context.

tf.summary.SummaryWriter
tf.summary.create_file_writer
tf.summary.create_noop_writer
tf.summary.scalar
tf.summary.image
tf.summary.histogram
tf.summary.text
tf.summary.record_if

Source: Module: tf.summary | TensorFlow v2.16.1. Checked 2026-10-15.