ModelComparisonSimulator: handle different outputs from individual simulators

bayesflow/simulators/model_comparison_simulator.py

@@ -6,6 +6,7 @@
 from bayesflow.utils.decorators import allow_batch_size
 
 from bayesflow.utils import numpy_utils as npu
+from bayesflow.utils import logging
 
 from types import FunctionType
 
@@ -22,6 +23,8 @@
         p: Sequence[float] = None,
         logits: Sequence[float] = None,
         use_mixed_batches: bool = True,
+        key_conflicts: str = "drop",
+        fill_value: float = np.nan,
         shared_simulator: Simulator | Callable[[Sequence[int]], dict[str, any]] = None,
     ):
         """
@@ -38,11 +41,21 @@
             A sequence of logits corresponding to model probabilities. Mutually exclusive with `p`.
             If neither `p` nor `logits` is provided, defaults to uniform logits.
         use_mixed_batches : bool, optional
-            If True, samples in a batch are drawn from different models. If False, the entire batch
-            is drawn from a single model chosen according to the model probabilities. Default is True.
+            Whether to draw samples in a batch from different models.
+
+            - If True (default), each sample in a batch may come from a different model.
+            - If False, the entire batch is drawn from a single model, selected according to model probabilities.
+        key_conflicts : str, optional
+            Policy for handling keys that are missing in the output of some models, when using mixed batches.
+
+            - "drop" (default): Drop conflicting keys from the batch output.
+            - "fill": Fill missing keys with the specified value.
+            - "error": An error is raised when key conflicts are detected.
+        fill_value : float, optional
+            If `key_conflicts=="fill"`, the missing keys will be filled with the value of this argument.
         shared_simulator : Simulator or Callable, optional
             A shared simulator whose outputs are passed to all model simulators. If a function is
-            provided, it is wrapped in a `LambdaSimulator` with batching enabled.
+            provided, it is wrapped in a :py:class:`~bayesflow.simulators.LambdaSimulator` with batching enabled.
         """
         self.simulators = simulators
 
@@ -68,6 +81,9 @@
 
         self.logits = logits
         self.use_mixed_batches = use_mixed_batches
+        self.key_conflicts = key_conflicts
+        self.fill_value = fill_value
+        self._keys = None
 
     @allow_batch_size
     def sample(self, batch_shape: Shape, **kwargs) -> dict[str, np.ndarray]:
@@ -105,6 +121,7 @@
             sims = [
                 simulator.sample(n, **(kwargs | data)) for simulator, n in zip(self.simulators, model_counts) if n > 0
             ]
+            sims = self._handle_key_conflicts(sims, model_counts)
             sims = tree_concatenate(sims, numpy=True)
             data |= sims
 
@@ -118,3 +135,58 @@
             model_indices = npu.one_hot(np.full(batch_shape, model_index, dtype="int32"), num_models)
 
         return data | {"model_indices": model_indices}
+
+    def _handle_key_conflicts(self, sims, batch_sizes):
+        batch_sizes = [b for b in batch_sizes if b > 0]
+
+        keys, all_keys, common_keys, missing_keys = self._determine_key_conflicts(sims=sims)
+
+        # all sims have the same keys
+        if all_keys == common_keys:
+            return sims
+
+        if self.key_conflicts == "drop":
+            sims = [{k: v for k, v in sim.items() if k in common_keys} for sim in sims]
+            return sims
+        elif self.key_conflicts == "fill":
+            combined_sims = {}
+            for sim in sims:
+                combined_sims = combined_sims | sim
+            for i, sim in enumerate(sims):
+                for missing_key in missing_keys[i]:
+                    shape = combined_sims[missing_key].shape
+                    shape = list(shape)
+                    shape[0] = batch_sizes[i]
+                    sim[missing_key] = np.full(shape=shape, fill_value=self.fill_value)
+            return sims
+        elif self.key_conflicts == "error":
+            raise ValueError("Key conflicts are found in simulator outputs, cannot combine them into one batch.")
+
+    def _determine_key_conflicts(self, sims):
+        # determine only once
+        if self._keys is not None:
+            return self._keys
+
+        keys = [set(sim.keys()) for sim in sims]
+        all_keys = set.union(*keys)
+        common_keys = set.intersection(*keys)
+        missing_keys = [all_keys - k for k in keys]
+
+        self._keys = keys, all_keys, common_keys, missing_keys
+
+        if all_keys == common_keys:
+            return self._keys
+
+        if self.key_conflicts == "drop":
+            logging.info(
+                f"Incompatible simulator output. \
+The following keys will be dropped: {', '.join(sorted(all_keys - common_keys))}."
+            )
+        elif self.key_conflicts == "fill":
+            logging.info(
+                f"Incompatible simulator output. \
+Attempting to replace keys: {', '.join(sorted(all_keys - common_keys))}, where missing, \
+with value {self.fill_value}."
+            )
+
+        return self._keys

