Add MOO mode for 1+1 + fix DE uniform sampling (#1058)

CHANGELOG.md

@@ -21,6 +21,7 @@
 - `Parameter` classes have now a layer structure [#1045](https://github.com/facebookresearch/nevergrad/pull/1045)
   which simplifies changing their behavior. In future PRs this system will take charge of bounds, other constraints,
   sampling etc.
+- `DE` initial sampling as been updated to take bounds into accounts [#1058](https://github.com/facebookresearch/nevergrad/pull/1058)
 
 ### Other changes
 

nevergrad/benchmark/test_xpbase.py

@@ -48,9 +48,7 @@ def test_run_packed_artificial_function() -> None:
     )
     xp = xpbase.Experiment(func, optimizer="OnePlusOne", budget=24, num_workers=2, batch_mode=True, seed=14)
     summary = xp.run()
-    np.testing.assert_almost_equal(
-        summary["loss"], -9784.829729792353, decimal=1
-    )  # makes sure seeding works!
+    np.testing.assert_almost_equal(summary["loss"], -9784.8, decimal=1)  # makes sure seeding works!
 
 
 def test_noisy_artificial_function_loss() -> None:

nevergrad/functions/images/core.py

@@ -3,10 +3,11 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
-import cv2
-from pathlib import Path
+import os
 import itertools
+from pathlib import Path
 
+import cv2
 import numpy as np
 import PIL.Image
 import torch.nn as nn
@@ -17,6 +18,7 @@
 
 import nevergrad as ng
 import nevergrad.common.typing as tp
+from nevergrad.common import errors
 from .. import base
 from . import imagelosses
 
@@ -259,6 +261,8 @@ def __init__(
         if not torch.cuda.is_available():
             use_gpu = False
         # Storing high level information..
+        if os.environ.get("CIRCLECI", False):
+            raise errors.UnsupportedExperiment("ImageFromPGAN is not well supported in CircleCI")
         self.pgan_model = torch.hub.load(
             "facebookresearch/pytorch_GAN_zoo:hub",
             "PGAN",

nevergrad/optimization/base.py

@@ -213,11 +213,14 @@ def pareto_front(
         ----
         During non-multiobjective optimization, this returns the current pessimistic best
         """
-        if self._hypervolume_pareto is None:
-            return [self.provide_recommendation()]
-        return self._hypervolume_pareto.pareto_front(
-            size=size, subset=subset, subset_tentatives=subset_tentatives
+        pareto = (
+            []
+            if self._hypervolume_pareto is None
+            else self._hypervolume_pareto.pareto_front(
+                size=size, subset=subset, subset_tentatives=subset_tentatives
+            )
         )
+        return pareto if pareto else [self.provide_recommendation()]
 
     def dump(self, filepath: tp.Union[str, Path]) -> None:
         """Pickles the optimizer into a file."""

nevergrad/optimization/differentialevolution.py

@@ -5,11 +5,10 @@
 
 import warnings
 import numpy as np
-from scipy import stats
 import nevergrad.common.typing as tp
 from nevergrad.parametrization import parameter as p
 from . import base
-from . import sequences
+from . import oneshot
 
 
 class Crossover:
@@ -100,7 +99,7 @@ def __init__(
         self._penalize_cheap_violations = True
         self._uid_queue = base.utils.UidQueue()
         self.population: tp.Dict[str, p.Parameter] = {}
-        self.sampler: tp.Optional[sequences.Sampler] = None
+        self.sampler: tp.Optional[base.Optimizer] = None
 
