tbexport extracts scalars, images, and audio from TensorBoard log directories and writes them out as static HTML files with accompanying assets. Scalar tags are rendered as plot images, and image/audio tags get interactive step scrubbers.

The output is plain HTML + assets, so it works well for embedding in static site generators (Hugo, Jekyll, MkDocs, etc.) or anywhere else you want to present training results without running a TensorBoard server.

Installation

tbexport is not yet on PyPI. To install it, clone the repo and use uv tool install .:

Usage

Say you want to show off the loss, generated images, and generated audio from a recent training run in your blog. You'd run

tbexport --log-dir /path/to/tensorboard/logs \
  --scalars loss \
  --images gen_img \
  --audio gen_audio

Each tag produces an HTML file (e.g. loss.html) and an accompanying _assets/ directory. When media elements must be scrubbed through different type-steps, an accompanying 'scrubber.js' file is generated as well, which should be included in the resulting page.

Options

• --log-dir — Path to the directory containing tfevents files.
• --scalars — Scalar tag(s) to plot as line charts.
• --images — Image tag(s) to extract with step scrubbers.
• --audio — Audio tag(s) to extract with playback controls.

Tags can be specified individually or as a prefix to match all sub-tags. For example, --scalars train will match train/loss, train/accuracy, etc., and group them into a single page.

Multiple tags of the same type can be passed as separate arguments:

tbexport --log-dir ./logs --scalars loss --scalars accuracy