HfArgumentParser defaults booleans to on

Environment info

• transformers version: 4.9.0.dev0 (initially discovered on 4.8.1)
• Platform: macOS-11.4-x86_64-i386-64bit
• Python version: 3.9.5
• PyTorch version (GPU?): not installed (NA)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed

Who can help

@sgugger

Information

HfArgumentParser when used on a dataclass with a bool field with no default turns the bool on unless it’s supplied with --<field_name> False or similar “false-y” value. I would expect that as the field has no default it should be false unless --<field_name> or --<field_name> True is supplied. This is a behaviour change from v3.0.0 where the booleans are parsed correctly as we’re looking at upgrading and this issue hit us.

To reproduce

Steps to reproduce the behavior:

1. Define a dataclass with a boolean field
2. Supply a list of arguments which does not include that field name
3. The field is turned on.

Appending this snippet to the bottom of test_basic at line 110 in test_hf_argparser.py fails the test.

        args = ["--foo", "1", "--baz", "quux", "--bar", "0.5"]

        example, = parser.parse_args_into_dataclasses(args, look_for_args_file=False)

        self.assertFalse(example.flag)

Extending args with ["--flag","False"] recovers the expected behaviour.

Expected behavior

The boolean should be set to false if the argument is not passed in.

[sgugger]

Yes, it does look like a bug. Fix is quite simple, I can make a PR.

[Craigacp]

Ok, I’ll check it tomorrow against our internal use case to make sure it fixes that too.