# pylint: disable=too-many-lines
"""
Perform transformations on Dimensions
=====================================
Provide functions and classes to build a Space which an algorithm can operate on.
"""
import copy
import functools
import itertools
from abc import ABCMeta, abstractmethod
import numpy
from orion.algo.space import Categorical, Dimension, Fidelity, Integer, Real, Space
from orion.core.utils import format_trials
from orion.core.utils.flatten import flatten
from orion.core.worker.trial import Trial
NON_LINEAR = ["loguniform", "reciprocal"]
# pylint: disable=unused-argument
@build_transform.register(Categorical)
def _(dim, type_requirement, dist_requirement):
transformers = []
if type_requirement == "real":
transformers.extend(
[Enumerate(dim.categories), OneHotEncode(len(dim.categories))]
)
elif type_requirement in ["integer", "numerical"]:
transformers.append(Enumerate(dim.categories))
return transformers
@build_transform.register(Fidelity)
def _(dim, type_requirement, dist_requirement):
return []
@build_transform.register(Integer)
def _(dim, type_requirement, dist_requirement):
transformers = []
if dist_requirement == "linear" and dim.prior_name[4:] in NON_LINEAR:
transformers.extend([Reverse(Quantize()), Linearize()])
# NOTE: we do not turn back to integer even though linearize outputs real
# otherwise the mapping from exp(int) to int squashes out lots of possible values.
elif type_requirement == "real":
transformers.append(Reverse(Quantize()))
return transformers
@build_transform.register(Real)
def _(dim, type_requirement, dist_requirement):
transformers = []
if dim.precision is not None:
transformers.append(Precision(dim.precision))
if dist_requirement == "linear" and dim.prior_name in NON_LINEAR:
transformers.append(Linearize())
elif type_requirement == "integer":
# NOTE: This may cause out-of-bound errors for rounded reals. Not fixed for now
# because there are no foreseeable algorithms that may require integer type.
transformers.append(Quantize())
return transformers
[docs]def reshape(space, shape_requirement):
"""Build a reshaped space"""
if shape_requirement is None:
return space
# We assume shape_requirement == 'flattened'
reshaped_space = ReshapedSpace(space)
for dim_index, dim in enumerate(space.values()):
if not dim.shape:
reshaped_space.register(
ReshapedDimension(
transformer=Identity(dim.type),
original_dimension=dim,
index=dim_index,
)
)
else:
for index in itertools.product(*map(range, dim.shape)):
key = f'{dim.name}[{",".join(map(str, index))}]'
reshaped_space.register(
ReshapedDimension(
transformer=View(dim.shape, index, dim.type),
original_dimension=dim,
name=key,
index=dim_index,
)
)
return reshaped_space
[docs]def build_required_space(
original_space, type_requirement=None, shape_requirement=None, dist_requirement=None
):
"""Build a :class:`orion.algo.space.Space` object which agrees to the `requirements` imposed
by the desired optimization algorithm.
It uses appropriate cascade of `Transformer` objects per `orion.algo.space.Dimension`
contained in `original_space`. `ReshapedTransformer` objects are used above
the `Transformer` if the optimizatios algorithm requires flattened dimensions.
Parameters
----------
original_space : `orion.algo.space.Space`
Original problem's definition of parameter space given by the user to Oríon.
type_requirement: str, None
String defining the requirement of the algorithm. It can be one of the following
- 'real', the dim should be transformed so type is `orion.algo.space.Real`
- 'integer', the dim should be transformed so type is `orion.algo.space.Integer`
- 'numerical', the dim should be transformed so type is either `orion.algo.space.Integer` or
`orion.algo.space.Real`
- None, no requirement
shape_requirement: str, None
String defining the shape requirement of the algorithm.
- 'flattened', any dimension with shape > 1 will be flattened
- None, no requirement
dist_requirement: str, None
String defining the distribution requirement of the algorithm.
- 'linear', any dimension with logarithmic prior while be linearized
- None, no requirement
"""
space = transform(original_space, type_requirement, dist_requirement)
space = reshape(space, shape_requirement)
return space
[docs]class Identity(Transformer):
"""Implement an identity transformation. Everything as it is."""
def __init__(self, domain_type=None):
self._domain_type = domain_type
@property
def first(self):
"""Signals to ReshapedSpace whether this dimension should be used for `reverse`"""
return True
# pylint:disable=unused-argument
[docs] def reverse(self, transformed_point, index=None):
"""Return `transformed_point` as it is."""
if index is not None:
return transformed_point[index]
return transformed_point
@property
def domain_type(self):
"""Return declared domain type on initialization."""
return self._domain_type
@property
def target_type(self):
"""Return domain type as this will be the target in a identity transformation."""
return self.domain_type
[docs]class Compose(Transformer):
"""Initialize composite transformer with a list of `Transformer` objects
and domain type on which it will be applied.
