Merge pull request #8 from jrapin/traverser_and_copy

Traverser and copy

nevergrad/parametrization/choice.py

@@ -228,15 +228,6 @@ def mutate(self) -> None:
         for ind in indices:
             self.choices[ind].mutate()
 
-    def _internal_spawn_child(self: C) -> C:
-        child = self.__class__(
-            choices=self.choices.spawn_child(),
-            deterministic=self._deterministic,
-            repetitions=self._repetitions,
-        )
-        child._content["weights"] = self.weights.spawn_child()
-        return child
-
 
 class TransitionChoice(BaseChoice):
     """Ordered categorical parameter, choosing one of the provided choice options as a value, with continuous transitions.
@@ -325,9 +316,3 @@ def mutate(self) -> None:
         indices = set(self.indices)
         for ind in indices:
             self.choices[ind].mutate()
-
-    def _internal_spawn_child(self: T) -> T:
-        child = self.__class__(choices=self.choices.spawn_child(), repetitions=self._repetitions)
-        child._content["positions"] = self.positions.spawn_child()
-        child._content["transitions"] = self.transitions.spawn_child()
-        return child

nevergrad/parametrization/core.py

@@ -4,6 +4,7 @@
 # LICENSE file in the root directory of this source tree.
 
 import uuid
+import copy
 import warnings
 import operator
 import functools
@@ -358,14 +359,25 @@ def spawn_child(self: P, new_value: tp.Optional[tp.Any] = None) -> P:
             Optionally, a new value will be set after creation
         """
         rng = self.random_state  # make sure to create one before spawning
-        child = self._internal_spawn_child()
-        child._set_random_state(rng)
-        child._constraint_checkers = list(self._constraint_checkers)
-        child._generation = self.generation + 1
-        child._descriptors = self._descriptors
-        child._name = self._name
-        child.parents_uids.append(self.uid)
+        child = copy.copy(self)
+        child.uid = uuid.uuid4().hex
+        child._frozen = False
+        child._generation += 1
+        child.parents_uids = [self.uid]
         child.heritage = dict(self.heritage)
+        child._treecall = self._treecall.new(child)
+        child._meta = {}
+        child.loss = None
+        child._losses = None
+        child._constraint_checkers = list(self._constraint_checkers)
+        attribute = self._treecall.attribute
+        container = getattr(child, self._treecall.attribute)
+        if attribute != "__dict__":  # make a copy of the container if different from __dict__
+            container = dict(container) if isinstance(container, dict) else list(container)
+            setattr(child, attribute, container)
+        for key, val in self._treecall._subitems():
+            container[key] = val.spawn_child()
+        child._set_random_state(rng)
         if new_value is not None:
             child.value = new_value
         return child
@@ -397,7 +409,6 @@ def copy(self: P) -> P:  # TODO test (see former instrumentation_copy test)
         This is used to run multiple experiments
         """
         child = self.spawn_child()
-        child._name = self._name
         child.random_state = None
         return child
 
@@ -581,11 +592,6 @@ def sample(self: D) -> D:
         child.heritage["lineage"] = child.uid
         return child
 
-    def _internal_spawn_child(self: D) -> D:
-        child = self.__class__()
-        child._content = {k: v.spawn_child() for k, v in self._content.items()}
-        return child
-
 
 class Dict(Container):
     """Dictionary-valued parameter. This Parameter can contain other Parameters,

nevergrad/parametrization/data.py

@@ -341,16 +341,6 @@ def _internal_set_standardized_data(
         if reference.bound_transform is not None:
             self._value = reference.bound_transform.forward(self._value)
 
-    def _internal_spawn_child(self: D) -> D:
-        child = self.__class__(init=self.value)
-        child.parameters._content = {
-            k: v.spawn_child() if isinstance(v, core.Parameter) else v
-            for k, v in self.parameters._content.items()
-        }
-        for name in ["integer", "exponent", "bounds", "bound_transform", "full_range_sampling"]:
-            setattr(child, name, getattr(self, name))
-        return child
-
     def _internal_get_standardized_data(self: D, reference: D) -> np.ndarray:
         return reference._to_reduced_space(self._value) - reference._get_ref_data()  # type: ignore
 
@@ -388,6 +378,13 @@ def recombine(self: D, *others: D) -> None:
         else:
             raise ValueError(f'Unknown recombination "{recomb}"')
 
+    def spawn_child(self: D, new_value: tp.Optional[tp.Any] = None) -> D:
+        child = super().spawn_child()  # dont forward the value
+        child._value = np.array(self._value, copy=True)
+        if new_value is not None:
+            child.value = new_value
+        return child
+
 
 class Array(Data):
 
@@ -428,7 +425,7 @@ class Scalar(Data):
     upper: optional float
         maximum value if any
     mutable_sigma: bool
-        whether the mutation standard deviation must mutate as well (for mutation based algorithms)
+        wheter the mutation standard deviation must mutate as well (for mutation based algorithms)
 
     Notes
     -----

nevergrad/parametrization/helpers.py

@@ -111,6 +111,7 @@ def split_as_data_parameters(
     # analyze results
     data = copied.get_standardized_data(reference=ref)
     order: tp.List[int] = []
+    print(data)
     for val, _ in itertools.groupby(data):
         num = int(np.round(val))
         if num in order:

nevergrad/parametrization/utils.py

@@ -232,6 +232,9 @@ def __init__(self, obj: tp.Any, cls: tp.Type[tp.Any], attribute: str = "__dict__
         self.cls = cls
         self.attribute = attribute
 
+    def new(self, obj: tp.Any) -> "Treecall":
+        return Treecall(obj, cls=self.cls, attribute=self.attribute)
+
     def _subitems(self) -> tp.Iterator[tp.Tuple[tp.Any, tp.Any]]:
         container = getattr(self.obj, self.attribute)
         if not isinstance(container, (list, dict)):