Source code for

# pylint: disable=eval-used,protected-access
Create Space objects from configuration

Functions which build ``Dimension`` and ``Space`` objects for defining problem's search space.

Replace actual hyperparam values in your script's config files or cmd
arguments with orion's keywords for declaring hyperparameter types
to be optimized.

Motivation for this way of orion's configuration is to achieve as
minimal intrusion to user's workflow as possible by:

   * Offering to user the choice to keep the original way of passing
     hyperparameters to their script, be it through some **config file
     type** (e.g. yaml, json, ini, etc) or through **command line

   * Instead of passing the actual hyperparameter values, use one of
     the characteristic keywords, names enlisted in :scipy.stats:`distributions`
     or :class:``,
     to describe distributions and declare the hyperparameters
     to be optimized. So that a possible command line argument
     like ``-lrate0=0.1`` becomes ``-lrate0~'uniform(-3, 1)'``.

.. note::
   Use ``~`` instead of ``=`` to denote that a variable "draws from"
   a distribution. We support limited Python syntax for describing distributions.

   * Module will also use the script's provided input file/args as a
     template to fill an appropriate input with proposed values for the
     script's execution in each hyperiteration.

import logging
import re
from collections import OrderedDict

from scipy.stats import distributions as sp_dists

from import Categorical, Fidelity, Integer, Real, Space
from orion.core.utils.flatten import flatten

log = logging.getLogger(__name__)

def _check_expr_to_eval(expr):
    if "__" in expr or ";" in expr:
        raise RuntimeError("Cannot use builtins, '__' or ';'. Sorry.")

def _get_arguments(*args, **kwargs):
    return args, kwargs

def _real_or_int(kwargs):
    return Integer if kwargs.pop("discrete", False) else Real

