DLLogger for Python

DLLogger - minimal logging tool

This project emerged from the need for a unified logging schema for Deep Learning Example modules. It provides a simple, extensible, and intuitive logging capabilities with API trimmed to an absolute minimum.

Table Of Contents

• Installation
• Quick Start Guide
• Available backends overview
  • StdOutBackend
  • JSONStreamBackend
• Advanced usage
  • Multiple loggers

Installation

To install DLLogger, run (separate commands):

pip install git+https://github.com/NVIDIA/dllogger#egg=dllogger

Quick Start Guide

To start using DLLogger, add the following two lines of code to your DL training script:

from dllogger import StdOutBackend, JSONStreamBackend, Verbosity
import dllogger as DLLogger

DLLogger.init(backends=[
    StdOutBackend(Verbosity.DEFAULT),
    JSONStreamBackend(Verbosity.VERBOSE, "log.json"),
])

To log data, call DLLogger.log(step=<TRAINING_STEP>, data=<DATA>, verbosity=<VERBOSITY>)

Where:

• <TRAINING_STEP> can be any number/string/tuple which would indicate where are we in the training process.

• Use step="PARAMETER" for script parameters (everything that is needed to reproduce the result).

• Use a tuple of numbers to indicate the training progress, for example:

  • step=tuple(epoch_number, iteration_number, validation_iteration_number) for a validation_iteration_number in a validation that happens after iteration iteration_number of epoch epoch_number

  • step=tuple(epoch_number,) for a summary of epoch epoch_number

  • step=tuple() for a summary of the entire training

• <DATA> should be a dictionary with metric names as keys and metric values as values.

To log metric metadata, for example, unit, description, ordering, and format, call DLLogger.metadata(metric_name, metric_metadata) where metric_metadata is a dictionary.

Backends can use the metadata information for logging purposes, for example, StdOutBackend uses the format and unit fields to format its output.

The log is automatically saved on the exit of the Python process (with exception of processes killed with SIGKILL). To flush the log file before training ends, run:

DLLogger.flush()

For usage examples, refer to the examples/dllogger_example.py and examples/dllogger_singleton_example.py files.

Available backends overview

DLLogger can use multiple backends for logging, for example, with one logging call, the logged data can be saved or printed in multiple formats.

StdOutBackend

A vanilla backend that holds no buffers; and that prints the provided values to stdout.

StdOutBackend(verbosity, step_format=..., metric_format=...)

Where:

• step_format is a function that formats step in the DLLogger.log call.

• metric_format is a function that formats a metric name and value given its metadata

For more information, see the default_*_format functions in the dllogger/logger.py script.

For example output, refer to the examples/stdout.txt file.

JSONStreamBackend

This backend saves JSON formatted lines into a file, adding time stamps for each line.

JsonBackend(verbosity, file_name)

For example output, refer to the examples/dummy_resnet_log.json file.

Advanced usage

By default, DLLogger takes advantage of the singleton pattern. It is possible to obtain a logger instance without referencing the DLLogger global instance, for example, to create multiple and different logs from one script.

from dllogger import Logger
logger = Logger(backends=BACKEND_LIST)