TensorBoard is a suite of web applications for inspecting and understanding your TensorFlow runs and graphs.
This README gives an overview of key concepts in TensorBoard, as well as how to interpret the visualizations TensorBoard provides. For an in-depth example of using TensorBoard, see the tutorial: TensorBoard: Getting Started. Documentation on how to use TensorBoard to work with images, graphs, hyper parameters, and more are linked from there, along with tutorial walk-throughs in Colab.
You may also be interested in the hosted TensorBoard solution at TensorBoard.dev. You can use TensorBoard.dev to easily host, track, and share your ML experiments for free. For example, this experiment shows a working example featuring the scalars, graphs, histograms, distributions, and hparams dashboards.
TensorBoard is designed to run entirely offline, without requiring any access to the Internet. For instance, this may be on your local machine, behind a corporate firewall, or in a datacenter.
Before running TensorBoard, make sure you have generated summary data in a log directory by creating a summary writer:
# sess.graph contains the graph definition; that enables the Graph Visualizer.
file_writer = tf.summary.FileWriter('/path/to/logs', sess.graph)For more details, see the TensorBoard tutorial. Once you have event files, run TensorBoard and provide the log directory. If you're using a precompiled TensorFlow package (e.g. you installed via pip), run:
tensorboard --logdir path/to/logs
Or, if you are building from source:
bazel build tensorboard:tensorboard
./bazel-bin/tensorboard/tensorboard --logdir path/to/logs
# or even more succinctly
bazel run tensorboard -- --logdir path/to/logsThis should print that TensorBoard has started. Next, connect to http://localhost:6006.
TensorBoard requires a logdir to read logs from. For info on configuring
TensorBoard, run tensorboard --help.
TensorBoard can be used in Google Chrome or Firefox. Other browsers might work, but there may be bugs or performance issues.
The first step in using TensorBoard is acquiring data from your TensorFlow run.
For this, you need
summary ops.
Summary ops are ops, just like
tf.matmul
and
tf.nn.relu,
which means they take in tensors, produce tensors, and are evaluated from within
a TensorFlow graph. However, summary ops have a twist: the Tensors they produce
contain serialized protobufs, which are written to disk and sent to TensorBoard.
To visualize the summary data in TensorBoard, you should evaluate the summary
op, retrieve the result, and then write that result to disk using a
summary.FileWriter. A full explanation, with examples, is in the
tutorial.
The supported summary ops include:
When you make a summary op, you will also give it a tag. The tag is basically
a name for the data recorded by that op, and will be used to organize the data
in the frontend. The scalar and histogram dashboards organize data by tag, and
group the tags into folders according to a directory/like/hierarchy. If you have
a lot of tags, we recommend grouping them with slashes.
summary.FileWriters take summary data from TensorFlow, and then write them to a
specified directory, known as the logdir. Specifically, the data is written to
an append-only record dump that will have "tfevents" in the filename.
TensorBoard reads data from a full directory, and organizes it into the history
of a single TensorFlow execution.
Why does it read the whole directory, rather than an individual file? You might have been using supervisor.py to run your model, in which case if TensorFlow crashes, the supervisor will restart it from a checkpoint. When it restarts, it will start writing to a new events file, and TensorBoard will stitch the various event files together to produce a consistent history of what happened.
You may want to visually compare multiple executions of your model; for example,
suppose you've changed the hyperparameters and want to see if it's converging
faster. TensorBoard enables this through different "runs". When TensorBoard is
passed a logdir at startup, it recursively walks the directory tree rooted at
logdir looking for subdirectories that contain tfevents data. Every time it
encounters such a subdirectory, it loads it as a new run, and the frontend
will organize the data accordingly.
For example, here is a well-organized TensorBoard log directory, with two runs, "run1" and "run2".
/some/path/mnist_experiments/
/some/path/mnist_experiments/run1/
/some/path/mnist_experiments/run1/events.out.tfevents.1456525581.name
/some/path/mnist_experiments/run1/events.out.tfevents.1456525585.name
/some/path/mnist_experiments/run2/
/some/path/mnist_experiments/run2/events.out.tfevents.1456525385.name
/tensorboard --logdir /some/path/mnist_experiments
You may also pass a comma separated list of log directories, and TensorBoard will watch each directory. You can also assign names to individual log directories by putting a colon between the name and the path, as in
tensorboard --logdir_spec name1:/path/to/logs/1,name2:/path/to/logs/2
This flag (--logdir_spec) is discouraged and can usually be avoided. TensorBoard walks log directories recursively; for finer-grained control, prefer using a symlink tree. Some features may not work when using --logdir_spec instead of --logdir.
