Logging and progress¶

nonconform uses standard Python logger levels to decide whether several long-running operations display tqdm progress bars. Raw-score aggregation has a separate ConformalDetector(verbose=...) switch.

Controls at a glance¶

Warnings use the same logger hierarchy. verbose=False does not suppress warnings, and changing the logger level does not disable an aggregation bar explicitly requested with verbose=True.

Set the package level explicitly¶

Configure logging before constructing and fitting detectors:

import logging

package_logger = logging.getLogger("nonconform")
package_logger.setLevel(logging.WARNING)

print(logging.getLevelName(package_logger.level))

Useful package-wide levels are:

• logging.DEBUG for parameter-normalization details;
• logging.INFO for strategy, weighting, and WCS progress;
• logging.WARNING for warnings and errors only; and
• logging.ERROR for errors only.

The effective behavior can also depend on handlers configured by the host application. Libraries should not call logging.basicConfig(...) on behalf of an application.

Complete progress example¶

import logging

import numpy as np
from sklearn.ensemble import IsolationForest

from nonconform import ConformalDetector, CrossValidation

logging.getLogger("nonconform").setLevel(logging.INFO)

rng = np.random.default_rng(42)
x_reference = rng.normal(size=(150, 3))
x_test = rng.normal(size=(20, 3))

detector = ConformalDetector(
    detector=IsolationForest(n_estimators=20, random_state=42),
    strategy=CrossValidation(k=3),
    verbose=True,
    seed=42,
).fit(x_reference)

p_values = detector.compute_p_values(x_test)
print(p_values[:3])

This can show Calibration during the three fold fits and Aggregation while the retained models score x_test.

Configure one subsystem¶

Logger names follow Python's dotted hierarchy:

import logging

logging.getLogger("nonconform").setLevel(logging.WARNING)
logging.getLogger("nonconform.resampling.bootstrap").setLevel(logging.INFO)

print(
    logging.getLevelName(
        logging.getLogger("nonconform.resampling.bootstrap").getEffectiveLevel()
    )
)

The relevant names are:

• nonconform
• nonconform.adapters
• nonconform.resampling.crossval
• nonconform.resampling.bootstrap
• nonconform.weighting.bagged
• nonconform.fdr

Set a child logger explicitly only when its output policy should differ from the package-level policy.

Application logging pattern¶

For scripts, configure a handler once at the application boundary, then set the package level:

import logging

handler = logging.StreamHandler()
handler.setFormatter(
    logging.Formatter("%(levelname)s %(name)s: %(message)s")
)

logger = logging.getLogger("nonconform")
logger.handlers.clear()
logger.addHandler(handler)
logger.setLevel(logging.INFO)
logger.propagate = False

logger.info("nonconform logging configured")

Clearing handlers is appropriate in a standalone script that owns its logging configuration. Do not do it inside reusable library code or a hosted runtime whose handlers belong to the caller.

Progress bars normally write to standard error. Account for that when capturing logs in notebooks, tests, job runners, or container platforms.