tensorflow/tensorboard

TensorBoard (tensorflow/tensorboard) – Explainer (grounded in README only) What it is TensorBoard is a suite of web applications for inspecting and understanding TensorFlow runs and graphs. It runs entirely offline (local machine, behind a firewall, or in a datacenter). How to use it 1. In your TensorFlow code create a summary writer, e.g. python file_writer = tf.summary.FileWriter('/path/to/logs', sess.graph) and write summary data with the summary ops (tf.summary.scalar, tf.summary.image, etc.). 2. Launch TensorBoard pointing at the log directory: bash tensorboard --logdir /path/to/logs (or via Bazel: bazel run tensorboard -- --logdir /path/to/logs). 3. Open http://localhost:6006 in Chrome or Firefox. Key concepts from the README - Summary ops: ops like tf.summary.scalar that produce serialized protobufs; evaluate them and write the result with a FileWriter. - Tags: names you give to each summary op; used to organize data in the frontend. - Event files & logdir: FileWriter writes append‑only records (files containing “tfevents”) to a directory (logdir). TensorBoard reads the whole directory, stitching together multiple event files if training restarts. - Runs: when TensorBoard starts, it walks the logdir tree; each subdirectory that contains tfevents files is treated as a separate run, enabling side‑by‑side comparison of experiments. Visualizations described - Scalar Dashboard: line charts of scalar statistics (loss, learning rate, etc.) over time, with zoom, pan, and cross‑hair tooltips. - Histogram Dashboard: shows how the distribution of a tensor changes over time; can view as offset slices or overlaid lines. - Distribution Dashboard: percentile lines (max, 93%, 84%, …, min) derived from histogram data, interpretable as ±σ bands on a normal distribution. - Image Dashboard: latest PNG for each tag, laid out in rows (tags) × columns (runs). - Audio Dashboard: playable audio widgets for the latest audio per tag. - Graph Explorer: visualizes the TensorFlow graph; use name scopes for clarity. - Embedding Projector: visualizes high‑dimensional embeddings from checkpoint files, optionally with metadata/vocab/sprites. - Text Dashboard: renders Markdown text snippets (hyperlinks, lists, tables). - Time Series Dashboard: unified view that can place scalar, histogram, and image cards side by side; supports run colours, pinning, settings persistence, and tag‑filter autocomplete. Additional notes from the README - TensorBoardğmen to run without internet access. - It expects only one active events file at a time; multiple summary writers should write to separate directories. - Data may be downsampled for RAM efficiency; control with --samples_per_plugin. - Orphaned data from checkpoints can be purged via SessionLog.START events (or disabled with --purge_orphaned_data false). - Plugins can be added; see the ADDING_A_PLUGIN guide and example plugins. All of the above is taken verbatim or paraphrased from the supplied README; no external facts are added.