Add a multiobjective adaptation for DE (#789)

facebookresearch · Aug 11, 2020 · e22212d · e22212d
1 parent 4f81eb8
commit e22212d
Show file tree

Hide file tree

Showing 16 changed files with 188 additions and 71 deletions.
diff --git a/.circleci/config.yml b/.circleci/config.yml
@@ -28,9 +28,9 @@ commands:
       - restore_cache:
           name: "[all] Restore cache"
           keys:
-          - v4-dependencies-{{ checksum "requirements/dev.txt" }}-{{ checksum "requirements/main.txt"}}-{{ checksum "requirements/bench.txt"}}
+          - v5-dependencies-{{ checksum "requirements/dev.txt" }}-{{ checksum "requirements/main.txt"}}-{{ checksum "requirements/bench.txt"}}
           # fallback to installing main requirements
-          - v4-dependencies-main-{{ checksum "requirements/main.txt"}}
+          - v5-dependencies-main-{{ checksum "requirements/main.txt"}}
 
       - run:
           name: "[all] Install dependencies"
@@ -53,7 +53,7 @@ jobs:
       - restore_cache:
           name: "[no-extra] Restore cache"
           keys:
-          - v4-dependencies-main-{{ checksum "requirements/main.txt"}}
+          - v5-dependencies-main-{{ checksum "requirements/main.txt"}}
           # fallback to using the latest cache if no exact match is found
 
       - run:
@@ -75,7 +75,7 @@ jobs:
           name: "[no-extra] Save cache"
           paths:
             - ./venv
-          key: v4-dependencies-main-{{ checksum "requirements/main.txt"}}
+          key: v5-dependencies-main-{{ checksum "requirements/main.txt"}}
 
       - run:
           name: "[no-extra] Run basic tests (checking dependencies)"
@@ -90,7 +90,7 @@ jobs:
           name: "[all] Save cache"
           paths:
             - ./venv
-          key: v4-dependencies-{{ checksum "requirements/dev.txt" }}-{{ checksum "requirements/main.txt"}}-{{ checksum "requirements/bench.txt"}}
+          key: v5-dependencies-{{ checksum "requirements/dev.txt" }}-{{ checksum "requirements/main.txt"}}-{{ checksum "requirements/bench.txt"}}
 
       - run:
           name: "[all] Print installation"
@@ -206,7 +206,7 @@ jobs:
       - restore_cache:
           name: "[all] Restore cache"
           keys:
-          - v4-win-dependencies-{{ checksum "requirements/dev.txt" }}-{{ checksum "requirements/main.txt"}}-{{ checksum "requirements/bench.txt"}}
+          - v5-win-dependencies-{{ checksum "requirements/dev.txt" }}-{{ checksum "requirements/main.txt"}}-{{ checksum "requirements/bench.txt"}}
 
       - run:
           name: Setup nevergrad in virtualenv
@@ -221,7 +221,7 @@ jobs:
           name: "[all] Save cache"
           paths:
             - ./venv
-          key: v4-win-dependencies-{{ checksum "requirements/dev.txt" }}-{{ checksum "requirements/main.txt"}}-{{ checksum "requirements/bench.txt"}}
+          key: v5-win-dependencies-{{ checksum "requirements/dev.txt" }}-{{ checksum "requirements/main.txt"}}-{{ checksum "requirements/bench.txt"}}
 
       - run:
           name: pytest

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -5,6 +5,7 @@
 **Cautious:** current `master` branch and `0.4.2.postX` version introduce tentative APIs which may be removed in the near future. Use version `0.4.2` for a more stable version.
 
 - as an **experimental** feature, `tell` method can now receive a list/array of losses for multi-objective optimization [#775](https://github.com/facebookresearch/nevergrad/pull/775). For now it is neither robust, nor scalable, nor stable, nor optimal so be careful when using it. More information in the [documentation](https://facebookresearch.github.io/nevergrad/optimization.html#multiobjective-minimization-with-nevergrad).
+- `DE` and its variants have been updated to make use of the multi-objective losses [#789](https://github.com/facebookresearch/nevergrad/pull/789). This is a **preliminary** fix since the initial `DE` implementaton was ill-suited for this use case.
 - `tell` argument `value` is renamed to `loss` for clarification [#774](https://github.com/facebookresearch/nevergrad/pull/774). This can be breaking when using named arguments!
 