     def recommend(self) -> p.Parameter:  # This is NOT the naive version. We deal with noise.
         if self._config.recommendation != "noisy":
@@ -117,18 +116,21 @@ def recommend(self) -> p.Parameter:  # This is NOT the naive version. We deal wi
     def _internal_ask_candidate(self) -> p.Parameter:
         if len(self.population) < self.llambda:  # initialization phase
             init = self._config.initialization
-            if self.sampler is None and init != "gaussian":
+            if self.sampler is None and init not in ["gaussian", "parametrization"]:
                 assert init in ["LHS", "QR"]
-                sampler_cls = sequences.LHSSampler if init == "LHS" else sequences.HammersleySampler
-                self.sampler = sampler_cls(
-                    self.dimension, budget=self.llambda, scrambling=init == "QR", random_state=self._rng
+                self.sampler = oneshot.SamplingSearch(
+                    sampler=init if init == "LHS" else "Hammersley", scrambled=init == "QR", scale=self.scale
+                )(
+                    self.parametrization,
+                    budget=self.llambda,
                 )
-            new_guy = self.scale * (
-                self._rng.normal(0, 1, self.dimension)
-                if self.sampler is None
-                else stats.norm.ppf(self.sampler())
-            )
-            candidate = self.parametrization.spawn_child().set_standardized_data(new_guy)
+            if init == "parametrization":
+                candidate = self.parametrization.sample()
+            elif self.sampler is not None:
+                candidate = self.sampler.ask()
+            else:
+                new_guy = self.scale * self._rng.normal(0, 1, self.dimension)
+                candidate = self.parametrization.spawn_child().set_standardized_data(new_guy)
             candidate.heritage["lineage"] = candidate.uid  # new lineage
             self.population[candidate.uid] = candidate
             self._uid_queue.asked.add(candidate.uid)
@@ -146,10 +148,11 @@ def _internal_ask_candidate(self) -> p.Parameter:
         # redefine the different parents in case of multiobjective optimization
         if self._config.multiobjective_adaptation and self.num_objectives > 1:
             pareto = self.pareto_front()
+            # can't use choice directly on pareto, because parametrization can be iterable
             if pareto:
-                best = parent if parent in pareto else self._rng.choice(pareto)
+                best = parent if parent in pareto else pareto[self._rng.choice(len(pareto))]
             if len(pareto) > 2:  # otherwise, not enough diversity
-                a, b = self._rng.choice(pareto, size=2, replace=False)
+                a, b = (pareto[idx] for idx in self._rng.choice(len(pareto), size=2, replace=False))
         # define donor
         data_a, data_b, data_best = (
             indiv.get_standardized_data(reference=self.parametrization) for indiv in (a, b, best)
@@ -228,8 +231,9 @@ class DifferentialEvolution(base.ConfiguredOptimizer):
 
     Parameters
     ----------
-    initialization: "LHS", "QR" or "gaussian"
-        algorithm/distribution used for the initialization phase
+    initialization: "parametrization", "LHS" or "QR"
+        algorithm/distribution used for the initialization phase. If "parametrization", this uses the
+        sample method of the parametrization.
     scale: float or str
         scale of random component of the updates
     recommendation: "pessimistic", "optimistic", "mean" or "noisy"
@@ -256,7 +260,7 @@ class DifferentialEvolution(base.ConfiguredOptimizer):
     def __init__(
         self,
         *,
-        initialization: str = "gaussian",
+        initialization: str = "parametrization",
         scale: tp.Union[str, float] = 1.0,
         recommendation: str = "optimistic",
         crossover: tp.Union[str, float] = 0.5,
@@ -268,7 +272,7 @@ def __init__(
     ) -> None:
         super().__init__(_DE, locals(), as_config=True)
         assert recommendation in ["optimistic", "pessimistic", "noisy", "mean"]
-        assert initialization in ["gaussian", "LHS", "QR"]
+        assert initialization in ["gaussian", "LHS", "QR", "parametrization"]
         assert isinstance(scale, float) or scale == "mini"
         if not isinstance(popsize, int):
             assert popsize in ["large", "dimension", "standard"]

