Typos import DAG_GNN, missing packages and ‘pcalg' problem

[Mathugo]

Hi Thibault, I noticed while I was trying to run your code: train_eval.ipynb that:

• from DCGL.model import DAG_GNN in train_eval.ipynb should be instead from DCGL.model import DAG_GCN, which causes an Import failure

Moreoever,
While executing the code for the Event Sequence and their associated model, in particular lines:

columns, event_table, ground_trues = DataLoader(name= "Microwave24V").data()
if data != "Wireless":
    new_event_table = event_table.loc[:, ["event_type", "time_stamp"]]
else:
    new_event_table = event_table
device = torch.device("cuda:0")

model = ADM4(device, len(columns), graph_threshold=0.05, max_iter=5, em_max_iter=5, rho=0.2, decay=1)
model = model.to(device)
model.fit(new_event_table, columns)
_SHD = model._eval(ground_trues, metric="SHD-C")
_SED = model._eval(ground_trues, metric="CSD")
_SID = model._eval(ground_trues, metric="SID")
_CED = model._eval(ground_trues, metric="CED")
print(f"ADM4, SHD:{_SHD}, SED:{_SED}, SID:{_SID}, CED:{_CED}")
    
model = RPPN(device, len(columns), init_scale=2, embedding_dim=64, hidden_size=64, graph_threshold=0.01, batch_size=64)
model = model.to(device)
model.fit(new_event_table, columns, epochs=20, num_workers=0, save_path="./", split_ratio=0.4   )
_SHD = model._eval(ground_trues, metric="SHD-C")
_SED = model._eval(ground_trues, metric="CSD")
_SID = model._eval(ground_trues, metric="SID")
_CED = model._eval(ground_trues, metric="CED")
print(f"RPPN, SHD:{_SHD}, SED:{_SED}, SID:{_SID}, CED:{_CED}")
    
model = SHP(device, len(columns), decay=3, time_interval=5, seed=42, reg=3, penalty="BIC")
model.fit(new_event_table, columns, threshold=0.9)
_SHD = model._eval(ground_trues, metric="SHD-C")
_SED = model._eval(ground_trues, metric="CSD")
_SID = model._eval(ground_trues, metric="SID")
_CED = model._eval(ground_trues, metric="CED")
print(f"SHP, SHD:{_SHD}, SED:{_SED}, SID:{_SID}, CED:{_CED}")

if data == "Wireless":
    node = [columns.index(event) for event in event_table["event_type"]]
    event = list(event_table["event_type"])
    timestamp = list(event_table["time_stamp"])
    event_dict = {"event": event,
                  "timestamp": timestamp,
                  "node": node}
    topology_matrix = np.mat(np.eye(len(node), dtype=int))
    max_hop = 0
else:
    node = list(event_table["node"])
    event = list(event_table["event_type"])
    timestamp = list(event_table["time_stamp"])
    event_dict = {"event": event,
                  "timestamp": timestamp,
                  "node": node}
    topology_matrix = np.load(f"./DCGL/data/dataset/{data}/Topology.npy")
    max_hop = 2
new_event_table = pd.DataFrame(event_dict)
model = TTPM(topology_matrix=topology_matrix, max_hop=max_hop)
model.learn(new_event_table)
causal_matrix = model.causal_matrix
_SHD = SHD_CPDAG(ground_trues, causal_matrix)
_SED = CSD(ground_trues, causal_matrix)
_SID = SID(ground_trues, causal_matrix)
_CED = CED(ground_trues, causal_matrix, ground_trues.shape[0])
print(f"THP, SHD:{_SHD}, SED:{_SED}, SID:{_SID}, CED:{_CED}")

of train_eval.ipynb, I'm stumbling upon an R package error:

 model._eval(ground_trues, metric="SHD-C")
     17 _SED = model._eval(ground_trues, metric="CSD")
     18 _SID = model._eval(ground_trues, metric="SID")

File ~/SageMaker/DOSEC/notebooks/research_pretraining/OCDB/model/BaseModel.py:33, in BaseModel._eval(self, ground_trues, metric)
     31     return SID(ground_trues, self.DAG)
     32 elif metric == "SHD-C":
---> 33     return SHD_CPDAG(ground_trues, self.DAG)
     34 elif metric == "CED":
     35     return CED(ground_trues, self.DAG, k=ground_trues.shape[0])

File ~/SageMaker/DOSEC/notebooks/research_pretraining/OCDB/metrics/Metrics.py:116, in SHD_CPDAG(G_true, G_pred)
    115 def SHD_CPDAG(G_true, G_pred):
--> 116     return cm.SHD_CPDAG(G_true, G_pred)

File ~/anaconda3/envs/python3/lib/python3.10/site-packages/cdt/metrics.py:185, in SHD_CPDAG(target, pred)
    181 true_labels = retrieve_adjacency_matrix(target)
    182 predictions = retrieve_adjacency_matrix(pred, target.nodes()
    183                                         if isinstance(target, nx.DiGraph) else None)
--> 185 true_labels = get_CPDAG(true_labels)
    186 predictions = get_CPDAG(predictions)
    188 return SHD(true_labels, predictions, False)

File ~/anaconda3/envs/python3/lib/python3.10/site-packages/cdt/metrics.py:131, in get_CPDAG(dag)
    116 R"""Compute the completed partially directed acyclic graph (CPDAG) of
    117 a given DAG
    118 
   (...)
    128 
    129 """
    130 if not RPackages.pcalg:
--> 131     raise ImportError("pcalg R package is not available. Please check your installation.")
    133 dag = retrieve_adjacency_matrix(dag)
    135 base_dir = Path('{0!s}/cdt_CPDAG_{1!s}'.format(gettempdir(), uuid.uuid4()))

ImportError: pcalg R package is not available. Please check your installation.

Which triggered a question ro me as why does this repo need a package that was added to R in early september 2024 (pcalg) whereas the paper came out on arxiv in June ? Do we really need this package in order to run the script ? If we need R, could you detail your overall installation and R version ?

Also wget and torch_geometric seems to be mandatory dependecies (I got errors or missing packages)

_My version:
Python 3.10.14
gcastle==1.0.3
cdt==0.6.0
wget==3.2
torch==2.4.1
torch-geometric==2.6.0

Merci pour le repo,
Best,
Hugo