TensorBoard

2.20.0 · active · verified Sun Mar 29

TensorBoard is a powerful visualization toolkit for machine learning experimentation, enabling tracking of metrics like loss and accuracy, visualization of model graphs, projection of embeddings, and much more. It is closely integrated with TensorFlow and PyTorch ecosystems, and its releases generally track TensorFlow versions. The current stable version is 2.20.0, and it is actively maintained with regular updates.

Warnings

breaking TensorBoard.dev, the hosted sharing service, has been shut down. The `tensorboard dev upload` command will fail and the website is no longer accessible.
Fix: Migrate to self-hosting TensorBoard or alternative experiment tracking platforms. There is no direct replacement for `tensorboard.dev` functionality within TensorBoard itself.
gotcha TensorBoard plugin compatibility with Keras 3. While TensorFlow 2.16+ uses Keras 3 by default, some TensorBoard plugins' implementations may still primarily support Keras 2. This can lead to unexpected behavior or missing visualizations for Keras 3 models.
Fix: Monitor official releases and documentation for Keras 3 compatibility updates. If issues arise, consider running Keras 2 compatible environments or alternative debugging strategies.
gotcha Protobuf dependency conflicts can occur. TensorBoard's `protobuf` requirements have varied across versions (e.g., tight restrictions, then relaxations). This can cause installation errors or runtime issues if other installed libraries have conflicting `protobuf` version requirements.
Fix: Ensure `protobuf` version is compatible with your TensorBoard installation. Use `pip check` to find conflicts and consider creating isolated virtual environments. If problems persist, try reinstalling `tensorboard` which often pulls a compatible `protobuf` version.
gotcha Python 3.13 compatibility requires TensorBoard version 2.20.0 or higher. Earlier versions will fail on Python 3.13 due to the removal of the `imghdr` module from the standard library, which TensorBoard previously used.
Fix: Upgrade TensorBoard to version 2.20.0 or newer if using Python 3.13.
gotcha When using `SummaryWriter` in notebook environments (e.g., Colab, Jupyter), it's highly recommended to call `writer.flush()` and `writer.close()` after logging data. This ensures all event files are properly written to disk and available for TensorBoard to render, preventing data loss or incomplete visualizations.
Fix: Always explicitly call `writer.flush()` and `writer.close()` at the end of your logging session, or use `with SummaryWriter(...) as writer:` context manager.

Install

pip install tensorboard Install stable version

Imports

SummaryWriter
```
from torch.utils.tensorboard import SummaryWriter
```
Commonly used for logging data from PyTorch models.
TensorBoard callback for Keras
```
from tensorflow.keras.callbacks import TensorBoard
```
Used to integrate TensorBoard logging with Keras `Model.fit()`.

tf.summary

import tensorflow as tf
writer = tf.summary.create_file_writer('logs/my_experiment')

For low-level summary logging in TensorFlow 2.x without Keras callbacks.

Quickstart

This example demonstrates how to use `SummaryWriter` from `torch.utils.tensorboard` to log scalar values, creating event files in a timestamped directory. After running the script, you can launch TensorBoard from your terminal to visualize the logged data.

import datetime
from torch.utils.tensorboard import SummaryWriter

log_dir = "runs/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
writer = SummaryWriter(log_dir)

# Log a scalar value
for i in range(100):
    writer.add_scalar('Loss/train', 100 / (i + 1), i)
    writer.add_scalar('Accuracy/train', i / 100, i)
writer.close()

print(f"TensorBoard logs saved to: {log_dir}")
print("To view, run in your terminal: tensorboard --logdir runs")

view raw JSON →