Minimal Python command-line parser inspired by Facebook's Hydra + dot-accessible nested dictionaries.
Easily parse arbitrary arguments from the command line without dependencies:
pip install minydraminydra is tested on Python 3.7, 3.8 and 3.9.
Getting Startedย ย โขย Forcing typesย ย โขย MinyDictย ย โขย Save configย ย โขย Prevent typosย ย โขย Use default configsย ย โขย Examples
from minydra.parser import Parser
if __name__ == "__main__":
parser = Parser(
verbose=0, # print received args
allow_overwrites=False, # allow repeating args in the command-line
warn_overwrites=True, # warn repeating args if they are allowed
parse_env=True, # get environment variable
warn_env=True, # warn if an environment variable is specified but not found
defaults=None, # path to a MinyDict-loadable dictionary of default values for the args
strict=True, # if `defaults` is provided, whether to allow new keys in the command-line
# or restrict to `defaults`' keys
keep_special_kwargs=True, # `defaults` and `strict` can be set from the command-line
# with `@defaults=` and `@strict=`. This argument decides if
# you want to keep those keys in the final arguments.
)
args = parser.args.pretty_print().resolve().pretty_print() # notice .resolve() transforms dotted.keys into nested dictsfrom minydra import resolved_args
if __name__ == "__main__":
args = resolved_args()
args.pretty_print()examples/demo.py
examples/demo.json
from minydra import MinyDict, resolved_args
from pathlib import Path
if __name__ == "__main__":
# parse arbitrary args in 1 line
args = resolved_args()
# override default conf
if args.default:
args = MinyDict.from_json(args.default).update(args)
# protect args in the rest of the code execution
args.freeze()
# print the args in a nice orderly fashion
args.pretty_print()
# access args with dot/attribute access
print(f'Using project "{args.log.project}" in {args.log.outdir}')
# save configuration
args.to_json(Path(args.log.outdir) / f"{args.log.project}.json")import minydra
from minydra.dict import MinyDict
@minydra.parse_args(verbose=0, allow_overwrites=False) # Parser's init args work here
def main(args: MinyDict) -> None:
args.resolve().pretty_print()
if __name__ == "__main__":
main()- Simple strings are parsed to
floatandintautomatically. - A single keyword will be interpreted as a positive flag.
- A single keyword starting with
-will be interpreted as a negative flag. - If
parse_envisTrue, environment variables are evaluated.
$ python examples/decorator.py outdir=$HOME/project save -log learning_rate=1e-4 batch_size=64
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ batch_size : 64 โ
โ learning_rate : 0.0001 โ
โ log : False โ
โ outdir : /Users/victor/project โ
โ save : True โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
- dotted keys will be resolved to nested dictionary keys:
$ python examples/decorator.py server.conf.port=8000
โญโโโโโโโโโโโโโโโโโโโโโฎ
โ server โ
โ โconf โ
โ โ โport : 8000 โ
โฐโโโโโโโโโโโโโโโโโโโโโฏ
- Using
ast.literal_eval(value),minydrawill try and parse more complex values for arguments as lists or dicts. Those should be specified as strings:
$ python examples/decorator.py layers="[1, 2, 3]" norms="{'conv': 'batch', 'epsilon': 1e-3}"
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ layers : [1, 2, 3] โ
โ norms : {'conv': 'batch', 'epsilon': 0.001} โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
Adding ___<type> to a key will force this type to the value. Notice how 01 is parsed to an integer 1 but 04 is parsed to a string (as specified) "04", and hello is parsed to a list, not kept as a string
$ python examples/decorator.py n_jobs___str=04 job=01 chips___list=hello
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ chips : ['h', 'e', 'l', 'l', 'o'] โ
โ job : 1 โ
โ n_jobs : 04 โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
Known types are defined in Parser.known_types and the separator (___) in Parser.type_separator
In [1]: from minydra import Parser
In [2]: Parser.known_types
Out[2]: {'bool', 'float', 'int', 'str'}
In [3]: Parser.type_separator
Out[3]: '___'You can configure the Parser from the command-line using special @ arguments. In other words, all __init__(self, ...) arguments can be set from the command-line with @argname=new_value.
