Skip to main content

PyTorch Distributed

The pytorch_distributed_example.py script demonstrates integrating ClearML into code that uses the PyTorch Distributed Communications Package (torch.distributed).

The script initializes a main Task and spawns subprocesses, each for an instance of that Task. The Task in each subprocess trains a neural network over a partitioned dataset (the torchvision built-in MNIST dataset), and reports (uploads) the following to the main Task:

  • Artifacts - A dictionary containing different key-value pairs.
  • Scalars - Loss reported as a scalar during training in each Task in a subprocess.
  • Hyperparameters - Hyperparameters created in each Task are added to the hyperparameters in the main Task.

Each Task in a subprocess references the main Task by calling Task.current_task, which always returns the main Task.

When the script runs, it creates an experiment named test torch distributed, which is associated with the examples project in the ClearML Web UI.

Artifacts#

The example uploads a dictionary as an artifact in the main Task by calling the Task.upload_artifact method on Task.current_task (the main Task). The dictionary contains the dist.rank of the subprocess, making each unique.

Task.current_task().upload_artifact(
'temp {:02d}'.format(dist.get_rank()), artifact_object={'worker_rank': dist.get_rank()})

All of these artifacts appear in the main Task under ARTIFACTS > OTHER.

image

Scalars#

Loss is reported to the main Task by calling the Logger.report_scalar method on Task.current_task().get_logger, which is the logger for the main Task. Since Logger.report_scalar is called with the same title (loss), but a different series name (containing the subprocess' rank), all loss scalar series are logged together.

Task.current_task().get_logger().report_scalar(
'loss', 'worker {:02d}'.format(dist.get_rank()), value=loss.item(), iteration=i)

The single scalar plot for loss appears in RESULTS > SCALARS.

image

Hyperparameters#

ClearML automatically logs the argparse command line options. Since the Task.connect method is called on Task.current_task, they are logged in the main Task. A different hyperparameter key is used in each subprocess, so they do not overwrite each other in the main Task.

param = {'worker_{}_stuff'.format(dist.get_rank()): 'some stuff ' + str(randint(0, 100))}
Task.current_task().connect(param)

All the hyperparameters appear in CONFIGURATIONS > HYPER PARAMETERS.

image

image

Console#

Output to the console, including the text messages printed from the main Task object and each subprocess appear in RESULTS > CONSOLE.

image