In the past, I use plain print() to display some statistics during training process of my models. However, it is not convenient to save the statistics in a file for later inspection. Once the program is finished, we do not have an overview of the training process. I am aware that there are packages such as tensorboardX and visdom, which are specifically designed for inspecting the various statistics during training. Right now, I would rather keep the program plain simple and usable. So I decided to give Python logging pacakge a try.

If we want to display messages on the terminal and save to a file, we can use the following setting:

import logging

logging.basicConfig(
    level=logging.INFO,
    datefmt="%Y-%m-%d %H:%M:%S",
    format="%(asctime)s %(message)s",
    handlers=[
        logging.FileHandler("{}/{}".format(log_path, log_name)),
        logging.StreamHandler()
    ]
)

The datefmt parameter is used to format the time string in the logging message. With the above format, the time is shown as something like the following:

2018-10-15 09:44:49 (…other message…)

The format string uses the standard Python time format. You can find the full list of format directives here.

The format paramter is used to set the format of the output message, i.e., what to show in the logging message. We can show time, current file name, process name etc. A list of predefined attributes are (excerpted from the logging documentation):

Click to see the doc.
%(name)s            Name of the logger (logging channel)
%(levelno)s         Numeric logging level for the message (DEBUG, INFO,
                    WARNING, ERROR, CRITICAL)
%(levelname)s       Text logging level for the message ("DEBUG", "INFO",
                    "WARNING", "ERROR", "CRITICAL")
%(pathname)s        Full pathname of the source file where the logging
                    call was issued (if available)
%(filename)s        Filename portion of pathname
%(module)s          Module (name portion of filename)
%(lineno)d          Source line number where the logging call was issued
                    (if available)
%(funcName)s        Function name
%(created)f         Time when the LogRecord was created (time.time()
                    return value)
%(asctime)s         Textual time when the LogRecord was created
%(msecs)d           Millisecond portion of the creation time
%(relativeCreated)d Time in milliseconds when the LogRecord was created,
                    relative to the time the logging module was loaded
                    (typically at application startup time)
%(thread)d          Thread ID (if available)
%(threadName)s      Thread name (if available)
%(process)d         Process ID (if available)
%(message)s         The result of record.getMessage(), computed just as
                    the record is emitted

The handler parameter is used to set up where logging messages should go. In the above example, we use a FileHandler to save the logging message to a disk file. We use StreamHandler to display messages in the terminal.

After configuring logging, we need a logger:

logger = logging.getLogger(__name__)

In our program, if we need to log something, use logger.info(). An example usage is shown below:

logger.info("Epoch %d, Accuracy: %f, Loss: %f", epoch, accuracy, loss)

The logging module relies on the old Python string formatting style. If you use the new .format() style string formatting. You will see a pylint warning complaining:

pylint: logging-not-lazy / Specify string format arguments as logging function parameters

The above formatting is also slightly different from the old Python formatting style. We do not need to write % between strings and the values. We can just append a list of needed values as the argument for the info method.

How to log exceptions

After catching an exception, we may want to log the exception message using our logger. Initially, I have the following code:

try:
    # my code
except SomeError as e:
    logger.debug("Exception: %s", e)

Unfortunately, this will only print an incomplete excpetion message and does not help to debug the error.

Using str(e) or repr(e) to represent the exception, you also won’t get the actual stack trace, so it is not helpful to find where the exception is.

After reading other answers and the logging package doc, the following two ways work great to print the actual stack trace for later debugging:

  1. use logger.debug() with parameter exc_info:

    try:
        # my code
    except SomeError as e:
        logger.debug(e, exc_info=True)
    
  2. use logger.exception() to print the exception directly:

    try:
        # my code
    except SomeError as e:
        logger.exception(e)
    

References