In particular if you run python examples/decorator.py @defaults=./examples/demo.json you will see:
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ @defaults : ./examples/demo.json โ
โ log โ
โ โlogger โ
โ โ โlog_level : DEBUG โ
โ โ โlogger_name : minydra โ
โ โoutdir : /some/path โ
โ โproject : demo โ
โ verbose : False โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
But if you add @strict=false @keep_special_kwargs=false you will now have:
$ python examples/decorator.py @defaults=./examples/demo.json @strict=false @keep_special_kwargs=false
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ log โ
โ โlogger โ
โ โ โlog_level : DEBUG โ
โ โ โlogger_name : minydra โ
โ โoutdir : /some/path โ
โ โproject : demo โ
โ verbose : False โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
(you need to have @strict=false since @keep_special_kwargs is unknown in demo.json. It would not be the case if strict=false had been used in the script itself (but it can be overridden from the command-line!))
Minydra's args are a custom lightweight wrapper around native dict which allows for dot access (args.key), resolving dotted keys into nested dicts and pretty printing sorted keys in a box with nested dicts indented. If a key does not exist, it will not fail, rather return None (as dict.get(key, None)).
a MinyDict inherits from dict so usual methods work .keys(), .items() etc.
In [1]: from minydra.dict import MinyDict
In [2]: args = MinyDict({"foo": "bar", "yes.no.maybe": "idontknow"}).pretty_print(); args
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ foo : bar โ
โ yes.no.maybe : idontknow โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
Out[2]: {'foo': 'bar', 'yes.no.maybe': 'idontknow'}
In [3]: args.resolve().pretty_print(); args
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ foo : bar โ
โ yes โ
โ โno โ
โ โ โmaybe : idontknow โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
Out[3]: {'foo': 'bar', 'yes': {'no': {'maybe': 'idontknow'}}}
In [4]: args.yes.no.maybe
Out[4]: "idontknow"
In [5]: "foo" in args
Out[5]: True
In [6]: "rick" in args
Out[6]: False
In [7]: args.morty is None
Out[7]: True
In [8]: args.items()
Out[8]: dict_items([('foo', 'bar'), ('yes', {'no': {'maybe': 'idontknow'}})])You can save and read MinyDict to/from disk in 3 formats: json and pickle without dependencies, yaml with the PyYAML dependency (pip install minydra[yaml]).
Methods to_pickle, to_json and to_yaml have 3 arguments:
file_pathas astrorpathlib.Pathwhich is resolved:- expand env variable (
$MYDIRfor instance) - expand user (
~) - make absolute
- expand env variable (
return_pathwhich defaults toTrue. If those methods return the path of the created fileallow_overwriteswhich defaults toTrue. IfFalseandpathexists, aFileExistsErrorwill be raised. Otherwise creates/overwrites the file atfile_pathverbosewhich defaults to0. If>0prints the path of the created object
Note:
to/from_yamlwill fail with aModuleNotFoundErrorifPyYAMLis not installed.- the
jsonstandard does not accept ints as keys in dictionaries so{3: 2}would be dumped -- and therefore loaded -- as{"3": 2}.
In [1]: from minydra.dict import MinyDict
In [2]: args = MinyDict({"foo": "bar", "yes.no.maybe": "idontknow"}).resolve(); args
Out[2]: {'foo': 'bar', 'yes': {'no': {'maybe': 'idontknow'}}}
In [3]: json_file_path = args.to_json("./args.json")
In [4]: yaml_file_path = args.to_yaml("./args.yaml")
In [5]: pkl_file_path = args.to_pickle("./args.pkl")
In [6]: _ = args.to_json("./args.json", verbose=1) # verbose argument prints the path
Json dumped to: /Users/victor/Documents/Github/vict0rsch/minydra/args.json
In [7]: MinyDict.from_json("args.json")
Out[7]: {'foo': 'bar', 'yes': {'no': {'maybe': 'idontknow'}}}
In [8]: assert (
MinyDict.from_yaml(yaml_file_path)
== MinyDict.from_json(json_file_path)
== MinyDict.from_pickle(pkl_file_path)
== args
)python examples/dumps.py path="./myargs.pkl" format=pickle cleanup
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ cleanup : True โ
โ format : pickle โ
โ path : ./myargs.pkl โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
Dumped args to /Users/victor/Documents/Github/vict0rsch/minydra/myargs.pkl
Cleaning up
To prevent typos from the command-line, the MinyDict.update method has a strict mode: updating a MinyDict with another one using strict=True will raise a KeyError if the key does not already exist:
from minydra import MinyDict, resolved_args
if __name__ == "__main__":
# parse arbitrary args in 1 line
args = resolved_args()
# override default conf
if args.default:
path = args.default
# delete otherwise it will be used to update the conf which does not have
# "default" as a key, therefore raising a KeyError in strict mode
del args.default
args = MinyDict.from_json(path).update(args, strict=True)
args.pretty_print()No typo:
$ python examples/strict.py default=./examples/demo.json log.logger.log_level=INFO
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ log โ
โ โlogger โ
โ โ โlog_level : INFO โ
โ โ โlogger_name : minydra โ
โ โoutdir : /some/path โ
โ โproject : demo โ
โ verbose : False โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
Typo:
$ python examples/strict.py default=./examples/demo.json log.logger.log_leveel=INFO
Traceback (most recent call last):
File "/Users/victor/Documents/Github/vict0rsch/minydra/examples/strict.py", line 13, in <module>
args = MinyDict.from_json(path).update(args, strict=True)
File "/Users/victor/Documents/Github/vict0rsch/minydra/minydra/dict.py", line 111, in update
self[k].update(v, strict=strict)
File "/Users/victor/Documents/Github/vict0rsch/minydra/minydra/dict.py", line 111, in update
self[k].update(v, strict=strict)
File "/Users/victor/Documents/Github/vict0rsch/minydra/minydra/dict.py", line 100, in update
raise KeyError(
KeyError: 'Cannot create a non-existing key in strict mode ({"log_leveel":INFO}).'