TensorBoard's Scalar Dashboard visualizes scalar statistics that vary over time; for example, you might want to track the model's loss or learning rate. As described in Key Concepts, you can compare multiple runs, and the data is organized by tag. The line charts have the following interactions:
-
Clicking on the small blue icon in the lower-left corner of each chart will expand the chart
-
Dragging a rectangular region on the chart will zoom in
-
Double clicking on the chart will zoom out
-
Mousing over the chart will produce crosshairs, with data values recorded in the run-selector on the left.
Additionally, you can create new folders to organize tags by writing regular expressions in the box in the top-left of the dashboard.
The HistogramDashboard displays how the statistical distribution of a Tensor
has varied over time. It visualizes data recorded via tf.summary.histogram.
Each chart shows temporal "slices" of data, where each slice is a histogram of
the tensor at a given step. It's organized with the oldest timestep in the back,
and the most recent timestep in front. By changing the Histogram Mode from
"offset" to "overlay", the perspective will rotate so that every histogram slice
is rendered as a line and overlaid with one another.
The Distribution Dashboard is another way of visualizing histogram data from
tf.summary.histogram. It shows some high-level statistics on a distribution.
Each line on the chart represents a percentile in the distribution over the
data: for example, the bottom line shows how the minimum value has changed over
time, and the line in the middle shows how the median has changed. Reading from
top to bottom, the lines have the following meaning: [maximum, 93%, 84%, 69%, 50%, 31%, 16%, 7%, minimum]
These percentiles can also be viewed as standard deviation boundaries on a
normal distribution: [maximum, μ+1.5σ, μ+σ, μ+0.5σ, μ, μ-0.5σ, μ-σ, μ-1.5σ, minimum] so that the colored regions, read from inside to outside, have widths
[σ, 2σ, 3σ] respectively.
The Image Dashboard can display pngs that were saved via a tf.summary.image.
The dashboard is set up so that each row corresponds to a different tag, and
each column corresponds to a run. Since the image dashboard supports arbitrary
pngs, you can use this to embed custom visualizations (e.g. matplotlib
scatterplots) into TensorBoard. This dashboard always shows you the latest image
for each tag.
The Audio Dashboard can embed playable audio widgets for audio saved via a
tf.summary.audio. The dashboard is set up so that each row corresponds to a
different tag, and each column corresponds to a run. This dashboard always
embeds the latest audio for each tag.
The Graph Explorer can visualize a TensorBoard graph, enabling inspection of the TensorFlow model. To get best use of the graph visualizer, you should use name scopes to hierarchically group the ops in your graph - otherwise, the graph may be difficult to decipher. For more information, including examples, see the graph visualizer tutorial.
The Embedding Projector allows you to visualize high-dimensional data; for example, you may view your input data after it has been embedded in a high- dimensional space by your model. The embedding projector reads data from your model checkpoint file, and may be configured with additional metadata, like a vocabulary file or sprite images. For more details, see the embedding projector tutorial.
The Text Dashboard displays text snippets saved via tf.summary.text. Markdown
features including hyperlinks, lists, and tables are all supported.
First, check that the directory passed to --logdir is correct. You can also
verify this by navigating to the Scalars dashboard (under the "Inactive" menu)
and looking for the log directory path at the bottom of the left sidebar.
If you're loading from the proper path, make sure that event files are present. TensorBoard will recursively walk its logdir, it's fine if the data is nested under a subdirectory. Ensure the following shows at least one result:
find DIRECTORY_PATH | grep tfevents
You can also check that the event files actually have data by running tensorboard in inspect mode to inspect the contents of your event files.
tensorboard --inspect --logdir DIRECTORY_PATH
Update: After 2.3.0 release, TensorBoard no longer auto reloads every 30 seconds. To re-enable the behavior, please open the settings by clicking the gear icon in the top-right of the TensorBoard web interface, and enable "Reload data".
Update: the experimental
--reload_multifile=trueoption can now be used to poll all "active" files in a directory for new data, rather than the most recent one as described below. A file is "active" as long as it received new data within--reload_multifile_inactive_secsseconds ago, defaulting to 4000.
This issue usually comes about because of how TensorBoard iterates through the
tfevents files: it progresses through the events file in timestamp order, and
only reads one file at a time. Let's suppose we have files with timestamps a
and b, where a<b. Once TensorBoard has read all the events in a, it will
never return to it, because it assumes any new events are being written in the
more recent file. This could cause an issue if, for example, you have two
FileWriters simultaneously writing to the same directory. If you have
multiple summary writers, each one should be writing to a separate directory.
