Neptune concepts
Run
In your Python code, the run logs metadata of a current model-training run. The metadata can be configs, metrics, and scores.
run = Run(run_id="intrepid-kittiwake-11")
run.log_configs(...)
for step in range(20):
run.log_metrics(...)
Experiment
An experiment is a lineage of related runs.
When you assign an experiment name to a run, it becomes part of the lineage:
run = Run(
run_id="astute-kittiwake-14",
experiment_name="seabird-flying-skills",
)
To create a new run in the lineage, you can fork the experiment at any step:
run = Run(
run_id="astute-kittiwake-17",
experiment_name="seabird-flying-skills",
fork_run_id="astute-kittiwake-14",
fork_step=102,
)
You can then focus on analyzing the latest run of your experiment, but also access the full history of related runs.
Namespaces and attributes
Inside a run, the logged metadata is organized into namespaces and attributes.
Each attribute holds a type of metadata, and attributes can be nested under one or more namespaces (folders).
API token
Each Neptune user or service account is associated with an API token.
When logging metadata, the token is used to check if the account is authorized to send the data to the target Neptune project.
Project
A Neptune project typically represents one machine learning task.
All of your tracked runs appear in the project, letting other team members explore and analyze experiments through shared views and reports.