update OptimizationLogger and callback tests for MOO (#1209)

* updated callback tests for MOO

* removed stray print

* Delete =

* updated optimization logger for MOO and added get_min_losses function

* edited get_min_losses

* type fixes

* removed testing statement

* removed testing statement

* more efficient get min losses

* added optimization logger test for single objective

Co-authored-by: Theodore Clarke <theoajclarke@devfair0149.h2.fair>

nevergrad/optimization/callbacks.py

@@ -87,8 +87,17 @@ def __call__(self, optimizer: base.Optimizer, *args: tp.Any, **kwargs: tp.Any) -
         if time.time() >= self._next_time or self._next_tell >= optimizer.num_tell:
             self._next_time = time.time() + self._log_interval_seconds
             self._next_tell = optimizer.num_tell + self._log_interval_tells
-            x = optimizer.provide_recommendation()
-            self._logger.log(self._log_level, "After %s, recommendation is %s", optimizer.num_tell, x)
+            if optimizer.num_objectives == 1:
+                x = optimizer.provide_recommendation()
+                self._logger.log(self._log_level, "After %s, recommendation is %s", optimizer.num_tell, x)
+            else:
+                losses = optimizer._hypervolume_pareto.get_min_losses()  # type: ignore
+                self._logger.log(
+                    self._log_level,
+                    "After %s, the respective minimum loss for each objective in the pareto front is %s",
+                    optimizer.num_tell,
+                    losses,
+                )
 
 
 class ParametersLogger:

nevergrad/optimization/multiobjective/core.py

@@ -174,6 +174,9 @@ def pareto_front(
         """
         return self._pf.get_front(size, subset, subset_tentatives)
 
+    def get_min_losses(self) -> tp.List[float]:
+        return np.min([p.losses for p in self._pf.get_raw()], axis=0)
+
 
 class ParetoFront:
     def __init__(

nevergrad/optimization/test_callbacks.py

@@ -103,20 +103,51 @@ def test_progressbar_dump(tmp_path: Path) -> None:
         optimizer.tell(cand, 0)
 
 
+class _EarlyStoppingTestee:
+    def __init__(self) -> None:
+        self.num_calls = 0
+
+    def __call__(self, *args, **kwds) -> float:
+        self.num_calls += 1
+        return np.random.rand()
+
+
 def test_early_stopping() -> None:
-    instrum = ng.p.Instrumentation(
-        None, 2.0, blublu="blublu", array=ng.p.Array(shape=(3, 2)), multiobjective=True
-    )
+    instrum = ng.p.Instrumentation(None, 2.0, blublu="blublu", array=ng.p.Array(shape=(3, 2)))
+    func = _EarlyStoppingTestee()
     optimizer = optimizerlib.OnePlusOne(parametrization=instrum, budget=100)
     early_stopping = ng.callbacks.EarlyStopping(lambda opt: opt.num_ask > 3)
     optimizer.register_callback("ask", early_stopping)
-    optimizer.minimize(_func, verbosity=2)
-    # below functions are inlcuded in the docstring
+    optimizer.minimize(func, verbosity=2)
+    # num_ask is set at the end of ask, so the callback sees the old value.
+    assert func.num_calls == 4
+    # below functions are included in the docstring of EarlyStopping
     assert optimizer.current_bests["minimum"].mean < 12
     assert optimizer.recommend().loss < 12  # type: ignore
 
 
 def test_optimization_logger(caplog) -> None:
+    instrum = ng.p.Instrumentation(
+        None, 2.0, blublu="blublu", array=ng.p.Array(shape=(3, 2)), multiobjective=False
+    )
+    logging.basicConfig(level=os.environ.get("LOGLEVEL", "INFO"))
+    logger = logging.getLogger(__name__)
+    optimizer = optimizerlib.OnePlusOne(parametrization=instrum, budget=3)
+    optimizer.register_callback(
+        "tell",
+        callbacks.OptimizationLogger(
+            logger=logger, log_level=logging.INFO, log_interval_tells=10, log_interval_seconds=0.1
+        ),
+    )
+    with caplog.at_level(logging.INFO):
+        optimizer.minimize(_func, verbosity=2)
+    assert (
+        "After 0, recommendation is Instrumentation(Tuple(None,2.0),Dict(array=Array{(3,2)},blublu=blublu,multiobjective=False))"
+        in caplog.text
+    )
+
+
+def test_optimization_logger_MOO(caplog) -> None:
     instrum = ng.p.Instrumentation(
         None, 2.0, blublu="blublu", array=ng.p.Array(shape=(3, 2)), multiobjective=True
     )
@@ -132,6 +163,6 @@ def test_optimization_logger(caplog) -> None:
     with caplog.at_level(logging.INFO):
         optimizer.minimize(_func, verbosity=2)
     assert (
-        "After 0, recommendation is Instrumentation(Tuple(None,2.0),Dict(array=Array{(3,2)},blublu=blublu,multiobjective=True))"
+        "After 0, the respective minimum loss for each objective in the pareto front is [12. 12.]"
         in caplog.text
     )