TensorBoard Book

tensorboard-book is a web application for organizing and comparing the runs in a TensorBoard log folder. It indexes every run, reports the best value of each scalar, and stores tags, notes, collections and fields alongside the runs.

Features

• A table of all runs with their best scalars, hyperparameters and filters.
• Tags, notes and stars for each run, collections of runs, and typed fields (numbers with units, choices, dates, links and more) that you define.
• Comparison of a collection at the step where a selection scalar is best, with mean ± std over seeds and CSV, Markdown or LaTeX export.
• Overlaid training curves with TensorBoard's smoothing.
• A timeline of when runs were active, and compute hours per day.
• A file browser with previews, diffs and downloads for each run.
• TensorBoard on any selection of runs.

tensorboard-book reads the folder you pass to tensorboard --logdir and never modifies it.

Getting started

1. Installation: install the command and start the app.
2. Quick start: a tour of the app on demo data.

Documentation

• Your runs folder: what counts as a run, and where the app stores its data.
• The views: each page of the app.
• Topics: filters, status and run time, fields, opening TensorBoard, scalars and settings.
• Password and deployment: sharing the app safely.
• Command line: all commands and options.

Disclaimer

tensorboard-book is an independent project. It is not affiliated with or endorsed by Google or the TensorFlow team.