nnlogging2

status-badge PyPI - Version PyPI - Downloads PyPI - Python Version

nnlogging2 is a non-invasive extension of the standard Python logging library. nnlogging2 is designed to track numerical metrics by time and execution context.

• Metric Tracking: Fast automatic metric statistics tracking.
• Extensibility: Fork metric, logger, handler or formatter at any time.
• Compatibility: Compatible with standard python logging library and third-party modules.

Installation

pip install nnlogging2

Usage

For open-box usage, nnlogging2 provides:

• MetricLogger (subclass of logging.Logger);
• ContextedMetricLogger (subclass of logging.LoggerAdapter);
• AttyStreamHandler (subclass of logging.StreamHandler);
• ContextedMetricFormatter (subclass of logging.Formatter).

A quick preview script can be found in the Preview Script subsection.

FallbackProtectorMeta

Despite designed classes mostly inherited from standard logging classes, they also chained with a custom metaclass FallbackProtectorMeta. FallbackProtectorMeta is designed to turn the default class attribute overwriting behavior into modify the default value (fallback).

An escaping window has been left for future usage:

FPM = FallbackProtectorMeta
class Cls(metaclass=FPM):
 attr = Fallback(1)
assert Cls.attr == 1
assert isinstance(Cls.__dict__["attr"], Fallback)
Cls.attr = 2
assert Cls.attr == 2
assert isinstance(Cls.__dict__["attr"], Fallback) # `FPM` prevents overwriting
del Cls.attr # delete class attribute to fully remove `Fallback` instance
Cls.attr = 3
assert Cls.attr == 3
assert isinstance(Cls.__dict__["attr"], int)

FallbackProtectorMeta allows different instances shares attribute with the same name but not necessarily same value. This is typical when one want different formatters responding to different record keys, or different loggers can be manipulated in different modules without affecting each other.

MetricLogger

Attributes:

• extra_key (fallbackable): Key to be injected into logging record.
• metric_factory (fallbackable): Factory used when tracer created.
• metric_attrs (fallbackable): Attributes this logger tracing.
• tracer: Tracer to trace metrics.

Note:

• Although metric_attrs is default as ("val",), nnlogging2 provides default metric class with the following attributes: median, mean, max, lastval, and val. For further extension, val is the only attribute you need to reimplement.
• metric_attrs is encouraged to record necessary attributes cause the access implementation is currenly running in loops and the time cost will grow linearily.
• tracer should be replaced manually, and the data recorded will not automatically preserved.

ContextedMetricLogger

ContextedMetricLogger is designed to inherit logging.LoggerAdapter since contexts are not necessary when tracing metrics and can be replaced by custom formatters. Despite many reasons, we still provide this class to make open-box usage happy.

Attributes:

• extra_key (fallbackable): Key to be injected into logging record.
• context: the contexts to be injected into logging record.
• cover: whether to cover contexts existed or not.

AttyStreamHandler

Because ANSI escaped code print nothing but hold length size when formatting, please increase 9 character length, e.g. '%(levelname)-8s' -> '%(levelname)-17s', when you know ANSI escaped codes will work.

Attributes:

• colored_level: Whether to show a colored level or not.

ContextedMetricFormatter

This formatter extracts context and metric data from the LogRecord, formats them according to the specified format strings, and temporarily injects them into the record (default as context_ext and metrics_ext attributes) so they can be used in the main log message format string.

Attributes:

• context_fmtstr (fallbackable): Context formatting string.
• context_delim (fallbackable): Context delimiter.
• context_key (fallbackable): Formatted context string entry.
• metrics_fmtstr (fallbackable): Metrics formatting string.
• metrics_delim (fallbackable): Metrics delimiter.
• metrics_key (fallbackable): Formatted metrics string entry.

Preview Script

from nnlogging2 import get_metric_logger, ContextedMetricFormatter, AttyStreamHandler
from nnlogging2 import helpers as hp
# 1. Initialize the promoted logger
logger = get_metric_logger("train")
logger.metric_attrs = ("lastval", "mean", "median")
logger.setLevel(hp.level_to_int("INFO"))
# 2. Setup formatting for context and metrics
# Inject "%(context_ext)s" and "%(metrics_ext)s" to formatter
# (8 showing characters, 9 ANSI escaped codes)
formatter = ContextedMetricFormatter(
 fmt="%(levelname)-17s%(context_ext)s%(message)s%(metrics_ext)s"
)
handler = AttyStreamHandler()
handler.setFormatter(formatter)
logger.addHandler(handler)
# 3. Attach context to logger (optional)
ctx_logger = logger.contexted({"epoch": "E1/1"})
# 4. Standard logging
ctx_logger.debug("debug ...")
ctx_logger.info("info ...")
ctx_logger.warning("warning ...")
ctx_logger.error("error ...")
ctx_logger.critical("critical ...")
# 5. Update & Logging
for i in range(3):
 # Update numerical values (automatically tracked in a moving window)
 ctx_logger.update(loss=0.5 / (i + 1), accuracy=0.8 + (i * 0.05))
 ctx_logger.log_metric(
 hp.level_to_int("INFO"),
 metrics=[
 ("loss", "[%(metricname)s: %(lastval).2f (%(mean).2f)]"),
 ("accuracy", "[%(metricname)s: %(median).2f]"),
 ],
 )

Preview

assets/preview.png

License

This project is licensed under the MIT License LICENSE.