nevergrad/optimization/experimentalvariants.py

@@ -22,7 +22,7 @@
 ParametrizationDE = DifferentialEvolution(crossover="parametrization").set_name(
     "ParametrizationDE", register=True
 )
-MiniDE = DifferentialEvolution(scale="mini").set_name("MiniDE", register=True)
+MiniDE = DifferentialEvolution(initialization="gaussian", scale="mini").set_name("MiniDE", register=True)
 MiniLhsDE = DifferentialEvolution(initialization="LHS", scale="mini").set_name("MiniLhsDE", register=True)
 MiniQrDE = DifferentialEvolution(initialization="QR", scale="mini").set_name("MiniQrDE", register=True)
 AlmostRotationInvariantDEAndBigPop = DifferentialEvolution(crossover=0.9, popsize="dimension").set_name(

nevergrad/optimization/multiobjective/core.py

@@ -166,8 +166,8 @@ def pareto_front(
         if size is None or size >= len(self._pareto):  # No limit: we return the full set.
             return self._pareto
         if subset == "random":
-            return self._rng.choice(self._pareto, size)  # type: ignore
-        tentatives = [self._rng.choice(self._pareto, size) for _ in range(subset_tentatives)]
+            return self._rng.choice(self._pareto, size).tolist()  # type: ignore
+        tentatives = [self._rng.choice(self._pareto, size).tolist() for _ in range(subset_tentatives)]
         if self._hypervolume is None:
             raise RuntimeError("Hypervolume not initialized, not supported")  # TODO fix
         hypervolume = self._hypervolume
@@ -192,4 +192,4 @@ def pareto_front(
                             raise ValueError(f'Unknown subset for Pareto-Set subsampling: "{subset}"')
                     score += best_score ** 2 if subset != "EPS" else max(score, best_score)
                 scores += [score]
-        return tentatives[scores.index(min(scores))]
+        return tentatives[scores.index(min(scores))]  # type: ignore

nevergrad/optimization/oneshot.py

@@ -75,14 +75,12 @@ def avg_of_k_best(archive: utils.Archive[utils.MultiValue], method: str = "dimfo
         raise ValueError(f"{method} not implemented as a method for choosing k in avg_of_k_best.")
     k = 1 if k < 1 else int(k)
     # Wasted time.
-    first_k_individuals = [
-        k for k in sorted(items, key=lambda indiv: archive[indiv[0]].get_estimation("pessimistic"))[:k]
-    ]
+    first_k_individuals = sorted(items, key=lambda indiv: archive[indiv[0]].get_estimation("pessimistic"))[:k]
     assert len(first_k_individuals) == k
     return np.array(sum(p[0] for p in first_k_individuals) / k)
 
 
-# # # # # classes of optimizers # # # # #
+# # # # # classes of optimizers # # # # #
 
 
 class OneShotOptimizer(base.Optimizer):
@@ -99,7 +97,7 @@ class OneShotOptimizer(base.Optimizer):
 # - Some variants use a rescaling depending on the budget and the dimension.
 
 
-# # # # # One-shot optimizers: all fitness evaluations are in parallel. # # # # #
+# # # # # One-shot optimizers: all fitness evaluations are in parallel. # # # # #
 
