TensorFlow · Class

callbacks.TensorBoard

Logs Keras training information for visualization in TensorBoard. It can record metrics, the graph, weight histograms, throughput, profiling data, and embeddings.

How it works

TensorBoard is a Keras callback that writes training information to event files for visualization. Configure its log directory and optional logging frequencies, then pass the callback to a model-training method such as fit.

Syntax

tf.keras.callbacks.TensorBoard(log_dir='logs', histogram_freq=0, write_graph=True, write_images=False, write_steps_per_second=False, update_freq='epoch', profile_batch=0, embeddings_freq=0, embeddings_metadata=None)

Parameters

Name	Type	Description
`log_dir`	`str`	Directory in which TensorBoard log files are written.
`histogram_freq`	`int`	Epoch interval for computing layer weight histograms; zero disables them.
`write_graph`	`bool`	Whether to write the model graph.
`write_images`	`bool`	Whether to write model weights for image visualization.
`write_steps_per_second`	`bool`	Whether to log training steps per second.
`update_freq`	`str or int`	Logging cadence: `"batch"`, `"epoch"`, or an integer batch interval.
`profile_batch`	`int or str`	Batches to profile; the documented default is 0.
`embeddings_freq`	`int`	Frequency in epochs for embedding visualization.
`embeddings_metadata`	`str or dict`	Metadata associated with embeddings.

Returns

A TensorBoard callback instance that can be passed to Keras model methods through their callbacks argument.

Key points

The class connects Keras training with TensorBoard visualization.
It writes TensorBoard events covering several training diagnostics.
Logged metrics can be displayed as summary plots.
The model's training graph can be visualized when graph writing is enabled.
write_steps_per_second=True adds training throughput information to the TensorBoard logs.

Common mistakes

The callback depends on TensorFlow being installed; it is not usable in an environment without TensorFlow.
histogram_freq=0 disables weight-histogram computation, so set a positive epoch frequency when those histograms are needed.
With the default epoch update mode, losses and metrics are emitted only after each epoch rather than after every batch.

Example: Log Keras training for TensorBoard

import tensorflow as tf

(x_train, y_train), _ = tf.keras.datasets.mnist.load_data()
x_train = x_train[..., None].astype("float32") / 255.0

model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28, 1)),
    tf.keras.layers.Dense(10, activation="softmax"),
])
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])

callback = tf.keras.callbacks.TensorBoard(log_dir="./logs")
model.fit(x_train, y_train, epochs=2, callbacks=[callback])

Creates TensorBoard event logs while fitting a Keras model; launch TensorBoard with tensorboard --logdir=./logs to inspect them.

When to use it

Monitor training progress

You need loss and metric charts while training a Keras model.

Visualize graph and weights

You want to inspect model structure and layer-weight distributions.

Profile selected batches

You need to investigate training throughput or execution performance.

Inspect embeddings

You want to examine learned representations.

Interview questions

QWhat does tf.keras.callbacks.TensorBoard do?

TensorBoard is a visualization tool, and this callback logs metrics summary plots, the training graph, weight histograms, and sampled profiling information. The callback is supplied to Keras training through the callbacks argument.

Follow-up: How would you inspect performance? Set profile_batch to select batches for profiling, then open the resulting logs in TensorBoard's profiling view.

QHow is the output location and logging cadence configured?

log_dir identifies the directory where TensorBoard-readable log files are saved. A typical setup creates the callback with tf.keras.callbacks.TensorBoard(log_dir="./logs") and passes it to model.fit.

Follow-up: How can logging frequency be changed? Use update_freq="batch", "epoch", or an integer such as 1000 for periodic batch-based logging.

QWhat controls weight-histogram logging?

histogram_freq is measured in epochs and controls how often layer weight histograms are computed. A value of zero turns histogram computation off.

Follow-up: What additional image logging is available? Set write_images=True to write model weights for visualization as images in TensorBoard.

tf.keras.callbacks.Callback
tf.keras.Model.fit
TensorBoard

Source: tf.keras.callbacks.TensorBoard | TensorFlow. Checked 2026-10-15.