[BUG] CGNN error: module 'networkx' has no attribute 'adj_matrix'

[JHPark9090]

Hello, I am trying to run CGNN model based on an undirected graph, using the "orient_directed_graph" function.
I am using the exact same codes presented in this example:
https://github.com/FenTechSolutions/CausalDiscoveryToolbox/blob/master/examples/example_cgnn.ipynb

Describe the bug
But whenever, I run the CGNN codes, I obtain this error, which indicates that the 'networkx' library has no attribute 'adj_matrix'.

I think it is maybe because the function 'adj_matrix' was deprecated in the 'networkx' library.

Below is the versions of my python and related libraries.

I am running my codes on GPUs.

[rafaela-amorim]

same here

[jason-huanghao]

Hello, I am trying to run CGNN model based on an undirected graph, using the "orient_directed_graph" function. I am using the exact same codes presented in this example: https://github.com/FenTechSolutions/CausalDiscoveryToolbox/blob/master/examples/example_cgnn.ipynb

Describe the bug But whenever, I run the CGNN codes, I obtain this error, which indicates that the 'networkx' library has no attribute 'adj_matrix'.

I think it is maybe because the function 'adj_matrix' was deprecated in the 'networkx' library.

Below is the versions of my python and related libraries.

I am running my codes on GPUs.

The problem is that the adj_matrix() is no more valid in new version of networkx, the new function is adjacency_matrix().

You can go to the cdt.causality.graph.CGNN, and change the function hill_climbing() as following:

def hill_climbing(data, graph, **kwargs):
    """Hill Climbing optimization: a greedy exploration algorithm."""
    if isinstance(data, th.utils.data.Dataset):
        nodelist = data.get_names()
    elif isinstance(data, pd.DataFrame):
        nodelist = list(data.columns)
    else:
        raise TypeError('Data type not understood')
    # tested_candidates = [nx.adj_matrix(graph, nodelist=nodelist, weight=None)]
    tested_candidates = [nx.adjacency_matrix(graph, nodelist=nodelist, weight=None)]
    best_score = parallel_graph_evaluation(data,
                                           tested_candidates[0].todense(),
                                           ** kwargs)
    best_candidate = graph
    can_improve = True
    while can_improve:
        can_improve = False
        for (i, j) in best_candidate.edges():
            test_graph = deepcopy(best_candidate)
            test_graph.add_edge(j, i, weight=test_graph[i][j]['weight'])
            test_graph.remove_edge(i, j)
            # tadjmat = nx.adj_matrix(test_graph, nodelist=nodelist, weight=None)
            tadjmat = nx.adjacency_matrix(test_graph, nodelist=nodelist, weight=None)
            if (nx.is_directed_acyclic_graph(test_graph) and not any([(tadjmat != cand).nnz ==
                                                                      0 for cand in tested_candidates])):
                tested_candidates.append(tadjmat)
                score = parallel_graph_evaluation(data, tadjmat.todense(),
                                                  **kwargs)
                if score < best_score:
                    can_improve = True
                    best_candidate = test_graph
                    best_score = score
                    break
    return best_candidate

Then it should be work.