 
 # pylint: disable=too-many-arguments,too-many-instance-attributes

nevergrad/optimization/optimizerlib.py

@@ -69,9 +69,11 @@ def __init__(
         noise_handling: tp.Optional[tp.Union[str, tp.Tuple[str, float]]] = None,
         mutation: str = "gaussian",
         crossover: bool = False,
+        use_pareto: bool = False,
     ) -> None:
         super().__init__(parametrization, budget=budget, num_workers=num_workers)
         self._sigma: float = 1
+        self.use_pareto = use_pareto
         all_params = paramhelpers.flatten_parameter(self.parametrization)
         arity = max(
             len(param.choices) if isinstance(param, p.TransitionChoice) else 500
@@ -153,6 +155,12 @@ def _internal_ask_candidate(self) -> p.Parameter:
         # crossover
         mutator = mutations.Mutator(self._rng)
         pessimistic = self.current_bests["pessimistic"].parameter.spawn_child()
+        if self.num_objectives > 1 and self.use_pareto:  # multiobjective
+            # revert to using a sample of the pareto front (not "pessimistic" though)
+            pareto = (
+                self.pareto_front()
+            )  # we can't use choice directly, because numpy does not like iterables
+            pessimistic = pareto[self._rng.choice(len(pareto))].spawn_child()
         ref = self.parametrization
         if self.crossover and self._num_ask % 2 == 1 and len(self.archive) > 2:
             data = mutator.crossover(
@@ -293,6 +301,8 @@ class ParametrizedOnePlusOne(base.ConfiguredOptimizer):
         - `"lengler"`: specific mutation rate chosen as a function of the dimension and iteration index.
     crossover: bool
         whether to add a genetic crossover step every other iteration.
+    use_pareto: bool
+        whether to restart from a random pareto element in multiobjective mode, instead of the last one added
 
     Notes
     -----
@@ -310,6 +320,7 @@ def __init__(
         noise_handling: tp.Optional[tp.Union[str, tp.Tuple[str, float]]] = None,
         mutation: str = "gaussian",
         crossover: bool = False,
+        use_pareto: bool = False,
     ) -> None:
         super().__init__(_OnePlusOne, locals())
 

nevergrad/optimization/test_callbacks.py

@@ -33,9 +33,9 @@ def test_log_parameters(tmp_path: Path) -> None:
     logs = logger.load_flattened()
     assert len(logs) == 32
     assert isinstance(logs[-1]["1"], float)
-    assert len(logs[-1]) == 35
+    assert len(logs[-1]) == 36
     logs = logger.load_flattened(max_list_elements=2)
-    assert len(logs[-1]) == 27
+    assert len(logs[-1]) == 28
     # deletion
     logger = callbacks.ParametersLogger(filepath, append=False)
     assert not logger.load()

nevergrad/optimization/test_optimizerlib.py

@@ -715,7 +715,7 @@ def _multiobjective(z: np.ndarray) -> tp.Tuple[float, float, float]:
     return (abs(x - 1), abs(y + 1), abs(x - y))
 
 
-@pytest.mark.parametrize("name", ["DE", "ES"])  # type: ignore
+@pytest.mark.parametrize("name", ["DE", "ES", "OnePlusOne"])  # type: ignore
 @testing.suppress_nevergrad_warnings()  # hides bad loss
 def test_mo_constrained(name: str) -> None:
     optimizer = optlib.registry[name](2, budget=60)
@@ -733,6 +733,26 @@ def constraint(arg: tp.Any) -> bool:  # pylint: disable=unused-argument
         assert optimizer._rank_method is not None  # make sure the nsga2 ranker is used
 
 
+@pytest.mark.parametrize("name", ["DE", "ES", "OnePlusOne"])  # type: ignore
+@testing.suppress_nevergrad_warnings()  # hides bad loss
+def test_mo_with_nan(name: str) -> None:
+    param = ng.p.Instrumentation(x=ng.p.Scalar(lower=0, upper=5), y=ng.p.Scalar(lower=0, upper=3))
+    optimizer = optlib.registry[name](param, budget=60)
+    optimizer.tell(ng.p.MultiobjectiveReference(), [10, 10, 10])
+    for _ in range(50):
+        cand = optimizer.ask()
+        optimizer.tell(cand, [-38, 0, np.nan])
+
+
+def test_de_sampling() -> None:
+    param = ng.p.Scalar(lower=-100, upper=100).set_mutation(sigma=1)
+    opt = optlib.LhsDE(param, budget=600, num_workers=100)
+    above_50 = 0
+    for _ in range(100):
+        above_50 += abs(opt.ask().value) > 50
+    assert above_50 > 20  # should be around 50
+
+
 def test_paraportfolio_de() -> None:
     workers = 40
     opt = optlib.ParaPortfolio(12, budget=100 * workers, num_workers=workers)