TensorBoard 0.1.4
TensorBoard is now released as a separate pip package, tensorflow-tensorboard
[1]. TensorFlow depends on this package, so no user action is necessary. The TensorBoard 0.1.x series corresponds to TensorFlow 1.3.x.
Features
- TensorBoard now has a fully featured plugin system. Existing first-party dashboards have been converted to use this API, and so are not more privileged than arbitrary third-party plugins. See https://github.com/tensorflow/tensorboard-plugin-example for details.
- Visualizations are now paginated, which prevents large datasets from locking up the CPU.
- We now offer better accessibility for color blind users. We thank Paul Tol for his help. See #288
- In the graph explorer, nodes representing TensorFlow functions (function.Defun) can now be expanded.
- In the graph explorer, nodes can be colored by TPU compatibility, which clarifies whether a model can run on tensor processing units.
- Only active dashboards appear in the main menu, which reduces clutter and helps especially on smaller screens.
Bug Fixes & Improvements
- TensorBoard now loads faster over the network, with fewer HTTP requests.
- Scalar chart tooltips no longer show misleading values when smoothing is disabled.
- The image dashboard now offers a dashboard-wide toggle for showing images at actual size.
- Downloading a graph from the graph explorer no longer results in a tiny image.
- Log output now looks nicer.
Known Issues
- The function
tensorboard.util.encode_wav
and the moduletensorboard.plugins.audio.summary
depend on the latest nightly version of TensorFlow, and will not work with TensorFlow version 1.3.0.
[1] Will be renamed tensorboard
on PyPi in the future. Please only install tensorflow-tensorboard
for now.