Improve tests

Author: vlad-moroshan

tests/test_optimization_over_graphs.py

@@ -19,7 +19,7 @@
 from grakel_replace.utils import min_max_scale
 
 
-class TestPipeline:
+class TestGraphOptimizationPipeline:
     @pytest.fixture
     def setup_data(self):
         """Fixture to set up common data for tests."""
@@ -79,11 +79,13 @@ def test_gp_fit_and_predict(self, setup_data):
         kernels = [
             ScaleKernel(MaternKernel(nu=2.5, ard_num_dims=setup_data["N_NUMERICAL"],
                                      active_dims=range(setup_data["N_NUMERICAL"]))),
-            ScaleKernel(CategoricalKernel(ard_num_dims=setup_data["N_CATEGORICAL"],
-                                          active_dims=range(setup_data["N_NUMERICAL"],
-                                                            setup_data["N_NUMERICAL"] +
-                                                            setup_data[
-                                                                "N_CATEGORICAL"]))),
+            ScaleKernel(
+                CategoricalKernel(ard_num_dims=setup_data["N_CATEGORICAL"],
+                                  active_dims=range(setup_data["N_NUMERICAL"],
+                                                    setup_data["N_NUMERICAL"] +
+                                                    setup_data["N_CATEGORICAL"])
+                                  )
+            ),
             ScaleKernel(
                 BoTorchWLKernel(graph_lookup=train_graphs, n_iter=5, normalize=True,
                                 active_dims=(train_x.shape[1] - 1,)))

tests/test_torch_wl_kernel.py

@@ -11,7 +11,7 @@ class TestTorchWLKernel:
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
     @pytest.fixture
-    def example_graphs(self):
+    def example_graphs_set(self):
         # Create example graphs for testing
         G1 = nx.Graph()
         G1.add_edges_from([(0, 1), (1, 2), (1, 3), (2, 3), (3, 4)])
@@ -30,35 +30,67 @@ def example_graphs(self):
 
         return [G1, G2, G3]
 
-    @pytest.mark.parametrize("n_iter", [1, 2, 3, 5, 10])
-    @pytest.mark.parametrize("normalize", [True, False])
-    def test_wl_kernel_against_grakel(self, n_iter, normalize, example_graphs):
-        adjacency_matrices, label_tensors = graphs_to_tensors(
-            example_graphs, device=self.device)
+    @pytest.fixture
+    def random_graphs_sets(self):
+        # Set a seed for reproducibility
+        seed = 100
+        np.random.seed(seed)
+        torch.manual_seed(seed)
+        random_graph_sets = []
 
-        # Initialize Torch WL Kernel
-        torch_kernel = TorchWLKernel(n_iter=n_iter, normalize=normalize)
-        torch_kernel_matrix = torch_kernel(adjacency_matrices,
-                                           label_tensors).cpu().numpy()
+        # Generate 10 random sets of graphs
+        for _ in range(10):
+            # Number of graphs in the set (2 to 10)
+            num_graphs = np.random.randint(2, 11)
+            graph_set = []
 
-        # Initialize GraKel WL Kernel
-        grakel_graphs = list(
-            graph_from_networkx(example_graphs, node_labels_tag="label", as_Graph=True))
-        grakel_kernel = WeisfeilerLehman(n_iter=n_iter, normalize=normalize)
-        grakel_kernel_matrix = grakel_kernel.fit_transform(grakel_graphs)
+            for _ in range(num_graphs):
+                # Number of nodes in the graph (3 to 50)
+                num_nodes = np.random.randint(3, 51)
+                G = nx.Graph()
 
-        # Define tolerances based on normalization
-        rtol = 1e-5 if normalize else 1e-4
-        atol = 1e-8 if normalize else 1e-7
+                # Add nodes with labels
+                for node in range(num_nodes):
+                    G.add_node(node, label=str(node))
 
-        # Compare the kernel matrices
-        np.testing.assert_allclose(
-            torch_kernel_matrix,
-            grakel_kernel_matrix,
-            rtol=rtol,
-            atol=atol,
-            err_msg=f"Kernel matrices differ for n_iter={n_iter}, normalize={normalize}"
-        )
+                # Add random edges
+                for u in range(num_nodes):
+                    for v in range(u + 1, num_nodes):
+                        if np.random.rand() > 0.5:  # 50% chance to add an edge
+                            G.add_edge(u, v)
+
+                graph_set.append(G)
+
+            random_graph_sets.append(graph_set)
+
+        return random_graph_sets
+
+    @pytest.mark.parametrize("n_iter", [1, 2, 3, 5, 10])
+    @pytest.mark.parametrize("normalize", [True, False])
+    def test_wl_kernel_against_grakel(self, n_iter, normalize, random_graphs_sets):
+        for graph_set in random_graphs_sets:
+            adjacency_matrices, label_tensors = graphs_to_tensors(
+                graph_set, device=self.device)
+
+            # Initialize Torch WL Kernel
+            torch_kernel = TorchWLKernel(n_iter=n_iter, normalize=normalize)
+            torch_kernel_matrix = torch_kernel(adjacency_matrices,
+                                               label_tensors).cpu().numpy()
+
+            # Initialize GraKel WL Kernel
+            grakel_graphs = list(
+                graph_from_networkx(graph_set, node_labels_tag="label", as_Graph=True))
+            grakel_kernel = WeisfeilerLehman(n_iter=n_iter, normalize=normalize)
+            grakel_kernel_matrix = grakel_kernel.fit_transform(grakel_graphs)
+
+            # Compare the kernel matrices
+            np.testing.assert_allclose(
+                torch_kernel_matrix,
+                grakel_kernel_matrix,
+                rtol=1e-5,
+                atol=1e-8,
+                err_msg=f"Kernel matrices differ for graph={graph_set}, n_iter={n_iter}"
+            )
 
     def test_empty_graph(self):
         G_empty = nx.Graph()
@@ -97,36 +129,47 @@ def test_kernel_on_single_node_graph(self):
         expected = torch.ones(1, 1, device=self.device)
         torch.testing.assert_close(K, expected)
 
