Skip to main content

PyTorch Lightning


If you are not already using ClearML, see Getting Started for setup instructions.

PyTorch Lightning is a framework that simplifies the process of training and deploying PyTorch models. ClearML seamlessly integrates with PyTorch Lightning, automatically logging PyTorch models, parameters supplied by LightningCLI, and more.

All you have to do is simply add two lines of code to your PyTorch Lightning script:

from clearml import Task

task = Task.init(task_name="<task_name>", project_name="<project_name>")

And that's it! This creates a ClearML Task which captures:

  • Source code and uncommitted changes
  • Installed packages
  • PyTorch Models
  • Parameters supplied by LightningCLI (when class is instantiated in script)
  • TensorBoard outputs
  • Console output
  • General details such as machine details, runtime, creation date etc.
  • And more

You can view all the task details in the WebApp.

LightningCLI params

See an example of PyTorch Lightning and ClearML in action here.

Automatic Logging Control

By default, when ClearML is integrated into your script, it automatically captures information from supported frameworks, and parameters from supported argument parsers. But, you may want to have more control over what your experiment logs.


To control a task's framework logging, use the auto_connect_frameworks parameter of Task.init(). Completely disable all automatic logging by setting the parameter to False. For finer grained control of logged frameworks, input a dictionary, with framework-boolean pairs.

For example, the following code will log PyTorch models, but will not log any information reported to TensorBoard:

'pytorch': True, 'tensorboard': False, 'matplotlib': True, 'tensorflow': True,
'xgboost': True, 'scikit': True, 'fastai': True, 'lightgbm': True,
'hydra': True, 'detect_repository': True, 'tfdefines': True, 'joblib': True,
'megengine': True, 'catboost': True

You can also input wildcards as dictionary values, so ClearML will log a model created by a framework only if its local path matches at least one wildcard.

For example, in the code below, ClearML will log PyTorch models only if their paths have the .pt extension. The unspecified frameworks' values default to true so all their models are automatically logged.

auto_connect_frameworks={'pytorch' : '*.pt'}

Argument Parsers

To control a task's logging of parameters from argument parsers, use the auto_connect_arg_parser parameter of Task.init(). Completely disable all automatic logging by setting the parameter to False (this includes disabling logging of LightningCLI parameters).


For finer grained control of logged parameters, input a dictionary with parameter-boolean pairs. The False value excludes the specified parameter. Unspecified parameters default to True.

For example, the following code will not log the Example_1 parameter, but will log all other arguments.

auto_connect_arg_parser={"Example_1": False}

To exclude all unspecified parameters, set the * key to False. For example, the following code will log only the Example_2 parameter.

auto_connect_arg_parser={"Example_2": True, "*": False}

Manual Logging

To augment its automatic logging, ClearML also provides an explicit logging interface.

See more information about explicitly logging information to a ClearML Task:

See Explicit Reporting Tutorial.

Remote Execution

ClearML logs all the information required to reproduce an experiment on a different machine (installed packages, uncommitted changes etc.). The ClearML Agent listens to designated queues and when a task is enqueued, the agent pulls it, recreates its execution environment, and runs it, reporting its scalars, plots, etc. to the experiment manager.

Deploy a ClearML Agent onto any machine (e.g. a cloud VM, a local GPU machine, your own laptop) by simply running the following command on it:

clearml-agent daemon --queue <queues_to_listen_to> [--docker]

Use the ClearML Autoscalers, to help you manage cloud workloads in the cloud of your choice (AWS, GCP, Azure) and automatically deploy ClearML agents: the autoscaler automatically spins up and shuts down instances as needed, according to a resource budget that you set.

Cloning, Editing, and Enqueuing

Cloning, editing, enqueuing gif

Use ClearML's web interface to edit task details, like configuration parameters or input models, then execute the task with the new configuration on a remote machine:

  • Clone the experiment
  • Edit the hyperparameters and/or other details
  • Enqueue the task

The ClearML Agent executing the task will use the new values to override any hard coded values.

Executing a Task Remotely

You can set a task to be executed remotely programmatically by adding Task.execute_remotely() to your script. This method stops the current local execution of the task, and then enqueues it to a specified queue to re-run it on a remote machine.

# If executed locally, process will terminate, and a copy will be executed by an agent instead
task.execute_remotely(queue_name='default', exit_process=True)

Hyperparameter Optimization

Use ClearML's HyperParameterOptimizer class to find the hyperparameter values that yield the best performing models. See Hyperparameter Optimization for more information.