TensorBoard Nightly

2.21.0a20251023 · active · verified Sun Apr 12

tb-nightly is the nightly build for TensorBoard, a suite of web applications that provides visualization and tooling for machine learning experimentation. It allows users to track and visualize metrics like loss and accuracy, view model graphs, project embeddings, and display various data types. This version tracks the latest development of TensorFlow, offering bleeding-edge features and bug fixes. It follows an active release cadence, typically updating daily with the latest changes from the main development branch.

Warnings

breaking TensorBoard.dev, the hosted service for sharing TensorBoard experiments, has been shut down as of January 1, 2024. The `tensorboard dev upload` command will no longer function.
Fix: Use TensorBoard as a local tool via the open-source project. For sharing results, consider alternatives like Google Colab's TensorBoard integration or third-party platforms like Weights & Biases.
gotcha Installing both `tensorboard` and `tb-nightly` packages in the same environment can lead to file overwrites and hard-to-debug errors, as they both ship the `tensorboard` Python package.
Fix: Always install either `tensorboard` (stable) or `tb-nightly` (development) but not both in the same Python environment. If you need the nightly features, stick to `tb-nightly`.
breaking Python 3.13 removed the built-in `imghdr` module, causing `ModuleNotFoundError` for older TensorBoard versions attempting to import it.
Fix: Upgrade `tb-nightly` to version 2.20.0 or later, which replaces `imghdr` with a `Pillow` dependency for MIME type detection. Alternatively, install `standard-imghdr` if using an older TensorBoard version with Python 3.13.
gotcha TensorBoard has historically faced compatibility issues with `protobuf` versions, leading to runtime errors. Conflicts can arise if other installed libraries pin `protobuf` to an incompatible version.
Fix: Ensure your `protobuf` installation is compatible with your `tb-nightly` version. Check the official TensorFlow/TensorBoard dependency requirements or update `tb-nightly` to the latest version, which often includes updated `protobuf` constraints. If conflicts persist, consider using a dedicated virtual environment.
gotcha Compatibility issues can arise with `numpy` 2.0. TensorBoard has released updates to ensure compatibility with changes in `numpy` 2.0. [cite: 2.18.0, 2.17.1 release notes]
Fix: Upgrade `tb-nightly` to version 2.18.0 or later to ensure full compatibility with `numpy` 2.0.

Install

pip install tb-nightly Install tb-nightly

Imports

SummaryWriter
```
from torch.utils.tensorboard import SummaryWriter
```
Commonly used for logging from PyTorch or general Python scripts.
TensorBoard (Keras Callback)
```
from tensorflow.keras.callbacks import TensorBoard
```
Used as a callback in TensorFlow/Keras training for automatic logging.
tf.summary
```
import tensorflow as tf; tf.summary
```
Direct API for logging various data types in TensorFlow.
Program
```
from tensorboard import program
```
While 'tensorboard.program' is the core entry point for launching, users typically invoke TensorBoard via the command line or `%tensorboard` magic in notebooks. Direct programmatic use is for advanced integrations.

Quickstart

This quickstart demonstrates how to train a simple Keras model and log its metrics and histograms to TensorBoard using the TensorBoard callback. After the script runs, it prints a command to launch the TensorBoard web interface to visualize the training process.

import datetime
import os

import tensorflow as tf

# Create a simple Keras model.
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

def create_model():
    return tf.keras.models.Sequential([
        tf.keras.layers.Flatten(input_shape=(28, 28)),
        tf.keras.layers.Dense(512, activation='relu'),
        tf.keras.layers.Dropout(0.2),
        tf.keras.layers.Dense(10, activation='softmax')
    ])

model = create_model()
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# Define the log directory
log_dir = os.path.join("logs", "fit", datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))

# Create a TensorBoard callback
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir, histogram_freq=1)

# Train the model while logging to TensorBoard
model.fit(x=x_train, y=y_train, epochs=5, validation_data=(x_test, y_test), callbacks=[tensorboard_callback])

print(f"TensorBoard logs saved to: {log_dir}")
print("To launch TensorBoard, run: tensorboard --logdir logs/")

view raw JSON →