-    def test_wl_kernel_with_empty_graph_and_reordered_edges(self, example_graphs):
+    def test_wl_kernel_with_empty_graph_and_reordered_edges(self, random_graphs_sets):
         """Test the TorchWLKernel with an empty graph and a graph with reordered edges."""
-        # Create example graphs for testing
-        G_empty = nx.Graph()
-        G_empty.add_node(0)
-        G_empty.nodes[0]["label"] = "0"
-
-        G = example_graphs[0]
-        G_reordered = nx.Graph()
-        G_reordered.add_edges_from([(1, 4), (2, 3), (1, 2), (0, 1), (1, 3)])
-        for node in G_reordered.nodes():
-            G_reordered.nodes[node]["label"] = str(node)
-
-        graphs = [G_empty, G, G_reordered]
-        adjacency_matrices, label_tensors = graphs_to_tensors(graphs,
-                                                              device=self.device)
-
-        wl_kernel = TorchWLKernel(n_iter=3, normalize=True)
-        K = wl_kernel(adjacency_matrices, label_tensors)
-
-        assert K.shape == (3, 3), "Kernel matrix shape is incorrect"
-        assert K[1, 1] == K[
-            2, 2], "Kernel value for original and reordered graphs should be the same"
+        for graph_set in random_graphs_sets:
+            # Create an empty graph
+            G_empty = nx.Graph()
+            G_empty.add_node(0)
+            G_empty.nodes[0]["label"] = "0"
+
+            # Select the first graph from the set to reorder its edges
+            G = graph_set[0]
+            G_reordered = nx.Graph()
+
+            # Add all nodes from the original graph to G_reordered
+            for node in G.nodes():
+                G_reordered.add_node(node, label=G.nodes[node]["label"])
+
+            # Reorder edges randomly
+            edges = list(G.edges())
+            np.random.shuffle(edges)  # Randomly shuffle the edges
+            G_reordered.add_edges_from(edges)
+
+            # Combine the empty graph, original graph, and reordered graph
+            graphs = [G_empty, G, G_reordered]
+            adjacency_matrices, label_tensors = graphs_to_tensors(
+                graphs, device=self.device
+            )
+
+            # Initialize and compute the kernel
+            wl_kernel = TorchWLKernel(n_iter=3, normalize=True)
+            K = wl_kernel(adjacency_matrices, label_tensors)
+
+            assert K.shape == (3, 3), "Kernel matrix shape is incorrect"
+            assert torch.allclose(K[1, 1], K[2, 2]), \
+                "Kernel value for original and reordered graphs should be the same"
 
     @pytest.mark.parametrize("n_iter", [1, 2, 3, 4, 5, 6, 7])
     @pytest.mark.parametrize("normalize", [True, False])
     def test_wl_kernel_with_different_node_labels(self, n_iter, normalize,
-                                                  example_graphs):
+                                                  example_graphs_set):
         graphs = []
-        for i, G in enumerate(example_graphs):
+        for i, G in enumerate(example_graphs_set):
             G_copy = G.copy()
             prefix = ["node_", "vertex_", "n"][i]
             for node in G_copy.nodes():
@@ -143,20 +186,15 @@ def test_wl_kernel_with_different_node_labels(self, n_iter, normalize,
         grakel_wl = WeisfeilerLehman(n_iter=n_iter, normalize=normalize)
         grakel_kernel_matrix = grakel_wl.fit_transform(grakel_graphs)
 
-        # Define tolerances based on normalization, matching the main test
-        rtol = 1e-5 if normalize else 1e-4
-        atol = 1e-8 if normalize else 1e-7
-
-        # Updated assertion with both rtol and atol
         np.testing.assert_allclose(
             torch_kernel_matrix,
             grakel_kernel_matrix,
-            rtol=rtol,
-            atol=atol,
+            rtol=1e-5,
+            atol=1e-8,
             err_msg=f"Kernel matrices differ for n_iter={n_iter}, normalize={normalize}"
         )
 
-    def test_wl_kernel_with_same_node_labels(self, example_graphs):
+    def test_wl_kernel_with_same_node_labels(self, example_graphs_set):
         """Test WL kernel behavior with same node labels but different structures.
 
         Even when all nodes have the same label, the WL kernel should:
@@ -166,7 +204,7 @@ def test_wl_kernel_with_same_node_labels(self, example_graphs):
         4. Maintain non-negative values (it's a valid kernel)
         """
         graphs = []
-        for G in example_graphs:
+        for G in example_graphs_set:
             G_copy = G.copy()
             for node in G_copy.nodes():
                 G_copy.nodes[node]["label"] = "A"