TensorBoard event files, hold the TensorFlow

What it does

tensorboardX is a logging library that generates TensorBoard-compatible event files from Python training loops. It exposes a SummaryWriter class with methods like add_scalar, add_image, and add_histogram, letting you emit visualizations—graphs, embeddings, audio, video, and even ONNX models—then open them in standard TensorBoard. The project also supports piping logs directly to Comet for cloud-based experiment tracking.

The interesting bit

The library is essentially a writer for TensorBoard’s event format, acting as diplomatic glue between PyTorch’s ecosystem and Google’s visualization tooling. It covers a surprisingly broad surface area—ONNX graphs, 3D meshes, and PR curves—suggesting it aims to be a complete stand-in for TensorFlow’s native logger.

Key highlights

• Supports a wide range of summary types: scalars, images, figures, histograms, audio, text, graphs, ONNX graphs, embeddings, PR curves, meshes, hyper-parameters, and video.
• Current release (v2.6.5) is tested against PyTorch 2.6, torchvision 0.21.0, and TensorBoard 2.19.0 on Python 3.10 through 3.13.
• Can export scalar data to JSON for processing outside the TensorBoard UI.
• Optional integration with Comet for cloud experiment tracking and dataset management.

Caveats

• Since version 2.1, audio logging requires the separate soundfile package; without it, add_audio() will not work.
• The README and test matrix are heavily PyTorch-centric, so support for Chainer and MXNet is claimed but not demonstrated in the provided documentation.

Verdict

A practical pick for PyTorch developers who want TensorBoard visualizations from their training scripts. If you already live inside TensorFlow or prefer an all-in-one experiment platform, you probably don’t need this bridge.