The minydra.Parser class takes a defaults= keyword argument. This can be:
- a
stror apathlib.Pathto ajsonyamlorpicklefile thatminydra.MinyDictcan load (from_X) - a
dictor aminydra.MinyDict - a
listof the above types, in which case the resulting defaults will be the result of sequential updates from those defaults, enabling hierarchical defaults (first defaults are the starting point, then each subsequent defaults updates it)
When defaults is provided, the resulting minydra.MinyDict serves as a reference for the arguments parsed from the command-line:
- If you setup the parser with
strict=True, arguments from the command-line will still have a higher priority but they will have to be present in thedefaultsto prevent typos or unknown arguments (see strict mode) - arguments not present in the command-line with fallback to values in
defaults
defaults can actually be a list and the update order is the same as the list's. For instance:
In [1]: from minydra import Parser
In [2]: Parser(defaults=["./examples/demo.json", "./examples/demo2.json"]).args.pretty_print();
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ log โ
โ โlogger โ
โ โ โlog_level : INFO โ
โ โ โlogger_name : minydra โ
โ โoutdir : /some/other/path โ
โ โproject : demo โ
โ new_key : 3 โ
โ verbose : False โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏIf you need to set defaults from the command-line, there's a special @defaults keyword you can use:
$ python examples/decorator.py @defaults=./examples/demo.json
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ @defaults : ./examples/demo.json โ
โ log โ
โ โlogger โ
โ โ โlog_level : DEBUG โ
โ โ โlogger_name : minydra โ
โ โoutdir : /some/path โ
โ โproject : demo โ
โ verbose : False โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
$ python examples/decorator.py @defaults="['./examples/demo.json', './examples/demo2.json']"
โญโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ @defaults : ['./examples/demo.json', './examples/demo2.json'] โ
โ log โ
โ โlogger โ
โ โ โlog_level : INFO โ
โ โ โlogger_name : minydra โ
โ โoutdir : /some/other/path โ
โ โproject : demo โ
โ new_key : 3 โ
โ verbose : False โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
Prints the MinyDict in a box, with dicts properly indented. A few arguments:
indents, which defaults to2: the amount of indentation for nested dictionariessort_keys, which defaults toTrue: whether or not to alphabetically sort the keys before printing
To produce a native Python dict, use args.to_dict()
MinyDict's methods (including the dict class's) are protected, they are read-only and you cannot therefore set attributes with there names, like args.get = 2. If you do need to have a get argument, you can access it through items: args["get"] = 2.
Try with examples/protected.py:
python examples/protected.py server.conf.port=8000 get=3
โญโโโโโโโโโโโโโโโโโโโโโฎ
โ get : 3 โ
โ server โ
โ โconf โ
โ โ โport : 8000 โ
โฐโโโโโโโโโโโโโโโโโโโโโฏ
<built-in method get of MinyDict object at 0x100ccd4a0>
3
dict_items([('get', 3), ('server', {'conf': {'port': 8000}})])
{'conf': {'port': 8000}}
Run tests and pre-commit checks (isort, black, flake8) with
$ pip install -r requirements-test.txt
$ pre-commit run --all-files
$ pytest -vv --cov=minydra tests/
