Fix batching network inputs of mixed type

Author: Robin Manhaeve

src/deepproblog/network.py

@@ -17,13 +17,13 @@ class Network(object):
     """Wraps a PyTorch neural network for use with DeepProblog"""
 
     def __init__(
-        self,
-        network_module: torch.nn.Module,
-        name: str,
-        optimizer: Optional[torch.optim.Optimizer] = None,
-        scheduler=None,
-        k: Optional[int] = None,
-        batching: bool = False,
+            self,
+            network_module: torch.nn.Module,
+            name: str,
+            optimizer: Optional[torch.optim.Optimizer] = None,
+            scheduler=None,
+            k: Optional[int] = None,
+            batching: bool = False,
     ):
         """Create a Network object
 
@@ -121,13 +121,17 @@ def __call__(self, to_evaluate: list) -> list:
         :return:
         """
         if self.batching:
-            batched_inputs: List[torch.Tensor] = [
-                self.function(*e)[0] for e in to_evaluate
-            ]
-            stacked_inputs = torch.stack(batched_inputs)
-            if self.is_cuda:
-                stacked_inputs = stacked_inputs.cuda(device=self.device)
-            evaluated = self.network_module(stacked_inputs)
+            inputs = (self.function(*e) for e in to_evaluate)
+            stacked_inputs = list()
+            for inputs in zip(*inputs):
+                try:
+                    inputs = torch.stack(inputs)
+                    if self.is_cuda:
+                        inputs.cuda(device=self.device)
+                except TypeError:
+                    inputs = list(inputs)
+                stacked_inputs.append(inputs)
+            evaluated = self.network_module(*stacked_inputs)
         else:
             evaluated = [self.network_module(*self.function(*e)) for e in to_evaluate]
         return evaluated
@@ -169,7 +173,6 @@ def get_hyperparameters(self):
         }
         return parameters
 
-
 # class NetworkEvaluation(object):
 #     """
 #     An object that keeps track of which inputs the neural networks need to be evaluated on.

src/deepproblog/tests/test_neural_predicate.py

@@ -12,7 +12,7 @@
 nn(dummy1,[X],Y,[a,b,c]) :: net1(X,Y).
 nn(dummy2,[X]) :: net2(X).
 nn(dummy3,[X],Y) :: net3(X,Y).
-nn(dummy4,[X,Y],Z,[a,b]) :: net4(X,Y,Z).
+nn(dummy4,[X,Y]) :: net4(X,Y).
 
 test1(X1,Y1,X2,Y2) :- net1(X1,Y1), net1(X2,Y2).
 test2(X1,X2) :- net2(X1), net2(X2).
@@ -28,9 +28,19 @@
 dummy_values3 = {Term("i1"): [1.0, 2.0, 3.0, 4.0], Term("i2"): [-1.0, 0.0, 1.0]}
 dummy_net3 = Network(DummyNet(dummy_values3), "dummy3")
 
-dummy_net4 = Network(DummyTensorNet(batching=True), "dummy4", batching=True)
 
-tensors = {(Constant(0),): torch.Tensor([0.2]), (Constant(1),): torch.Tensor([0.8])}
+dummy_tensors = {(Term("a"),): torch.Tensor([0.1, 0.2, 0.3, 0.4]), (Term("b"),): torch.Tensor([0.25, 0.25, 0.25, 0.25])}
+
+
+class IndexNet(torch.nn.Module):
+
+    def forward(self, t, index):
+        # index = int(index)
+        index = torch.LongTensor([int(i) for i in index])
+        return t.index_select(dim=1, index=index)
+
+
+dummy_net4 = Network(IndexNet(), "dummy4", batching=True)
 
 
 @pytest.fixture(
@@ -53,7 +63,7 @@ def model(request) -> Model:
     model = Model(program, [dummy_net1, dummy_net2, dummy_net3, dummy_net4], load=False)
     engine = request.param["engine_factory"](model)
     model.set_engine(engine, cache=request.param["cache"])
-    model.add_tensor_source('dummy', tensors)
+    model.add_tensor_source('dummy', dummy_tensors)
     return model
 
 
@@ -108,13 +118,12 @@ def test_det_network_substitution(model: Model):
         assert all(r1.detach().numpy() == [1.0, 2.0, 3.0, 4.0])
         assert all(r2.detach().numpy() == [-1.0, 0.0, 1.0])
 
-def test_double_input(model: Model):
-    terms = lambda x: Term("net4",
-                           Term("tensor",Term("dummy", Constant(0))),
-                           Term("tensor",Term("dummy", Constant(1))),
-                           x)
-    results = model.solve([Query(terms(Var("X")))])
-    r1 = float(results[0].result[terms(Term("a"))])
-    r2 = float(results[0].result[terms(Term("b"))])
+def test_multi_input_network(model: Model):
+    dummy_tensor = lambda x: Term("tensor", Term("dummy", x))
+    q1 = Query(Term("net4", dummy_tensor(Term("a")), Constant(1)))
+    q2 = Query(Term("net4", dummy_tensor(Term("b")), Constant(2)))
+    results = model.solve([q1, q2])
+    r1 = float(results[0].result[q1.query])
+    r2 = float(results[1].result[q2.query])
     assert pytest.approx(0.2) == r1
-    assert pytest.approx(0.8) == r2
+    assert pytest.approx(0.25) == r2