Importing tensorflow_hub interferes with python logging

[matteby]

I've run into an issue where importing tensorflow_hub interferes with and prevents python logging.

For example, this script runs fine:

#!/usr/bin/env python3
import argparse
import logging
import time

import tensorflow as tf
#import tensorflow_hub as hub

logger = logging.getLogger(__name__)
    
parser = argparse.ArgumentParser()
parser.add_argument('--version', action='version', version='1.0.0')
subparsers = parser.add_subparsers()

subparser = subparsers.add_parser('train', help='train the Is Data Plate model')
subparser.add_argument('--batch-size', help='training batch size', type=int)
subparser.add_argument('--validation-batch-size', help='validation batch size', type=int)

if __name__ == '__main__':
    log_fmt = '%(asctime)s - %(name)s - %(levelname)s - %(message)s'
    logging.basicConfig(level=logging.INFO, format=log_fmt)
    args = parser.parse_args()

    logger.info("Processing...")

    logger.info(f'Done - elapsed time: {time.process_time()} seconds')

and produces the following output:

$ ./tf-hub-logging.py train
2019-03-28 17:15:31,635 - __main__ - INFO - Processing...
2019-03-28 17:15:31,635 - __main__ - INFO - Done - elapsed time: 2.383287 seconds

However, when tensorflow_hub is imported as in this script:

#!/usr/bin/env python3
import argparse
import logging
import time

import tensorflow as tf
import tensorflow_hub as hub

logger = logging.getLogger(__name__)
    
parser = argparse.ArgumentParser()
parser.add_argument('--version', action='version', version='1.0.0')
subparsers = parser.add_subparsers()

subparser = subparsers.add_parser('train', help='train the Is Data Plate model')
subparser.add_argument('--batch-size', help='training batch size', type=int)
subparser.add_argument('--validation-batch-size', help='validation batch size', type=int)

if __name__ == '__main__':
    log_fmt = '%(asctime)s - %(name)s - %(levelname)s - %(message)s'
    logging.basicConfig(level=logging.INFO, format=log_fmt)
    args = parser.parse_args()

    logger.info("Processing...")

    logger.info(f'Done - elapsed time: {time.process_time()} seconds')

The python logger is broken and produces the following output which displays a warning and also causes lost logging messages:

$ ./tf-hub-logging.py train
WARNING: Logging before flag parsing goes to stderr.
W0328 17:15:19.787199 4598539712 __init__.py:56] Some hub symbols are not available because TensorFlow version is less than 1.14

Has anyone run into a similar issue?

[arnoegw]

Hi matteby. I can reproduce your problem. The crucial point is that import tensorflow_hub triggers from absl import logging, which in turn fiddles with logging.root.handlers; see
absl/logging/__init__.py:1151.

Here is a minimal example:

import logging
# from absl import logging as absl_logging  # Uncomment to trigger the issue.
logger = logging.getLogger("yourlogger")
logging.basicConfig(level=logging.INFO)
logger.info("Hello, world")

This is a problem with absl.logging and cannot be fixed at the level of tensorflow_hub. We do as all of TensorFlow does for version 2.0 and switch from tf.logging to absl.logging (see tensorflow.org/alpha/guide/effective_tf2).

Please report at https://github.com/abseil/abseil-py/issues