 ## 0.4.2 (2020-08-04)

diff --git a/docs/optimization.rst b/docs/optimization.rst
@@ -245,6 +245,8 @@ We are currently working on an **new experimental API** allowing users to direct
     :start-after: DOC_MULTIOBJ_OPT_0
     :end-before: DOC_MULTIOBJ_OPT_1
 
+Note that `DE` and its variants have been updated to make use of the multi-objective losses [#789](https://github.com/facebookresearch/nevergrad/pull/789). This is a **preliminary** fix since the initial `DE` implementaton was ill-suited for this use case.
+
 
 Reproducibility
 ---------------

diff --git a/nevergrad/benchmark/experiments.py b/nevergrad/benchmark/experiments.py
@@ -1131,25 +1131,25 @@ def multiobjective_example(seed: tp.Optional[int] = None) -> tp.Iterator[Experim
 def new_multiobjective_example(seed: tp.Optional[int] = None) -> tp.Iterator[Experiment]:
     """Optimization of 2 and 3 objective functions in Sphere, Ellipsoid, Cigar, Hm.
     Dimension 6 and 7.
-    Budget 2000, 2400, 2800, 3200, 3600, 4000.
+    Budget 100 to 3200
     """
     seedg = create_seed_generator(seed)
-    optims = ["NaiveTBPSA", "PSO", "DE", "LhsDE", "RandomSearch", "NGO", "Shiwa", "DiagonalCMA",
-              "CMA", "OnePlusOne", "TwoPointsDE"]
+    optims: tp.List[tp.Any] = ["NaiveTBPSA", "PSO", "DE", "LhsDE", "RandomSearch", "NGO", "Shiwa", "DiagonalCMA",
+                               "CMA", "OnePlusOne", "TwoPointsDE"]
+    optims += [ng.families.DifferentialEvolution(multiobjective_adaptation=False).set_name("DE-noadapt"),
+               ng.families.DifferentialEvolution(crossover="twopoints", multiobjective_adaptation=False).set_name("TwoPointsDE-noadapt")]
     mofuncs: tp.List[MultiExperiment] = []
-    for name1 in ["sphere", "cigar"]:
-        for name2 in ["sphere", "cigar", "hm"]:
-            mofuncs.append(MultiExperiment([ArtificialFunction(name1, block_dimension=7),
-                                            ArtificialFunction(name2, block_dimension=7)],
-                                           upper_bounds=np.array((50., 50.))))
-            for name3 in ["sphere", "ellipsoid"]:
-                mofuncs.append(MultiExperiment([ArtificialFunction(name1, block_dimension=6),
-                                                ArtificialFunction(name3, block_dimension=6),
-                                                ArtificialFunction(name2, block_dimension=6)],
-                                               upper_bounds=np.array((100, 100, 1000.))))
+    for name1, name2 in itertools.product(["sphere"], ["sphere", "hm"]):
+        mofuncs.append(MultiExperiment([ArtificialFunction(name1, block_dimension=7),
+                                        ArtificialFunction(name2, block_dimension=7)],
+                                       upper_bounds=[100, 100]))
+        mofuncs.append(MultiExperiment([ArtificialFunction(name1, block_dimension=6),
+                                        ArtificialFunction("sphere", block_dimension=6),
+                                        ArtificialFunction(name2, block_dimension=6)],
+                                       upper_bounds=[100, 100, 100.]))
     for mofunc in mofuncs:
         for optim in optims:
-            for budget in list(range(2000, 4001, 400)):
+            for budget in [100, 200, 400, 800, 1600, 3200]:
                 for nw in [1, 100]:
                     yield Experiment(mofunc, optim, budget=budget, num_workers=nw, seed=next(seedg))
 

diff --git a/nevergrad/benchmark/plotting.py b/nevergrad/benchmark/plotting.py
@@ -313,8 +313,12 @@ def create_plots(
             description = description[:140] + hash_ + description[-140:]
         out_filepath = output_folder / "xpresults{}{}.png".format("_" if description else "", description.replace(":", ""))
         data = XpPlotter.make_data(subdf)
-        xpplotter = XpPlotter(data, title=description, name_style=name_style, xaxis=xpaxis)
-        xpplotter.save(out_filepath)
+        try:
+            xpplotter = XpPlotter(data, title=description, name_style=name_style, xaxis=xpaxis)
+        except Exception as e:
+            warnings.warn(f"Bypassing error in xpplotter:\n{e}", RuntimeWarning)
+        else:
+            xpplotter.save(out_filepath)
     plt.close("all")
 
