TensorBoard¶

TensorBoard is a widely used tool for visualizing and inspecting deep learning models. Determined makes it easy to use TensorBoard to examine a single experiment or to compare multiple experiments.

TensorBoard instances can be launched via the WebUI or the CLI. To launch TensorBoard instances from the CLI, first install the CLI on your development machine.

Analyzing Experiments¶

To launch TensorBoard to analyze a single Determined experiment, use det tensorboard start <experiment-id>:

$ det tensorboard start 7
Scheduling TensorBoard (rarely-cute-man) (id: aab49ba5-3357-4145-861c-7e6ff2d702c5)...
TensorBoard (rarely-cute-man) was assigned to an agent...
Scheduling tensorboard tensorboard (id: c68c9fc9-7eed-475b-a50f-fd78406d7c83)...
TensorBoard is running at: http://localhost:8080/proxy/c68c9fc9-7eed-475b-a50f-fd78406d7c83/
disconnecting websocket

The Determined master will schedule a TensorBoard instance in the cluster. The Determined CLI will wait until the TensorBoard instance is running, and then it will open the TensorBoard web interface in a local browser window.

You view information about scheduled and running TensorBoard instances by executing the following command:

$ det tensorboard list
 Id                                   | Owner      | Description                         | State      | Experiment Id   | Trial Ids   | Exit Status
--------------------------------------+------------+-------------------------------------+------------+-----------------+-------------+--------------
 aab49ba5-3357-4145-861c-7e6ff2d702c5 | determined | TensorBoard (rarely-cute-man)       | RUNNING    | 7               | N/A         | N/A

TensorBoard can also be used to analyze multiple experiments. To launch TensorBoard for multiple experiments use det tensorboard start <experiment-id> <experiment-id> ....

Note

Initially, TensorBoard may not contain metrics when the browser window opens. Data will be available after a trial workload is completed. TensorBoard pulls metrics from persistent storage. It may take up to 5 minutes for TensorBoard to receive data and render visualizations.

Customizing Tensorboards¶

Determined supports initializing TensorBoard with a YAML configuration file. For example, this feature can be useful for running TensorBoard with a specific container image or for enabling access to additional data with a bind-mount.

environment:
  image: determinedai/environments:cuda-10.0-pytorch-1.4-tf-1.15-cpu-0.5.0
bind_mounts:
  - host_path: /my/agent/path
    container_path: /my/container/path
    read_only: true

Details of configuration settings can be found in the Determined Task Configuration.

To launch Tensorboard with a config file, use det tensorboard start <experiment-id> --config-file=my_config.yaml.

To view the configuration of a running Tensorboard instance, use det tensorboard config <tensorboard_id>.

Analyzing Specific Trials¶

Determined also supports using TensorBoard to analyze specific trials from one or more experiments. This can be useful if an experiment has many trials but you would like to only compare a small number of them. This capability can also be used to compare trials from different experiments.

To launch TensorBoard to analyze specific trials, use det tensorboard start --trial-ids <trial_id 1> <trial_id 2> ....

Data in TensorBoard¶

In this section, we summarize how Determined captures data from TensorFlow models. For a more in depth discussion of how TensorBoard visualizes data see the TensorBoard documentation.

TensorBoard visualizes data captured during model training and validation. Data is captured in tfevent files by writing TensorFlow summary operations to disk via a tf.summary.FileWriter. We provide support in each deep learning framework to write and upload metrics as tfevent files. See below for details on how to configure Determined with TensorBoard for your desired framework.

FileWriters are configured to write log files, called tfevent files, to a directory known as the logdir. TensorBoard watches this directory for changes and updates accordingly. The Determined-supported logdir is /tmp/tensorboard. All tfevent files written to /tmp/tensorboard in a trial are uploaded to persistent storage when a trial is configured with Determined TensorBoard support.

Determined Batch Metrics¶

At the end of every training workload, batch metrics are collected and stored in the database, providing a granular view of model metrics over time. Batch metrics will appear in TensorBoard under the Determined group. The x-axis of each plot corresponds to the batch number. For example, a point at step 5 of the plot is the metric associated with the fifth batch seen.

Framework-specific Configuration¶

The following examples demonstrate how to configure TensorBoard for each framework.

TensorFlow Keras¶

To add TensorBoard support for models that use TFKerasTrial, add a determined.keras.callabacks.TensorBoard callback to your trial class:

from determined.keras import TFKerasTrial
from determined.keras.callbacks import TensorBoard


class MyModel(TFKerasTrial):
    ...

    def keras_callbacks(self):
        return [TensorBoard()]

Estimator¶

There is no configuration necessary for trials using EstimatorTrial. Unless configured otherwise, Estimators automatically log TensorBoard events to the model_dir, which Determined then moves to /tmp/tensorboard.

PyTorch¶

To add TensorBoard support for models that use PyTorch, use the writer field in an instance of the TorchWriter class:

from determined.tensorboard.metric_writers.pytorch import TorchWriter


class MyModel(PyTorchTrial):
    def __init__(self, context):
        ...
        self.logger = TorchWriter()

    def train_batch(self, batch, epoch_idx, batch_idx):
        self.logger.writer.add_scalar("my_metric", np.random.random(), batch_idx)

For a full-length example of using TensorBoard with PyTorch, see the mnist-GAN model.

Lifecycle Management¶

Determined will automatically terminate idle TensorBoard instances. A TensorBoard instance is considered idle if it is does not receive HTTP traffic (a TensorBoard that is still being viewed by a web browser will not be considered idle). By default, idle TensorBoards will be terminated after 5 minutes; the timeout duration can be changed by editing tensorboard_timeout in the master config file.

You can also terminate TensorBoard instances by hand using det tensorboard kill <tensorboard-id>:

$ det tensorboard kill aab49ba5-3357-4145-861c-7e6ff2d702c5

To open a web browser window connected to a previously launched TensorBoard instance, use det tensorboard open. To view the logs of an existing TensorBoard instance, use det tensorboard logs.

Implementation Details¶

Determined schedules TensorBoard instances in containers that run on agent machines. The Determined master will proxy HTTP requests to and from the TensorBoard container. TensorBoard instances are hosted on agent machines but they do not occupy GPUs.