Process Killed with Minimal Settings on GPU (inj_flickr)

[gbay7]

Description:
I am experiencing an issue where the process is unexpectedly killed while benchmarking several graph anomaly detection methods from PyGOD: "dominant," "done," "adone," "anomalydae," and "conad," on the dataset "inj_flickr."

The problem persists even with minimal settings that are much smaller than those used in related works (e.g., the BOND paper, page 20). Despite running on a system with a 12GB NVIDIA A2000 GPU, the process fails to complete.

Settings:
Below are the configurations I used:

• num_layers: 1
• batch_size: 1
• num_neigh: 1
• epoch: 1
• hid_dim: [2]

These are all used settings for "dominant":

{'hid_dim': 2, 'weight_decay': 0.01, 'dropout': 0.0, 'lr': 0.01, 'epoch': 1, 'gpu': '0', 'weight': 0.2, 'batch_size': 1, 'num_neigh': 1, 'verbose': 0}

System Information:

• OS: Linux
• Python version: 3.12.4
• CUDA version: 12
• PyGOD version: 1.1.0
• PyTorch version: 2.4.0+cu124
• GPU: NVIDIA RTX A2000 (12GB)

Additional Context:
GCNAE runs successfully with BOND settings on the same dataset and system.