Improved Decoding Mechanism :microscope:

Improved decoding mechanism to support inherited types, one can still use the existing register functionality

pyrallis.decode.register(SomeClass, lambda x: SomeClass(x))

But can also register parent classes as well

pyrallis.decode.register(BaseClass, lambda t, x: t(x), include_subclasses=True)

Also revised the dispatch mechanism from singledispatch to a new dedicated wrapper (singledispatch tries to dispatch by class instead of value which isn't suited for decoding, but worked only thanks to the default behavior)

Added pyrallis.set_config_type to support json and toml

Added an option to change the configuration format used by pyrallis globally, or for a specific context.

This will affect pyrallis.parse, pyrallis.dump and pyrallis.load to use the desired format

import pyrallis
pyrallis.set_config_type('json')
# or
with pyrallis.config_type('json'):
        pyrallis.dump(cfg)

Changed enum parsing mechanism to be name based + bug fixes

Changed enum parsing to be based on enum field names instead of values.

As discussed in #8 , this allows one to support enums with non-string values such as

class LRMethod1(Enum):
    onecycle: torch.optim.lr_scheduler = OneCycleLR
    lambdalr: torch.optim.lr_scheduler = LambdaLR

Instead of allowing only enums of the form:

class LRMethod2(Enum):
    onecycle: str = "onecycle"
    constant: str = "constant"

The release also includes:

• Minor improvement in error mechanism
• Bug fixes when using the default collections such as Dict and List

Added omit_defaults in dump and improved --help

Implemented the suggestion from issue #1 to add an omit_defaults option to the dump function

pyrallis.dump(cfg, omit_defaults=True)

Also simplified the wrapper classes in the process and improved the --help command to be more concise.

Modified API with pyrallis.parse and --config_path

Two breaking API changes

1. Added the new pyrallis.parse call instead of the pyrallis.ArgumentParser.parse_args()

cfg = pyrallis.parse(config_class=TrainConfig)

2. Modified the configuration argument name from --CONFIG to --config_path

$ python train_model.py --config_path=some_config.yaml --exp_name=my_first_exp

Required Arguments Bug Fix

Fixed a bug with nested dataclasses, where if the dataclass had a required argument the dataclass wrapper could not initialize the dataclass to define its default value and raised an exception.

@dataclass
class LogConfig:
    exp_dir: str

@dataclass
class TrainConfig:
    log: LogConfig= field(default_factory=LogConfig)

Type Hint Fixes

Fixed type hints in pyrallis.decode and pyrallis.ArgumentParser for better code analysis

Improved Error Mechanism

Improved error mechanism to be more informative, showing the field that caused the exception

$  python -m examples.demo --compute.workers=five

pyrallis.utils.ParsingError: Failed when parsing value='five' into field "<class '__main__.ComputeConfig'>.workers" of type <class 'int'>.
        Underlying error: invalid literal for int() with base 10: 'five'

Bug fixes for different python versions

Fixed two bugs found when testing different python versions:

• Started using collections.abc as expected (python 3.10)
• Fixed a bug where assumed .name field does not exist (python 3.6)

Updates for PyPi Publication

Updated readme to use raw resources so that PyPi renders them correctly