Add utility function for batched simulations

bayesflow/simulators/simulator.py

@@ -66,3 +66,32 @@ def rejection_sample(
                 result = tree_concatenate([result, samples], axis=axis, numpy=True)
 
         return result
+
+    @allow_batch_size
+    def sample_batched(
+        self,
+        batch_shape: Shape,
+        *,
+        sample_size: int,
+        **kwargs,
+    ):
+        """Sample the desired number of simulations in smaller batches.
+
+        Limited resources, especially memory, can make it necessary to run simulations in smaller batches.
+        The number of samples per simulated batch is specified by `sample_size`.
+
+        Parameters
+        ----------
+        batch_shape : Shape
+            The desired output shape, as in :py:meth:`sample`. Will be rounded up to the next complete batch.
+        sample_size : int
+            The number of samples in each simulated batch.
+        kwargs
+            Additional keyword arguments passed to :py:meth:`sample`.
+
+        """
+
+        def accept_all_predicate(x):
+            return np.full((sample_size,), True)
+
+        return self.rejection_sample(batch_shape, predicate=accept_all_predicate, sample_size=sample_size, **kwargs)

tests/test_simulators/test_simulators.py

@@ -41,6 +41,26 @@ def test_sample(simulator, batch_size):
         assert not np.allclose(value, value[0])
 
 
+def test_sample_batched(simulator, batch_size):
+    sample_size = 2
+    samples = simulator.sample_batched((batch_size,), sample_size=sample_size)
+
+    # test output structure
+    assert isinstance(samples, dict)
+
+    for key, value in samples.items():
+        print(f"{key}.shape = {keras.ops.shape(value)}")
+
+        # test type
+        assert isinstance(value, np.ndarray)
+
+        # test shape (sample_batched rounds up to complete batches)
+        assert value.shape[0] == int(np.ceil(batch_size / sample_size)) * sample_size
+
+        # test batch randomness
+        assert not np.allclose(value, value[0])
+
+
 def test_fixed_sample(composite_gaussian, batch_size, fixed_n, fixed_mu):
     samples = composite_gaussian.sample((batch_size,), n=fixed_n, mu=fixed_mu)