Update: the experimental
--reload_multifile=trueoption can now be used to poll all "active" files in a directory for new data, defined as any file that received new data within--reload_multifile_inactive_secsseconds ago, defaulting to 4000.
No. TensorBoard expects that only one events file will be written to at a time, and multiple summary writers means multiple events files. If you are running a distributed TensorFlow instance, we encourage you to designate a single worker as the "chief" that is responsible for all summary processing. See supervisor.py for an example.
If you are seeing data that seems to travel backwards through time and overlap with itself, there are a few possible explanations.
-
You may have multiple execution of TensorFlow that all wrote to the same log directory. Please have each TensorFlow run write to its own logdir.
Update: the experimental
--reload_multifile=trueoption can now be used to poll all "active" files in a directory for new data, defined as any file that received new data within--reload_multifile_inactive_secsseconds ago, defaulting to 4000. -
You may have a bug in your code where the global_step variable (passed to
FileWriter.add_summary) is being maintained incorrectly. -
It may be that your TensorFlow job crashed, and was restarted from an earlier checkpoint. See How to handle TensorFlow restarts, below.
As a workaround, try changing the x-axis display in TensorBoard from steps to
wall_time. This will frequently clear up the issue.
TensorFlow is designed with a mechanism for graceful recovery if a job crashes or is killed: TensorFlow can periodically write model checkpoint files, which enable you to restart TensorFlow without losing all your training progress.
However, this can complicate things for TensorBoard; imagine that TensorFlow
wrote a checkpoint at step a, and then continued running until step b, and
then crashed and restarted at timestamp a. All of the events written between
a and b were "orphaned" by the restart event and should be removed.
To facilitate this, we have a SessionLog message in
tensorflow/core/util/event.proto which can record SessionStatus.START as an
event; like all events, it may have a step associated with it. If TensorBoard
detects a SessionStatus.START event with step a, it will assume that every
event with a step greater than a was orphaned, and it will discard those
events. This behavior may be disabled with the flag
--purge_orphaned_data false (in versions after 0.7).
The Scalar Dashboard supports exporting data; you can click the "enable download links" option in the left-hand bar. Then, each plot will provide download links for the data it contains.
If you need access to the full dataset, you can read the event files that
TensorBoard consumes by using the summary_iterator
method.
Yes! You can clone and tinker with one of the examples and make your own, amazing visualizations. More documentation on the plugin system is described in the ADDING_A_PLUGIN guide. Feel free to file feature requests or questions about plugin functionality.
Once satisfied with your own groundbreaking new plugin, see the distribution section on how to publish to PyPI and share it with the community.
Using the custom scalars plugin, you can create scalar plots with lines for custom run-tag pairs. However, within the original scalars dashboard, each scalar plot corresponds to data for a specific tag and contains lines for each run that includes that tag.
Margin plots (that visualize lower and upper bounds) may be created with the custom scalars plugin. The original scalars plugin does not support visualizing margins.
This isn't yet possible. As a workaround, you could create your custom plot in
your own code (e.g. matplotlib) and then write it into an SummaryProto
(core/framework/summary.proto) and add it to your FileWriter. Then, your
custom plot will appear in the TensorBoard image tab.
TensorBoard uses reservoir
sampling to downsample your
data so that it can be loaded into RAM. You can modify the number of elements it
will keep per tag by using the --samples_per_plugin command line argument (ex:
--samples_per_plugin=scalars=500,images=20).
See this Stack Overflow question
for some more information.
Versions of TensorBoard prior to TensorBoard 2.0 would by default serve on host
0.0.0.0, which is publicly accessible. For those versions of TensorBoard, you
can stop the popups by specifying --host localhost at startup.
In TensorBoard 2.0 and up, --host localhost is the default. Use --bind_all
to restore the old behavior of serving to the public network on both IPv4 and
IPv6.
TensorBoard 1.14+ can be run with a reduced feature set if you do not have TensorFlow installed. The primary limitation is that as of 1.14, only the following plugins are supported: scalars, custom scalars, image, audio, graph, projector (partial), distributions, histograms, text, PR curves, mesh. In addition, there is no support for log directories on Google Cloud Storage.
See DEVELOPMENT.md.
First, try searching our GitHub issues and Stack Overflow. It may be that someone else has already had the same issue or question.
General usage questions (or problems that may be specific to your local setup) should go to Stack Overflow.
If you have found a bug in TensorBoard, please file a GitHub issue with as much supporting
information as you can provide (e.g. attaching events files, including the output
of tensorboard --inspect, etc.).