Tracking and logging¶

Logging belongs to ScenarioSpec.logging. Keep experiment semantics in the scenario while each backend owns its external service/runtime.

TensorBoard¶

TensorBoard support is part of the core installation:

from nexuml.core.types import LoggingSpec, TensorBoardSpec

logging = LoggingSpec(
    experiment_name="MyProject",
    tensorboard=TensorBoardSpec(log_dir=".experiments/tensorboard"),
)

Run TensorBoard against the configured directory.

MLflow¶

Install the tracking extra:

uv pip install "nexuml[tracking]"

Then configure the URI explicitly:

from nexuml.core.types import LoggingSpec, MLflowSpec

logging = LoggingSpec(
    experiment_name="MyProject",
    run_name="baseline",
    mlflow=MLflowSpec(
        tracking_uri="sqlite:///.experiments/mlflow.db",
        experiment_name="MyProject",
        log_model=False,
    ),
)

Remote HTTP tracking URIs are also passed through to MLflow. Credentials/service deployment are external to NexuML configuration.

DVCLive¶

LoggingSpec also supports DVCLiveSpec. Install dvclive in the consuming environment when you select that backend.

from nexuml.core.types import DVCLiveSpec

DVCLiveSpec(dir=".experiments/dvclive")

Pipeline diagrams¶

LoggingSpec.diagram controls Mermaid export independently from scalar loggers. See Pipeline diagrams.

No external logger¶

ScenarioSpec.logging=None disables configured NexuML loggers. A plain LoggingSpec() has no TensorBoard/MLflow/DVCLive backend by default, though its diagram configuration is enabled by default.

The optional base library's default_logging() helper intentionally chooses its own convenience defaults; do not confuse those helper defaults with core LoggingSpec defaults.

Paths¶

Relative logging paths are resolved through NEXUML_LOGS_ROOT where the relevant NexuML path helper is used. See Environment roots for the exact rule.