Metrics

For metrics we recommend using Tensorboard to log metrics directly to cloud storage along side your model. As the model trains you can launch a tensorboard instance locally to monitor your model progress:

$ tensorboard --log-dir provider://path/to/logs

Or you can use the torchx.components.metrics.tensorboard() component as part of your pipeline.

Reference

• PyTorch Tensorboard Tutorial https://pytorch.org/tutorials/intermediate/tensorboard_tutorial.html

• PyTorch Lightning Loggers https://pytorch-lightning.readthedocs.io/en/stable/extensions/logging.html

torchx.components.metrics.tensorboard(logdir: str, image: str = 'ghcr.io/pytorch/torchx:0.8.0dev0', timeout: float = 3600, port: int = 6006, start_on_file: str = '', exit_on_file: str = '') → AppDef

This component runs a Tensorboard server which will render the logs specified by logdir.

Since Tensorboard runs as a service you need to specify the termination conditions. This consists of a timeout as well as an optional exit_on_file which will cause the service to quit when that path is created.

The files are periodically polled for existence via fsspec and will trigger the corresponding behavior when created.

Parameters:

• logdir – fsspec path to the Tensorboard logs

• image – image to use

• timeout – maximum time to run before exiting (seconds)

• start_on_file – start the server when the fsspec path is created

• exit_on_file – shutdown the server when the fsspec path is created