cfgopt 0.7.4

You only configure your deep learning experiment once with cfgopt.

Introduction

cfgopt is a solution for elegance of configuring complicated deep learning projects. Here, you configure everything only once in json, without need to write configuration and commandline parsing boilerplate code in python anymore. The solution is extremly simple, elegant, and compact. Without any 3rd-party dependency, the core implementation of cfgopt is only around 200 LOC, and will be (todo) fully tested.

News

• 2022/06/21 cfgopt code released.

Getting Started

Installation

pip install cfgopt

Basic Usage

Generally, you should have some json config files in a folder, and a python program that import cfgopt to parse them. The program can have command line options that update specific field in the parsed results for this running. This is also why this library is called "cfgopt".

To be concrete, you have two choices using cfgopt:

• Use the cfgopt.parse_configs(cfg_root) API, just pass in the directory path of all the configuration files, and explore with the results! Or you should read the examples described in the Features section.
• Use cfgoptrun command directly or wrap cfgopt.main() function, for details please refer to the follow-up main function section.

Features

basic json binding

cfgopt has a main api parse_configs(cfg_root) that accepts the path of a folder of json files. All files will be (recursively) load in python and added to a root dict, keeping the hierarchy and data types unchanged, with some exceptions described in follow-up sections.

Feature first added in v0.1.

block-level reference

cfgopt release you from repeating yourself with a mountain pile of configuration files by borrowing the concept of block-level reference and embedding in many modern note-taking apps such as Logseq, or more well-known hyperlinks for webpages.

The cfg:// format URI:

cfgopt follows and expands string values matching a special syntax cfg://<file-path>/<intra-file-uri> in the configuration files during parsing. This is one of the most repealing feature for cfgopt. You can also specify relative uri which contains no substring ".json" in it.

Example:

file structure

.
└── test_blockref_in_list
    ├── cfg
    │   ├── data.json
    │   └── recipes.json
    └── test_blockref_in_list.py

data.json:

{
    "data1": {
        "meta": {
            "location": "/data/1/loc"
        }
    },
    "data2": {
        "meta": {
            "location": "/data/2/loc"
        }
    }
}

recipes.json:

{
    "recipe1": {
        "use_data": [
            "cfg://data.json/data1"
        ]
    },
    "recipe2": {
        "use_data": [
            "cfg://data.json/data1",
            "cfg://data.json/data2"
        ]
    }
}

test_blockref_in_list.py

import cfgopt

def test_blockref_in_list():
    cfg = cfgopt.parse_configs(cfg_root='test_blockref_in_list/cfg')

    # following lines are equivalent
    assert cfg["recipes.json"]["recipe2"]["use_data"][1]["meta"]["location"] == "/data/2/loc"
    assert cfg["recipes.json"]["recipe2/use_data/1/meta/location"] == "/data/2/loc"
    assert cfg["recipes.json/recipe2/use_data/1/meta/location"] == "/data/2/loc"

__hostname__ in URI will be translated into what you get by executing hostname in shell. Since v0.5.5. Relative URI support added in v0.3.0. Feature first added in v0.1.

command-line update

cfgopt will automatically parse command line options matching the python regex format ^--(.*?)=(.*), and interpret it as an update of the parsed configuration folder. The right hand side of = should be valid json, and you might need to take care of shell escaping your special characters.

For example, you can write --train.json/max_epochs=100 or --train.json/max_epochs="100", since your shell would escape double-quotes, you will get an integer 100 in both cases.

But if you write --train.json/resume=\"/path/to/ckpt\" or --train.json/resume='"/path/to/ckpt"', you should probably get a string value, which depends on your shell implementation.

Unmatched uri/keys are treated as errors since v0.7.0. Regex format updated as ^--(.*?)=(.*) to match the first = symbol since v0.7.2.

Behavior changed to update existing uri before expansion and then update thereafter, this makes both features, i.e., command-line specifying expansion and assigning values after expansion, work. Since v0.5.0.

Updated regex format from ^--(.*)=(.*) to ^--(.*\.json.*)=(.*) in v0.2. But then changed back since v0.4.3, be careful that unmatched uri/keys would just be ignored without warning.

Feature first added in v0.1.

json objects inheritance

Users can specify a json dict, that contains a __base__ field, linking to another base json dict objects with the cfg:// reference format (described in block-level reference). Then the current dict would inherit the base object, and also has its own values in normal fields. This feature is also critical to eliminate repeating, with which now you can develop multiple simillar configs from some prototypes.

TODO: an example.

Bugfix: changed from referencing to deepcopying the base dict in v0.2.

Feature first added in v0.1.

parse python objects

cfgopt has a extremly flexible feature, that parse an json dict to ALMOST any python objects defined in user's code or any code python can find in PYTHON_PATH.

Users can specify a json dict, that contains __module__ and __class__ field. The __module__ field will be imported by importlib.import_module() during parsing, and __class__ field naming any python callable in the imported module will be passed to functools.partial() along with other fields as keyword arguments. Finally, the mapped "dict" in python would be directly callable to instantiate corresponding class or get result of corresponding functions.

pseudo-code of parsing:

module = importlib.import_module(data["__module__"])
klass = getattr(module, data["__class__"])
data["__class__"] = partial(klass, **{k:v for k, v in data.items() if not k.startswith("__")})

NOTE: The python object should not use VAR_POSITIONAL and POSITIONAL_ONLY arguments, and keywords arguments are always recommended than positional arguments.

By default, when arguments has nested object that is meant to be constructed by cfgopt and you call a higher level object, the nested ones in arguments will be automatically and recursively instantiated as long as all its required arguments are defined. If you want certain object not to be automatically created, add a "__as_type__": true field alongside the __class__ field.

Let us refer to these "dict"s with __class__ and __module__ fields as python object builders. You can call into these builders with extra arguments to instantiate them. You can also pass only partial arguments for multiple times, and it won't actually instantiate until all required arguments are ready, similar to above. This is very handy to pass around the builder everywhere deep into the code to collect scattered arguments. For example define a bachnorm builder, and pass it as argument to your networks, the latter can pass the channel argument when it is figured out from the preceding layer. The builder can be used multiple times.

You can also create a builder from python code, by using cfgopt.PartialClass() function, very similar to standard functools.partial(), but again it allows being called multiple times until all required args are given.

TODO: an example.

__as_type__ keyword added in v0.5.3.

Important Feature: PartialClass and lazily instantiation since v0.5.0.

Supported auto instantiate of nested python objects since v0.4.0.

API enhance: user now can directly call the mapped "dict" object instead of its __class__ field in v0.2.

Feature first added in v0.1.

main function

import argparse
import cfgopt

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "recipe",
        help="a main function uri to be execute."
    )
    parser.add_argument(
        "-d", "--cfgdir",
        default="cfg",
        help="config directory that maps to cfg:// root. (default: `cfg`)"
    )
    args, unknown_args = parser.parse_known_args()
    cfgs = cfgopt.parse_configs(
        cfg_root=args.cfgdir,
        args=unknown_args,
        args_root=args.recipe,
    )
    _main = cfgs[args.recipe]
    return _main(recursive=False)

if __name__ == "__main__":
    main()

This example main() function accepts the first argument as a previous described cfg://-format uri, that parses a python callable function, and call it as the main function. This is also implemented as cfgopt.main(), and user can use cfgoptrun command direct from shell, or just import this main function and call it in usercode with extra python arguments.

TODO: an example.

Behaviour changed: cfgoptrun interprets commandline arguments as relative path to recipe instead of absolute(full) uri since v0.6.0.

Add cfgoptrun command (entrypoint) in v0.2.

Feature first added in v0.1.