 

diff --git a/nevergrad/benchmark/test_xpbase.py b/nevergrad/benchmark/test_xpbase.py
@@ -130,3 +130,13 @@ def test_equality() -> None:
     xp1 = xpbase.Experiment(func, optimizer="OnePlusOne", budget=300, num_workers=2)
     xp2 = xpbase.Experiment(func, optimizer="RandomSearch", budget=300, num_workers=2)
     assert xp1 != xp2
+
+
+def test_multiobjective_experiment() -> None:
+    mofunc = MultiExperiment([ArtificialFunction("sphere", block_dimension=7),
+                              ArtificialFunction("cigar", block_dimension=7)],
+                             upper_bounds=np.array((50., 50.)))
+    xp = xpbase.Experiment(mofunc, optimizer="TwoPointsDE", budget=100, num_workers=1)
+    summary = xp.run()
+    loss: float = summary["loss"]
+    assert loss < 1e9
diff --git a/nevergrad/benchmark/xpbase.py b/nevergrad/benchmark/xpbase.py
@@ -246,7 +246,7 @@ def _run_with_error(self, callbacks: tp.Optional[tp.Dict[str, obase._OptimCallBa
                     self._optimizer,
                     pfunc,
                     batch_mode=executor.batch_mode,
-                    executor=executor
+                    executor=executor,
                 )
             except Exception as e:  # pylint: disable=broad-except
                 self.recommendation = self._optimizer.provide_recommendation()  # get the recommendation anyway

diff --git a/nevergrad/optimization/base.py b/nevergrad/optimization/base.py
@@ -14,7 +14,7 @@
 from nevergrad.common import tools as ngtools
 from nevergrad.common.decorators import Registry
 from . import utils
-from .multiobjective import HypervolumePareto
+from . import multiobjective as mobj
 
 
 registry: Registry[tp.Union["ConfiguredOptimizer", tp.Type["Optimizer"]]] = Registry()
@@ -112,12 +112,14 @@ def __init__(self, parametrization: IntOrParameter, budget: tp.Optional[int] = N
             num_workers=num_workers, dimension=self.parametrization.dimension
         )
         # multiobjective
-        self._hypervolume_pareto: tp.Optional[HypervolumePareto] = None
+        self._MULTIOBJECTIVE_AUTO_BOUND = mobj.AUTO_BOUND
+        self._hypervolume_pareto: tp.Optional[mobj.HypervolumePareto] = None
         # instance state
         self._asked: tp.Set[str] = set()
+        self._first_tell_done = False  # set to True at the beginning of the first tell
         self._suggestions: tp.Deque[p.Parameter] = deque()
         self._num_ask = 0
-        self._num_tell = 0
+        self._num_tell = 0  # increases after each successful tell
         self._num_tell_not_asked = 0
         self._callbacks: tp.Dict[str, tp.List[tp.Any]] = {}
         # to make optimize function stoppable halway through
@@ -139,7 +141,7 @@ def dimension(self) -> int:
 
     @property
     def num_objectives(self) -> int:
-        if not self._num_tell and self._hypervolume_pareto is None:
+        if not self._first_tell_done:
             raise RuntimeError('Unknown number of objectives, provide a "tell" first.')
         return 1 if self._hypervolume_pareto is None else self._hypervolume_pareto.num_objectives
 