[docs]def replace_key_in_order(odict, key_prev, key_after): """Replace ``key_prev`` of ``OrderedDict`` ``odict`` with ``key_after``, while leaving its value and the rest of the dictionary intact and in the same order. """ tmp = OrderedDict() for k, v in odict.items(): if k == key_prev: tmp[key_after] = v else: tmp[k] = v return tmp
def _should_not_be_built(expression): return expression.startswith("-") or expression.startswith(">") def _remove_marker(expression, marker="+"): return ( expression.replace(marker, "", 1) if expression.startswith(marker) else expression )
[docs]class DimensionBuilder: """Create `Dimension` objects using a name for it and an string expression which encodes prior and dimension information. Basically, one must provide a string like a function call to a method that has the name of a distribution, .e.g. ``alpha``, and then provide settings about that distributions and information about the `Dimension`, if it cannot be inferred. One example for the latter case is: ``uniform(-3, 5)`` will return a :class:`` dimension, while ``uniform(-3, 5, discrete=True)`` will return an :class:`` dimension. Sometimes there is also a separate name for the same distribution in integers, for the 'uniform' example: ``randint(-3, 5)`` will return a uniform :class:`` dimension. For categorical dimensions, one can use either ``enum`` or ``random`` name. ``random`` however can be used for uniform reals or integers as well. Most names are taken from instances contained in :scipy.stats:`distributions`. So, if the distribution you are searching for is there, then `DimensionBuilder` can build one dimension with that prior! Examples -------- >>> dimbuilder = DimensionBuilder() >>>'learning_rate', 'loguniform(0.001, 1, shape=10)') Real(name=learning_rate, prior={reciprocal: (0.001, 1), {}}, shape=(10,)) >>>'something_else', 'poisson(mu=3)') Integer(name=something_else, prior={poisson: (), {'mu': 3}}, shape=()) >>> dim ='other2', 'uniform(-5, 2)') >>> dim Real(name=other2, prior={uniform: (-5, 7), {}}, shape=()) >>> dim.interval() (-5.0, 2.0) """ def __init__(self): """Init of `DimensionBuilder`.""" = None
[docs] def choices(self, *args, **kwargs): """Create a :class:`` dimension.""" name = try: if isinstance(args[0], (dict, list)): return Categorical(name, *args, **kwargs) except IndexError as exc: raise TypeError( f"Parameter '{name}': " "Expected argument with categories." ) from exc return Categorical(name, args, **kwargs)
[docs] def fidelity(self, *args, **kwargs): """Create a :class:`` dimension.""" name = return Fidelity(name, *args, **kwargs)
[docs] def uniform(self, *args, **kwargs): """Create an :class:`` or :class:`` uniformly distributed dimension. .. note:: Changes scipy convention for uniform's arguments. In scipy, ``uniform(a, b)`` means uniform in the interval [a, a+b). Here, it means uniform in the interval [a, b]. """ name = klass = _real_or_int(kwargs) if len(args) == 2: return klass(name, "uniform", args[0], args[1] - args[0], **kwargs) return klass(name, "uniform", *args, **kwargs)
[docs] def randint(self, *args, **kwargs): """Create an :class:`` or :class:`` uniformly distributed dimension. .. note:: Changes scipy convention for uniform's arguments. In scipy, ``uniform(a, b)`` means uniform in the interval [a, a+b). Here, it means uniform in the interval [a, b]. """ raise NotImplementedError( "`randint` is not supported. Use uniform(discrete=True) instead." )
[docs] def gaussian(self, *args, **kwargs): """Synonym for :scipy.stats:`distributions.norm`.""" return self.normal(*args, **kwargs)
[docs] def normal(self, *args, **kwargs): """Another synonym for :scipy.stats:`distributions.norm`.""" name = klass = _real_or_int(kwargs) return klass(name, "norm", *args, **kwargs)
[docs] def loguniform(self, *args, **kwargs): """Return a `Dimension` object with :scipy.stats:`distributions.reciprocal` prior distribution. """ name = klass = _real_or_int(kwargs) return klass(name, "reciprocal", *args, **kwargs)
def _build(self, name, expression): """Build a `Dimension` object using a string as its `name` and another string, `expression`, from configuration as a "function" to create it. """ = name _check_expr_to_eval(expression) prior, arg_string = re.findall(r"([a-z][a-z0-9_]*)\((.*)\)", expression)[0] globals_ = {"__builtins__": {}} import numpy as np globals_["np"] = np try: dimension = eval("self." + expression, globals_, {"self": self}) return dimension except AttributeError: pass # If not found in the methods of `DimensionBuilder`. # try to see if it is legit scipy stuff and call a `Dimension` # appropriately. args, kwargs = eval( "_get_arguments(" + arg_string + ")", globals_, {"_get_arguments": _get_arguments}, ) if hasattr(sp_dists._continuous_distns, prior): klass = _real_or_int(kwargs) elif hasattr(sp_dists._discrete_distns, prior): klass = Integer else: raise TypeError( f"Parameter '{name}': " f"'{prior}' does not correspond to a supported distribution." ) dimension = klass(name, prior, *args, **kwargs) return dimension
[docs] def build(self, name, expression): """Check ``DimensionBuilder._build`` for documentation. .. note:: Warm-up: Fail early if arguments make object not usable. """ try: dimension = self._build(name, expression) except ValueError as exc: raise TypeError(f"Parameter '{name}': Incorrect arguments.") from exc except IndexError as exc: error_msg = ( f"Parameter '{name}': Please provide a valid form for prior:\n" "'distribution(*args, **kwargs)'\n" f"Provided: '{expression}'" ) raise TypeError(error_msg) from exc try: dimension.sample() except TypeError as exc: error_msg = ( f"Parameter '{name}': " f"Incorrect arguments for distribution '{dimension._prior_name}'.\n" f"Scipy Docs::\n\n{dimension.prior.__doc__}" ) raise TypeError(error_msg) from exc except ValueError as exc: raise TypeError(f"Parameter '{name}': Incorrect arguments.") from exc return dimension
[docs]class SpaceBuilder: """Build a :class:`` object form user's configuration.""" def __init__(self): self.dimbuilder = DimensionBuilder() = None self.commands_tmpl = None self.converter = None self.parser = None
[docs] def build(self, configuration): """Create a definition of the problem's search space. Using information from the user's script configuration (if provided) and the command line arguments, will create a :class:`` object defining the problem's search space. Parameters ---------- configuration: OrderedDict An OrderedDict containing the name and the expression of the parameters. Returns ------- :class:`` The problem's search space definition. """ = Space() for namespace, expression in flatten(configuration).items(): if _should_not_be_built(expression): continue expression = _remove_marker(expression) dimension =, expression) try: except ValueError as exc: error_msg = f"Conflict for name '{namespace}' in parameters" raise ValueError(error_msg) from exc return
[docs] def build_to(self, config_path, trial, experiment=None): """Create the configuration for the user's script. Using the configuration parser, create the commandline associated with the user's script while replacing the correct instances of parameter distributions by their actual values. If needed, the parser will also create a configuration file. Parameters ---------- config_path: str Path in which the configuration file instance will be created. trial: `orion.core.worker.trial.Trial` Object with concrete parameter values for the defined :class:``. experiment: :class:`orion.core.worker.experiment.Experiment`, optional Object with information related to the current experiment. Returns ------- list The commandline arguments that must be given to script for execution. """ return self.parser.format(config_path, trial, experiment)