search

sisaman / ProGAP

ProGAP: Progressive Graph Neural Networks with Differential Privacy Guarantees (WSDM 2024)
Creative Commons Zero v1.0 Universal
5 stars 2 forks source link

ProGAP Code Execution Errors #2

Open Zening-Li opened 1 month ago

Zening-Li commented 1 month ago

Hi Sisaman: So happy to see your code about ProGAP. I'm attempting to execute the code, but I encounter errors. Specifically, when the model is set to GNN-DP and the graph convolution layer is either GCN or GAT, an error arises during runtime: ValueError: Attempted to use an uninitialized parameter in <built-in method empty_like of type object at 0x7f9ba553a140>. This error happens when you are using a LazyModule or explicitly manipulating torch.nn.parameter.UninitializedParameter objects. When using LazyModules Call forward with a dummy batch to initialize the parameters before calling torch functions. Can you tell me how to fix the bug? Thanks.

sisaman commented 1 month ago

Can you give me the exact CLI command you use to reproduce the error?

Zening-Li commented 1 month ago

Thank you for your response. The exact CLI command I used is: python train.py gnn node --dataset facebook --hidden_dim 16 --activation relu --optimizer adam --learning_rate 0.01 --repeats 1 --epochs 100 --batch_size 64 --verbose False --base_layers 1 --head_layers 1 --jk cat --epsilon 8.0 --project GNN.

To utilize the GCN architecture as the backbone GNN model, I modified core/methods/gnn/node.py at line 36 to conv='gcn'.