@@ -184,9 +186,13 @@ def pareto_front(
         --------
         list
             the list of Parameter of the pareto front
+
+        Note
+        ----
+        During non-multiobjective optimization, this returns the current pessimistic best
         """
         if self._hypervolume_pareto is None:
-            raise RuntimeError("No pareto front with a single objective")
+            return [self.current_bests["pessimistic"].parameter]
         return self._hypervolume_pareto.pareto_front(size=size, subset=subset, subset_tentatives=subset_tentatives)
 
     def dump(self, filepath: tp.Union[str, Path]) -> None:
@@ -302,16 +308,14 @@ def tell(self, candidate: p.Parameter, loss: tp.Loss) -> None:
             )
         # checks are done, start processing
         candidate.freeze()  # make sure it is not modified somewhere
-        # call callbacks for logging etc...
-        for callback in self._callbacks.get("tell", []):
-            callback(self, candidate, loss)
+        self._first_tell_done = True
         # add reference if provided
         if isinstance(candidate, p.MultiobjectiveReference):
             if self._hypervolume_pareto is not None:
                 raise RuntimeError("MultiobjectiveReference can only be provided before the first tell.")
             if not isinstance(loss, np.ndarray):
                 raise RuntimeError("MultiobjectiveReference must only be used for multiobjective losses")
-            self._hypervolume_pareto = HypervolumePareto(upper_bounds=loss)
+            self._hypervolume_pareto = mobj.HypervolumePareto(upper_bounds=loss)
             if candidate.value is None:
                 return
             candidate = candidate.value
@@ -320,8 +324,13 @@ def tell(self, candidate: p.Parameter, loss: tp.Loss) -> None:
             candidate._losses = loss
         if not isinstance(loss, float):
             loss = self._preprocess_multiobjective(candidate)
+        # call callbacks for logging etc...
         candidate.loss = loss
         assert isinstance(loss, float)
+        for callback in self._callbacks.get("tell", []):
+            # multiobjective reference is not handled :s
+            # but this allows obtaining both scalar and multiobjective loss (through losses)
+            callback(self, candidate, loss)
         if isinstance(loss, float):
             self._update_archive_and_bests(candidate, loss)
         if candidate.uid in self._asked:
@@ -334,7 +343,7 @@ def tell(self, candidate: p.Parameter, loss: tp.Loss) -> None:
 
     def _preprocess_multiobjective(self, candidate: p.Parameter) -> tp.FloatLoss:
         if self._hypervolume_pareto is None:
-            self._hypervolume_pareto = HypervolumePareto()
+            self._hypervolume_pareto = mobj.HypervolumePareto(auto_bound=self._MULTIOBJECTIVE_AUTO_BOUND)
         return self._hypervolume_pareto.add(candidate)
 
     def _update_archive_and_bests(self, candidate: p.Parameter, loss: tp.FloatLoss) -> None:

diff --git a/nevergrad/optimization/callbacks.py b/nevergrad/optimization/callbacks.py
@@ -8,9 +8,9 @@
 import warnings
 import inspect
 import datetime
-import typing as tp
 from pathlib import Path
 import numpy as np
+import nevergrad.common.typing as tp
 from nevergrad.parametrization import parameter as p
 from nevergrad.parametrization import helpers
 from . import base
@@ -80,7 +80,7 @@ def __init__(self, filepath: tp.Union[str, Path], append: bool = True, order: in
             self._filepath.unlink()  # missing_ok argument added in python 3.8
         self._filepath.parent.mkdir(exist_ok=True, parents=True)
 
-    def __call__(self, optimizer: base.Optimizer, candidate: p.Parameter, value: float) -> None:
+    def __call__(self, optimizer: base.Optimizer, candidate: p.Parameter, loss: tp.FloatLoss) -> None:
         data = {"#parametrization": optimizer.parametrization.name,
                 "#optimizer": optimizer.name,
                 "#session": self._session,
@@ -91,7 +91,10 @@ def __call__(self, optimizer: base.Optimizer, candidate: p.Parameter, value: flo
                 "#lineage": candidate.heritage["lineage"],
                 "#generation": candidate.generation,
                 "#parents_uids": [],
-                "#loss": value}
+                "#loss": loss}
+        if optimizer.num_objectives > 1:  # multiobjective losses
+            data.update({f"#losses#{k}": val for k, val in enumerate(candidate.losses)})
+            data["#pareto-length"] = len(optimizer.pareto_front())
         if hasattr(optimizer, "_configured_optimizer"):
             configopt = optimizer._configured_optimizer  # type: ignore
             if isinstance(configopt, base.ConfiguredOptimizer):