tests/test_simulators/conftest.py

@@ -167,6 +167,56 @@ def likelihood(mu, n):
     return make_simulator([prior, likelihood], meta_fn=context)
 
 
+@pytest.fixture()
+def multimodel():
+    from bayesflow.simulators import make_simulator, ModelComparisonSimulator
+
+    def context(batch_size):
+        return dict(n=np.random.randint(10, 100))
+
+    def prior_0():
+        return dict(mu=0)
+
+    def prior_1():
+        return dict(mu=np.random.standard_normal())
+
+    def likelihood(n, mu):
+        return dict(y=np.random.normal(mu, 1, n))
+
+    simulator_0 = make_simulator([prior_0, likelihood])
+    simulator_1 = make_simulator([prior_1, likelihood])
+
+    simulator = ModelComparisonSimulator(simulators=[simulator_0, simulator_1], shared_simulator=context)
+
+    return simulator
+
+
+@pytest.fixture(params=["drop", "fill", "error"])
+def multimodel_key_conflicts(request):
+    from bayesflow.simulators import make_simulator, ModelComparisonSimulator
+
+    rng = np.random.default_rng()
+
+    def prior_1():
+        return dict(w=rng.uniform())
+
+    def prior_2():
+        return dict(c=rng.uniform())
+
+    def model_1(w):
+        return dict(x=w)
+
+    def model_2(c):
+        return dict(x=c)
+
+    simulator_1 = make_simulator([prior_1, model_1])
+    simulator_2 = make_simulator([prior_2, model_2])
+
+    simulator = ModelComparisonSimulator(simulators=[simulator_1, simulator_2], key_conflicts=request.param)
+
+    return simulator
+
+
 @pytest.fixture()
 def fixed_n():
     return 5

tests/test_simulators/test_simulators.py

@@ -1,3 +1,4 @@
+import pytest
 import keras
 import numpy as np
 
@@ -47,3 +48,24 @@ def test_fixed_sample(composite_gaussian, batch_size, fixed_n, fixed_mu):
     assert samples["mu"].shape == (batch_size, 1)
     assert np.all(samples["mu"] == fixed_mu)
     assert samples["y"].shape == (batch_size, fixed_n)
+
+
+def test_multimodel_sample(multimodel, batch_size):
+    samples = multimodel.sample(batch_size)
+
+    assert set(samples) == {"n", "mu", "y", "model_indices"}
+    assert samples["mu"].shape == (batch_size, 1)
+    assert samples["y"].shape == (batch_size, samples["n"])
+
+
+def test_multimodel_key_conflicts_sample(multimodel_key_conflicts, batch_size):
+    if multimodel_key_conflicts.key_conflicts == "drop":
+        samples = multimodel_key_conflicts.sample(batch_size)
+        assert set(samples) == {"x", "model_indices"}
+    elif multimodel_key_conflicts.key_conflicts == "fill":
+        samples = multimodel_key_conflicts.sample(batch_size)
+        assert set(samples) == {"x", "model_indices", "c", "w"}
+        assert np.sum(np.isnan(samples["c"])) + np.sum(np.isnan(samples["w"])) == batch_size
+    elif multimodel_key_conflicts.key_conflicts == "error":
+        with pytest.raises(Exception):
+            samples = multimodel_key_conflicts.sample(batch_size)