The custom scalar dashboard lets users
- Create line charts with custom combinations of runs and tags by assigning each chart a list of regular expressions for tags.
- Create margin charts (for visualizing confidence intervals).
- Lay out the dashboard in a customized way.
To use this dashboard, users first collect scalar data and then lay out the UI.
The (original) scalar dashboard shows line charts for data collected from scalar summaries: Each line chart visualizes data for 1 tag, and tags are automatically organized into categories based on how tags are prefixed.
Hence, during data collection (adding summaries to the model), some users might be compelled to construct tag strings in ways that codge the scalar dashboard to render in desirable layouts. This is suboptimal: The display of data is coupled with the collection of data.
For instance, a user might introduce 2 tags (say rms/foo and rms/bar) that
each pertain to loss for different models. The scalar dashboard introduces 1
chart for each tag, making it harder to compare loss for the 2 models.
In addition, some users desire to organize the dashboard in ways that differ from the default. For instance, some users would like to quickly see charts for certain loss functions above the fold of the viewport.
As expected, users first collect scalar data with summaries. For example:
from tensorboard import summary as summary_lib
...
loss = tf.losses.softmax_cross_entropy(
onehot_labels=tf.one_hot(labels, depth=num_classes),
logits=logits,
weights=weights)
summary_lib.scalar('cross_entropy', loss)Some code might use tf.summary.scalar from the TensorFlow library in lieu of
summary_lib.scalar. That works as well and will continue to be supported by
TensorBoard.
To lay out the dashboard, users pass a layout_pb2.Layout proto to the
summary_lib.custom_scalar_pb method. This method does not return a TensorFlow
op but rather a Summary proto that be passed to the SummaryWriter (See sample
code.).
The Layout proto encapsulates the organization of the dashboard. See
layout.proto for documentation.
A user could optionally use the summary_lib.custom_scalar TensorFlow op write
the layout to disk as part of executing the graph.
This is often not needed because in many cases, users need only to specify the layout once (which also saves disk space). However, sometimes, the user may seek to change the layout over time.
See custom_scalar_demo.py for an example of collecting scalar data and then laying out the dashboard in a customized way.
The example layout contains both multi-line charts and margin charts.
The logic within the demo produces this custom scalar dashboard.
Note that the layout mirrors the layout_pb2.Layout proto that we had passed to
the summary_lib.custom_scalar_pb method. Specifically, we have 2 categories:
(1) losses in 1 chart and (2) trig functions (which is closed by default).
Within the "losses" chart are 2 lines. They correspond to tags that had been
obtained by the regular expression r'loss.*'. Because the 2 tags are for the
same run, the 2 lines differ in markers (One uses squares, while the other uses
diamonds.) to be distinct from each other. Color still encodes the run.
The baz chart is a margin chart. Zoom in (by dragging a rectangle) or hover over points to view margin values. Smoothing does not affect margin charts.
If we expand the "trig functions" category, we find that lines for sine and cosine are within one chart, and the line for tangent resides in a separate one (as expected).
The left hand side of the dashboard is identical to that of the normal scalar dashboard and should be already familiar to users of that dashboard.
For instance, users can opt to change the X-axis to say wall or relative time as well as control how much to smooth charts.
Users can download data for a specific run-tag combination (that is, a specific line within a chart) by turning on "Show data download links" on the left hand side, then selecting the data series within a chart:
Formats supported include JSON and CSV. Each entry corresponds to a point in time.