"""
def __init__(self, transformers, base_domain_type=None):
try:
self.apply = transformers[-1]
except IndexError:
self.apply = Identity()
if len(transformers) > 1:
self.composition = Compose(transformers[:-1], base_domain_type)
else:
self.composition = Identity(base_domain_type)
assert (
self.apply.domain_type is None
or self.composition.target_type == self.apply.domain_type
)
# pylint:disable=unused-argument
[docs] def reverse(self, transformed_point, index=None):
"""Reverse transformation by reversing in the opposite order of the `transformers` list."""
transformed_point = self.apply.reverse(transformed_point)
return self.composition.reverse(transformed_point)
[docs] def interval(self, alpha=1.0):
"""Return interval of composed transformation."""
if hasattr(self.apply, "interval"):
return self.apply.interval(alpha)
return None
[docs] def infer_target_shape(self, shape):
"""Return the shape of the dimension after transformation."""
shape = self.composition.infer_target_shape(shape)
return self.apply.infer_target_shape(shape)
@property
def domain_type(self):
"""Return base domain type."""
return self.composition.domain_type
@property
def target_type(self):
"""Infer type of the transformation target."""
type_before = self.composition.target_type
type_after = self.apply.target_type
return type_after if type_after else type_before
# pylint:disable=protected-access
def _get_hashable_members(self):
return (
(self.__class__.__name__,)
+ self.apply._get_hashable_members()
+ self.composition._get_hashable_members()
)
[docs]class Reverse(Transformer):
"""Apply the reverse transformation that another one would do."""
def __init__(self, transformer: Transformer):
assert not isinstance(
transformer, OneHotEncode
), "real to categorical is pointless"
self.transformer = transformer
# pylint:disable=unused-argument
[docs] def reverse(self, transformed_point, index=None):
"""Use `transform` of composed `transformer`."""
return self.transformer.transform(transformed_point)
@property
def target_type(self):
"""Return `domain_type` of composed `transformer`."""
return self.transformer.domain_type
@property
def domain_type(self):
"""Return `target_type` of composed `transformer`."""
return self.transformer.target_type
[docs]class Precision(Transformer):
"""Round real numbers to requested precision."""
domain_type = "real"
target_type = "real"
def __init__(self, precision=4):
self.precision = precision
# pylint:disable=unused-argument
[docs] def reverse(self, transformed_point, index=None):
"""Cast `transformed_point` to floats, as numpy arrays."""
return self.transform(transformed_point)
[docs]class Quantize(Transformer):
"""Transform real numbers to integers, violating injection."""
domain_type = "real"
target_type = "integer"
# pylint:disable=unused-argument
[docs] def reverse(self, transformed_point, index=None):
"""Cast `transformed_point` to floats, as numpy arrays."""
return numpy.asarray(transformed_point).astype(float)
[docs]class Enumerate(Transformer):
"""Enumerate categories.
Effectively transform from a list of objects to a range of integers.
"""
domain_type = "categorical"
target_type = "integer"
def __init__(self, categories):
self.categories = categories
map_dict = {cat: i for i, cat in enumerate(categories)}
self._map = numpy.vectorize(lambda x: map_dict[x], otypes="i")
self._imap = numpy.vectorize(lambda x: categories[x], otypes=[object])
def __deepcopy__(self, memo):
"""Make a deepcopy"""
return type(self)(self.categories)
# pylint:disable=unused-argument
[docs] def reverse(self, transformed_point, index=None):
"""Return categories corresponding to their positions inside `transformed_point`."""
return self._imap(transformed_point)
# pylint:disable=unused-argument
[docs] def interval(self, alpha=1.0):
"""Return the interval for the enumerated choices."""
return (0, len(self.categories) - 1)
[docs]class OneHotEncode(Transformer):
"""Encode categories to a 1-hot integer space representation."""
domain_type = "integer"
target_type = "real"
def __init__(self, bound: int):
self.num_cats = bound
# pylint:disable=unused-argument
[docs] def reverse(self, transformed_point, index=None):
"""Match real vector representations to integers using an argmax function.
If the number of dimensions is exactly 2, then use 0.5 as a decision boundary,
and convert representation to integers 0 or 1.
If the number of dimensions is exactly 1, then return zeros.
.. note:: This reverse transformation possibly removes the last tensor dimension
from `transformed_point`.
"""
point_ = numpy.asarray(transformed_point)
if self.num_cats == 2:
return (point_ > 0.5).astype(int)
elif self.num_cats == 1:
return numpy.zeros_like(point_, dtype=int)
assert point_.shape[-1] == self.num_cats
return point_.argmax(axis=-1)
# pylint:disable=unused-argument
[docs] def interval(self, alpha=1.0):
"""Return the interval for the one-hot encoding in proper shape."""
if self.num_cats == 2:
return 0, 1
else:
low = numpy.zeros(self.num_cats)
high = numpy.ones(self.num_cats)
return low, high
[docs] def infer_target_shape(self, shape):
"""Infer that transformed points will have one more tensor dimension,
if the number of supported integers to transform is larger than 2.
