Source of model parameter

[jinyilun718]

Hi, thanks for your awesome work. However, I met some problems when replicating your work. If I did not use the model parameters you provided, the results could not be aligned with the paper. How are these model parameters file trained? I am looking forward to your reply.

[jinyilun718]

In ogbn_arxiv we obtain incorrect results, if we comment out the code on line 37:

# load pre-trained model parameter

model.load_statedict(torch.load("./models/DinkNet{}.pt".format(args.dataset)))

results: epoch 009 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 019 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 029 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 039 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 049 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 059 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 069 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 079 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 089 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 epoch 099 ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73 100%|█████████████████████████████████████████| 100/100 [01:23<00:00, 1.19it/s] test ｜ acc:16.14 ｜ nmi:0.00 ｜ ari:0.00 ｜ f1:0.73

Do you know what the cause is and how to solve it, thanks for your help.

[yueliu1999]

Hello, thanks for your attention. Our proposed method contains pre-training and fine-tuning. During the pre-training process, the network and cluster center embeddings are pre-trained. If you comment out the model loading code and do not use the pre-training, the model easily leads to the collapse problem. The pre-training process aims to obtain promising node and cluster center embeddings, which alleviates the collapse problem during the fine-tuning process. Thanks again. Feel free to contact me on WeChat: ly13081857311. :)