tensorboardX

What it is and what it does

TensorBoardX is a standalone library that writes TensorBoard event files without requiring TensorFlow as a dependency. It provides a SummaryWriter interface for logging scalars, histograms, images, audio, embeddings, graphs, and other experiment artifacts to disk or cloud storage (S3, GCS). The library acts as a bridge between training code and TensorBoard visualization, letting you monitor experiments in TensorBoard regardless of which deep learning framework you use.

The package supports PyTorch, JAX, PaddlePaddle, and other tensor libraries through a framework-agnostic API. It handles writing to local disk, Amazon S3, and Google Cloud Storage, with optional optimizations like crc32c acceleration. Recent versions maintain compatibility with modern NumPy and protobuf releases, and the codebase remains actively maintained with regular bug fixes and dependency updates.

Use it for:

• Log training metrics (loss, accuracy) and visualize them in TensorBoard during model training without TensorFlow.
• Track hyperparameter experiments and compare multiple runs side-by-side in TensorBoard's hparams plugin.
• Store experiment artifacts (model graphs, embeddings, sample images) to cloud storage for distributed team access.
• Monitor non-PyTorch deep learning frameworks (JAX, PaddlePaddle) using TensorBoard's visualization interface.
• Archive training logs to S3 or GCS for long-term experiment reproducibility and audit trails.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Writes experiment metrics, graphs, embeddings, and media to TensorBoard event files without requiring TensorFlow, supporting multiple tensor frameworks and cloud storage backends.

Yes. Actively maintained, low install friction, no security vulnerabilities, and MIT-licensed. Use it if you want TensorBoard visualization without TensorFlow, or need to log experiments from multiple frameworks. The recent protobuf version bump and NumPy 1.25+ fixes show the maintainers keep pace with ecosystem changes. Stable choice for experiment tracking in non-TensorFlow workflows.

Install

tensorboardx on PyPI

pip install tensorboardx

uv add tensorboardx

poetry add tensorboardx

Installing tensorboardX

Before you install

Low friction installation with stable maintenance. Active development (last commit 2026-07-14) and recent release (2026-04-03). Depends on three common packages (numpy, packaging, protobuf) with no compiled dependencies. Protobuf version constraint recently bumped to >=5.29.6 in latest release.

License in practice

MIT license permits unrestricted use, modification, and distribution in both open-source and commercial projects with minimal attribution requirements.

Quickstart

pip install tensorboardx

from tensorboardX import SummaryWriter

writer = SummaryWriter()
writer.add_scalar('loss', 0.5, global_step=1)
writer.close()

Requires Python 3.9 or later.

Verify before relying

• Whether the package works with all frameworks it claims to support (JAX, PaddlePaddle, Caffe2, mxnet, Chainer) in current version 2.6.5.
• Performance characteristics and memory overhead when logging large-scale experiments with many metrics.
• Compatibility with recent TensorBoard UI versions and whether all event types render correctly.

Package facts

Evidence: tensorboardx-2.6.5-py3-none-any.whl

Tags

More Monitoring packages