Polyaxon tracking is a high level api for logging parameters, code versions, metrics, and outputs when running your machine learning code, both on a Polyaxon deployment or on a different platform/environment.
The tracked information will be later visualized and compared on the Polyaxon dashboard.
Polyaxon tracking lets you log and interact with REST API in a very convenient way.
In previous versions of Polyaxon, we only exposed metrics reporting through, polyaxon-helper.
The new polyaxon-client allows more advanced workflows, both managed by Polyaxon, in-cluster runs, or on external environment (e.g. your local machine).
This section will guide you how to track:
In addition to this high level tracking APIs, when running an experiment/job inside Polyaxon, some paths and other information are exposed:
$ pip install -U polyaxon-client
$ pip3 install -U polyaxon-client
If you want to delegate the installation to polyaxon during the build process,
add a new step to the
run section in your polyaxonfile:
... build: image: ... build_steps: - ... - pip install -U polyaxon-client - ... run: cmd: ...
Since using the Polyaxon client and the tracking api requires code change, e.g.
# Polyaxon experiment experiment = Experiment() # training code ... # Metrics reporting experiment.log_metrics(step=1000, loss=0.01, accuracy=0.97)
Users might need to run the same code outside of a Polyaxon context, which will break since Polyaxon related operations perform api calls.
Starting from v0.3.8, users won't need to perform any change to their code,
they just need to set an environment variable
POLYAXON_NO_OP to true/1, and the Polyaxon code will be ignored.