Tensorboard

The 1.15 minor series tracks TensorFlow 1.15.

Features

- Embeddings projector now shows sprite images in the nearest neighbors list
(2543) - thanks beasteers
- When recording hyperparameters, the trial ID can now be customized, for easier
integration with existing tuner systems (2442)
- Improvements to Colab and Jupyter notebook integration:
- The `TENSORBOARD_BINARY` environment variable can now be set to invoke a
non-default `tensorboard` binary (2386)
- Error messages are now clearer when the TensorBoard binary fails to launch
(2395)
- The `%tensorboard` magic no longer spams log messages when a different
version of TensorBoard is already running on the same machine (2470)
- The `%tensorboard` magic can now be used in Jupyter notebooks running on
hosts other than `localhost` (2407)
- What-If Tool improvements:
- Errors running inference are now surfaced in the What-If Tool UI (2414)
- Median error stats are now displayed in addition to mean error stats (2434)
- Mesh plugin improvements:
- Now compatible with TensorFlow 2.0 via a new `summary_v2` module (2443)
- The number of vertices in the mesh can now be dynamic (2373)
- Profile dashboard improvements:
- Wasted time now appears in the node table, and can be used as a sort key
(2525)
- Memory bandwidth utilization now appears in the dashboard header (2525)
- Improvements for plugin developers:
- Plugins can now be rendered in an iframe whose source is served from the
plugin backend, eliminating the need to bundle a frontend with the
TensorBoard binary
- Plugins can now be discovered dynamically and loaded at runtime, by defining
a `tensorboard_plugins` entry point
- See our [example dynamically loaded plugin][example-plugin] for a plugin to
use as a starting point, plus documentation
- TensorBoard now uses Polymer 2.7 (2392, et al.)

[example-plugin]: https://github.com/tensorflow/tensorboard/tree/1.15/tensorboard/examples/plugins/example_basic#readme

Bug fixes

- 2614 - "Toggle All Runs" button now behaves correctly on the first click when
many runs are loaded (PR 2633)
- Scalar charts should no longer "become tiny" on certain kinds of rendering
failures (PR 2605)
- 2028 - TensorBoard now logs less verbosely with Werkzeug 0.15.0 and up; it
now behaves the same across Werkzeug versions (PR 2383)
- The What-If Tool can now properly compare two regression models in the initial
Facets Dive view (PR 2414)
- Embedding projector metadata view now wraps long strings correctly (PR 2198)

Features

- New hyperparameters dashboard: see [tutorial and demo][hparams-docs] and
[summary APIs][hparams-apis]
- New dashboard for visualizing meshes and point clouds: see
[README][mesh-readme]
- Graph dashboard now shows Keras conceptual graph: see [tutorial and
demo][conceptual-graph-docs]
- Embedding projector now supports the [UMAP dimensionality reduction
algorithm][umap] ([learn more about UMAP here][umap-tutorial]) (1901) -
thanks cannoneyed
- [TensorBoard notebook support][notebook-docs] is stabilized: in a Colab or
Jupyter notebook, run `%load_ext tensorboard` followed by `%tensorboard
--logdir ./path/to/logs`
- Profile dashboard improvements:
- New pod viewer tool to analyze TPU performance (2111)
- Now allows capturing profiles from TensorBoard (1894)
- What-If Tool improvements:
- Now available as a notebook widget for Jupyter and Colab: see
[demo][witwidget-demo]
- Now shows PR curves and F1 score (2264)
- Now supports Cloud AI Platform, including XGBoost models (2194)
- Now shows feature-level attributions for individual predictions, as
applicable (2252)
- Image dashboard now allows scrolling for large images (2164) - thanks lr1d
- Scalar chart smoothing now caps at 0.999 for convenience (1974) - thanks
flostim
- Scalar chart scroll-to-zoom behavior now requires holding `Alt` (2221)
- `tensorboard` now supports a `--version` command line argument (2097) -
thanks shashvatshahi1998
- Python API now defines `tensorboard.__version__` in addition to
`tensorboard.version.VERSION` (2026)

Bug fixes

- Projector metadata card now formats long words properly (PR 2016) - thanks
makseq
- 2010 - `.tensorboard-info` is now world-writable for multi-user \*nix systems
(PR 2131)
- 1989 - `importlib.reload(tensorboard)` now works properly (PR 2005)

[conceptual-graph-docs]: https://www.tensorflow.org/tensorboard/r2/graphs
[hparams-apis]: https://github.com/tensorflow/tensorboard/blob/1.14/tensorboard/plugins/hparams/api.py#L15
[hparams-docs]: https://www.tensorflow.org/tensorboard/r2/hyperparameter_tuning_with_hparams
[mesh-readme]: https://github.com/tensorflow/tensorboard/blob/1.14/tensorboard/plugins/mesh/README.md#mesh-plugin
[notebook-docs]: https://www.tensorflow.org/tensorboard/r2/tensorboard_in_notebooks
[umap-tutorial]: https://umap-learn.readthedocs.io/en/latest/how_umap_works.html
[umap]: https://github.com/lmcinnes/umap#umap
[witwidget-demo]: https://colab.research.google.com/github/tensorflow/tensorboard/blob/1.14/tensorboard/plugins/interactive_inference/What_If_Tool_Notebook_Usage.ipynb

Bug fixes

- 1895 - Fix `strftime`-related launch error on Windows (PR 1900)
- 1794 - Fix What-If Tool loading examples without inference (PR 1898)
- 1914 - Disable the profile dashboard inside Colab, where it doesn’t work
- 1945 - Fix profile dashboard loading behavior

The 1.13 minor series tracks TensorFlow 1.13.

Compatibility note: As of 1.13, TensorBoard has begun transitioning its own use
of some TensorFlow APIs to align with their naming in TF 2.0, and as a result
TensorBoard 1.13+ strictly requires TensorFlow 1.13+.

Features

- What-If tool notebook mode and general improvements
- Now usable directly inside Jupyter and Colab notebooks (1662, 1745, 1788)
- Added comparison of multiple models (1589, 1672)
- Added CSV loading model (1597)
- Added global partial dependence plots (1604)
- Added custom prediction function support (1842)
- (Alpha) TensorBoard can be embedded inside Juptyer and Colab notebooks via a
`%tensorboard` magic, after loading the `tb.notebook` extension (1813, 1822)
- Profile dashboard overview page now shows step time breakdown (PR 1683)
- Line chart "log" scale is now a true log scale (1507)
- When no --port flag is specified, TensorBoard will now search for open ports
near the default port (6006) if that port is already in use (1851)

Performance improvements

- Faster event file loading by caching runtime check (PR 1686) - thanks abiro

Bug fixes

- 786 (partial) - Avoid trying to smooth plots of constant y-value (PR 1698)
- 1515 - Fix image right-click accessiblity in non-Chromium browsers (PR 1561)
- 1541 - Fix --event_file flag when using --inspect
- 1566 - Fix error on trying to import "google.protobuf.pyext" (PR 1887)
- 1567 - Fix display bug on line chart after toggling series selection
- 1598 - Fix clipping in graph dashboard PNG download (PR 1600)
- 1601 - Fix chart SVG download option in Firefox
- 1623 - Fix --path_prefix interpretation
- 1838 - Fix run selector synchronization across already-loaded dashboards

Bug fixes

- 1620 - Fix path_prefix flag regression (PR 1623)
- 1704 - Fix debugger sidebar resizer

Bug fixes

- 1549 - Run names wrap at all character (PR 1602) - thanks dgrahn
- 1610 - Fix Download as PNG for large graph
- 1684 - Fix bug rendering debugger plugin (PR 1550) - thanks longouyang