"""
return tuple(list(shape) + [self.num_cats]) if self.num_cats > 2 else shape
def _get_hashable_members(self):
return super()._get_hashable_members() + (self.num_cats,)
[docs]class Linearize(Transformer):
"""Transform real numbers from loguniform to linear."""
domain_type = "real"
target_type = "real"
# pylint:disable=unused-argument
[docs] def reverse(self, transformed_point, index=None):
"""Turn linear distribution to logarithmic distribution."""
return numpy.exp(numpy.asarray(transformed_point))
[docs]class View(Transformer):
"""Look-up single index in a dimensions with shape > 1"""
def __init__(self, shape, index, domain_type=None):
self.shape = shape
self.index = index
self._domain_type = domain_type
@property
def first(self):
"""Signals to ReshapedSpace whether this dimension should be used for `reverse`"""
return sum(self.index) == 0
[docs] def reverse(self, transformed_point, index=None):
"""Only return packend point if view of first element, otherwise drop."""
subset = transformed_point[index : index + numpy.prod(self.shape)]
return numpy.array(subset).reshape(self.shape)
[docs] def interval(self, interval):
"""Return corresponding view from interval"""
return (interval[0][self.index], interval[1][self.index])
@property
def domain_type(self):
"""Return declared domain type on initialization."""
return self._domain_type
@property
def target_type(self):
"""Return domain type as this will be the target in flatten transformation."""
return self.domain_type
[docs]class ReshapedDimension(TransformedDimension):
"""Duck-type :class:`orion.algo.space.Dimension` to mimic its functionality."""
def __init__(self, transformer, original_dimension, index, name=None):
super().__init__(transformer, original_dimension)
if name is None:
name = original_dimension.name
self._name = name
self.index = index
@property
def first(self):
"""Signals to ReshapedSpace whether this dimension should be used for `reverse`"""
return self.transformer.first
[docs] def reverse(self, transformed_point, index=None):
"""Expose `Transformer.reverse` interface from underlying instance."""
return self.transformer.reverse(transformed_point, index)
[docs] def interval(self, alpha=1.0):
"""Map the interval bounds to the transformed ones."""
interval = self.original_dimension.interval(alpha)
if hasattr(interval[0], "shape") and numpy.prod(interval[0].shape) > 1:
return self.transformer.interval(interval)
return interval
@property
def cardinality(self):
"""Compute cardinality"""
cardinality = super().cardinality
if isinstance(self.transformer, View):
cardinality /= numpy.prod(self.transformer.shape)
return cardinality
[docs] def cast(self, point):
"""Cast a point according to original_dimension and then transform it"""
return self.original_dimension.cast(point)
@property
def shape(self):
"""Shape is fixed to ()."""
return ()
@property
def name(self):
"""Name of the view"""
return self._name
[docs]class ReshapedSpace(Space):
"""Wrap the `TransformedSpace` to support reshape methods.
Parameters
----------
space: `orion.core.worker.TransformedSpace`
Transformed version of the orinigal problem's definition of parameter space.
"""
contains = ReshapedDimension
def __init__(self, original_space, *args, **kwargs):
super().__init__(*args, **kwargs)
self._original_space = original_space
@property
def original(self):
"""Original space without reshape or transformations"""
return self._original_space
[docs] def reverse(self, transformed_trial: Trial) -> Trial:
"""Reverses transformation so that a point from this `ReshapedSpace` to be in the original
one.
"""
return self.original.reverse(self.restore_shape(transformed_trial))
[docs] def reshape(self, trial):
"""Reshape the point"""
point = format_trials.trial_to_tuple(trial, self._original_space)
reshaped_point = []
for dim in self.values():
reshaped_point.append(dim.transform(point[dim.index]))
return change_trial_params(trial, reshaped_point, self)
[docs] def restore_shape(self, transformed_trial):
"""Restore shape."""
transformed_point = format_trials.trial_to_tuple(transformed_trial, self)
original_keys = self._original_space.keys()
point = [None for _ in original_keys]
for index, dim in enumerate(self.values()):
if dim.first:
point_index = original_keys.index(dim.original_dimension.name)
point[point_index] = dim.reverse(transformed_point, index)
return change_trial_params(transformed_trial, point, self._original_space)
[docs] def sample(self, n_samples=1, seed=None):
"""Sample from the original dimension and forward transform them."""
trials = self.original.sample(n_samples=n_samples, seed=seed)
return [self.reshape(trial) for trial in trials]
[docs] def assert_contains(self, trial):
"""Check if the trial or key is contained inside the space, if not an exception is raised
Raises
------
TypeError when a dimension is not compatible with the space
"""
if isinstance(trial, str):
super().assert_contains(trial)
return self.original.assert_contains(self.restore_shape(trial))
def __contains__(self, key_or_trial):
"""Check whether `trial` is within the bounds of the space.
Or check if a name for a dimension is registered in this space.
Parameters
----------
key_or_trial: str or `orion.core.worker.trial.Trial`
If str, test if the string is a dimension part of the search space.
If a Trial, test if trial's hyperparameters fit the current search space.
"""
try:
self.assert_contains(key_or_trial)
return True
except ValueError:
return False
@property
def cardinality(self):
"""Reshape does not affect cardinality"""
return self.original.cardinality
[docs]def change_trial_params(trial, point, space):
"""Convert params in Param objects and update trial"""
new_trial = copy.copy(trial)
# pylint: disable=protected-access
new_trial._params = format_trials.tuple_to_trial(point, space)._params
return new_trial