Debug trials

Most trial failures come from one of three places: the image cannot start, the CLI arguments do not match, or no valid metrics were printed.

Start with the logs

The trial log should answer three questions:

  • Did the container start?
  • Did the program receive the expected --hpo-* arguments?
  • Did stdout include at least one valid hpo.metrics.* line?

Common failures

  • Name
    unknown argument
    Type
    parser
    Description

    The dashboard parameter name does not match your CLI parser.

  • Name
    missing metric
    Type
    collector
    Description

    The program finished without printing a valid hpo.metrics.* line.

  • Name
    invalid JSON
    Type
    collector
    Description

    The value after = is not JSON-serializable.

  • Name
    data not found
    Type
    runtime
    Description

    The image cannot access a dataset or configuration file it needs.

Debug locally

Run the same command shape on your machine before launching a large experiment.

docker run --rm your-image:latest \
  --hpo-lookback-window=50 \
  --hpo-risk-multiplier=1.4

Confirm the output contains metric lines:

hpo.metrics.objective=0.84
hpo.metrics.duration_seconds=42.1

For exact collector rules, see Metric format.

Was this page helpful?