Importing TensorFlow causes argparse_flags.ArgumentParser to fail when specifying an argument_default

Basically just the title. This code fails:

import tensorflow as tf
from absl.flags.argparse_flags import ArgumentParser

ArgumentParser(argument_default=42)

Here it is in Colab.

Edit: I should clarify, it only fails if tensorflow is imported

Thanks for filing the bug. Looks like argument_default= simply doesn't work if you have any absl.flags flag defined. This is because a default= argument is passed to custom actions when argument_default= is supplied, but our _FlagAction and _BooleanFlagAction don't support it. Since absl.flags use dest=argparse.SUPPRESS and the default value is never seen through the argparse namespace, we can simply add a default= argument and ignore it.

Is there a workaround?

There isn't a good workaround before we put out a fix, but you could monkey patch these methods to have an extra default= argument:

abseil-py/absl/flags/argparse_flags.py

Line 240 in 4d8ea27

    
           def __init__(self, option_strings, dest, help, metavar, flag_instance):  # pylint: disable=redefined-builtin

abseil-py/absl/flags/argparse_flags.py

Line 267 in 4d8ea27

    
           def __init__(self, option_strings, dest, help, metavar, flag_instance):  # pylint: disable=redefined-builtin