Iterations per job has different semantics for comb MCShap

Author: mdbenito

src/pydvl/shapley/montecarlo.py

@@ -322,7 +322,6 @@ def combinatorial_montecarlo_shapley(
     """
     parallel_backend = init_parallel_backend(config)
     u_id = parallel_backend.put(u)
-    iterations_per_job = max_iterations // n_jobs
 
     def reducer(results_it: Iterable[MonteCarloResults]) -> MonteCarloResults:
         values = np.zeros(len(u.data))
@@ -335,13 +334,14 @@ def reducer(results_it: Iterable[MonteCarloResults]) -> MonteCarloResults:
             stderr += std
         return MonteCarloResults(values=values, stderr=stderr)
 
+    # FIXME? max_iterations has different semantics in permutation-based methods
     map_reduce_job: MapReduceJob["NDArray", MonteCarloResults] = MapReduceJob(
         map_func=_combinatorial_montecarlo_shapley,
         reduce_func=reducer,
         map_kwargs=dict(
             u=u_id,
             dist=dist,
-            max_iterations=iterations_per_job,
+            max_iterations=max_iterations,
             progress=progress,
         ),
         chunkify_inputs=True,

tests/shapley/test_montecarlo.py

@@ -53,11 +53,6 @@ def test_analytic_montecarlo_shapley(
 def test_hoeffding_bound_montecarlo(
     analytic_shapley, fun, n_jobs, delta, eps, tolerate
 ):
-    """FIXME: This test passes but there are several unclear points.
-    For example, map_reduce is called with num_jobs=num_runs. Is this correct?
-    If I put num_jobs=jobs_per_run, map_reduce encounters errors since a utility
-    is passed.
-    Before coming back to this test, fix map_reduce interface."""
     u, exact_values = analytic_shapley
 
     max_iterations = lower_bound_hoeffding(delta=delta, eps=eps, score_range=1)
@@ -93,8 +88,8 @@ def test_hoeffding_bound_montecarlo(
             12,
             combinatorial_montecarlo_shapley,
             "explained_variance",
-            0.5,
-            5000,
+            0.2,
+            2**11,
         ),
     ],
 )
@@ -116,7 +111,7 @@ def test_linear_montecarlo_shapley(
     )
     values, _ = fun(
         linear_utility,
-        max_iterations=max_iterations,
+        max_iterations=int(max_iterations),
         progress=False,
         n_jobs=n_jobs,
     )