Skip to content

Quick start

This page walks through tensorboard-book on a demo folder. It assumes that you know TensorBoard and have installed tensorboard-book.

1. Open the demo

tensorboard-book demo tbook-demo    # writes 14 example runs to ./tbook-demo
tensorboard-book tbook-demo         # starts the app on them

The demo is a standard TensorBoard log folder. It contains a learning-rate sweep (resnet_…), an ablation with three seeds per setting (aug_…), a Keras run and two runs that are still active. Some tags, notes, collections and fields (the dataset, the GPU, the parameter count, a rating) are already set.

2. Find the best run

The app opens on Experiments. Each row is a run, with the best value of each scalar. ↑ means higher is better, ↓ means lower is better.

  • Type optimizer.lr < 1e-3 and val/acc > 0.9 in Filter.
  • In the sidebar, set Collection to lr_sweep.
  • Add Params with Show, or filter on it: Params < 20M.

Experiments filtered with optimizer.lr < 1e-3 and val/acc > 0.9 Experiments filtered with optimizer.lr < 1e-3 and val/acc > 0.9

3. Compare an ablation

Open Compare, pick aug_ablation and turn on Aggregate seeds.

Each run is read at the step where val/acc was best, and the other scalars are taken at that same step. Seeds are averaged into mean ± std. You can export the table as CSV, Markdown or LaTeX.

Further down, Which settings lead to the best result works like TensorBoard's HParams dashboard.

4. Look at curves and at one run

  • Curves overlays runs, with the same smoothing as TensorBoard, and can draw several scalars on one chart.
  • Run details shows one run. Pick resnet_lr3e-3: the Timing tab shows that it crashed and was resumed. The Overview tab shows its fields, the Artifacts tab its files, and Edit changes its tags, notes and fields.

5. Open TensorBoard

For images, histograms and other plugins, tick some rows in Experiments and press Open in TensorBoard.

6. Use your own runs

Stop the demo with Ctrl+C and start the app on the folder that you pass to tensorboard --logdir:

tensorboard-book /path/to/runs

Each folder one level down with event files is a run. Use --depth 2 if your runs are nested deeper. The files in the folder are never modified.

Coming from TensorBoard

In TensorBoard In tensorboard-book
tensorboard --logdir runs tensorboard-book runs
Scalars dashboard Curves, and best values in Experiments
HParams dashboard Compare → Which settings lead to the best result
Run filter (regex) Filter and Search
Images, histograms, graphs Open in TensorBoard
Tags, notes, collections, fields and Timeline

Next