Hybrid Graph Dataset

[pouya-taghipour]

Hello everyone,

I need to create a PyTorch Geometric dataset so that each sample consists of two graphs but with the same label.
I know the naive solution is to create two separate datasets and load them separately to feed into my model.

But, is there any way to make a single dataset?

it should be noted that the node and edge feature size/structure in these two datasets is different.

Here is my Example Code:

# A function to generate a random graph
def generate_random_graph(node_min, node_max, node_feature_size, edge_feature_size, bias=0):
    num_nodes = torch.randint(node_min, node_max, size=(1,)).item()
    num_edges = torch.randint(num_nodes, num_nodes*(num_nodes-1)//2, size=(1,)).item()
    graph = nx.gnm_random_graph(num_nodes, num_edges)
    
    # Generate and set random features for nodes
    node_features=dict()
    for i in range(graph.number_of_nodes()):
        temp = []
        for a in range(node_feature_size):
            temp.append(round(torch.rand(1).item()+bias, 3))
        node_features[i]={'x':temp}
    nx.set_node_attributes(graph, node_features)

    # Generate and set random features for edges
    edge_features = dict()
    for u,v,a in graph.edges(data=True):
        temp = []
        for b in range(edge_feature_size):
            temp.append(round(torch.rand(1).item()+bias, 3))
            edge_features[(u,v)]={'edge_attr':temp}
    nx.set_edge_attributes(graph, edge_features)
    # label = torch.rand(1).item()*bias
    return graph,node_features,edge_features

# List of graphs in PyTorch geometric Data format
graph_list1 = []
graph_list2 = []

#List of labels (same for both graphs)
label_list = []

# Generate 2 set of 20 graphs
for i in range(20):
    graph1, node_features1, edge_features1 = generate_random_graph(5,30,6,2,i)
    graph2, node_features2, edge_features2 = generate_random_graph(7,60,5,3,i)

    # Set labels
    label_random = torch.rand(1).item()*i
    graph1.y = graph2.y = label_random

    # Convert to Pytorch Geometric Data Format
    graph_list1.append(from_networkx(graph1))
    graph_list2.append(from_networkx(graph2))
    label_list.append(label_random)

[wsad1]

Check if this https://pytorch-geometric.readthedocs.io/en/latest/advanced/batching.html#pairs-of-graphs helps.