robinlu1209/st-gfsl

st-gfsl's Issues

Transfer the knowledge between graphs with different number of nodes?

Dear authors,

Firstly, I would like to express my gratitude for sharing the code for your wonderful work, which I find quite inspiring.

After thoroughly reading your paper, I have a question regarding the transfer of non-shared parameters to the target city with a graph of different numbers of nodes. While the paper explains the meta-knowledge and the Parameter Generation process in Section 4.2, I am still unclear about how the source and target nodes are aligned. For instance, how the meta-knowledge $Z^{MK} \in R^{N \times d_{MK}}$ can be transferred to the target city with a graph of $N'$ nodes?

I would be grateful if you could provide clarification on this matter.

Thank you in advance for your time and assistance.

Best regards,
Jingwei

Which feature extractor do you use to get the result in table2?

It seems that the feature extractor is not claimed for the results in table2. Should it be graph wavenet (GWN)?
I would be grateful if you could provide clarification on this problem.

Why is it necessary to load "shenzhen" state dict for GWN?

Dear authors,

I followed your code and run it with graph-wavenet. In the file maml.py, I saw

        if self.model_name == 'GWN':
            maml_model = MetaGWN(self.model_args, self.task_args)
            maml_model.load_state_dict(torch.load('shenzhen_gwn_model.pkl'))
            print("model load successfully.")
            maml_model = maml_model.cuda()

I wonder why is it necessary to load the state dict, especially for GWN model?

关于maml.py中第149行

作者您好，有些元学习训练时的疑惑，采用deepcopy不会导致优化器没办法优化嘛

Why the result for both MAML(M2) and ST-GFSL different between Table2 and Table3?

Most of the result are the same between MAML in Table2 and M2 in Table3, is it an accidentally made mistake for the rest differences?
What is more, do you choose a different random seed for ST-GFSL in Table2 and Table3?
I would be grateful if you could provide clarification on this problem.

Why add target dataset when construct the dataset?

Hello author, thank you for your excellent work. I've been studying your work recently and I've run into some problems that I hope you can help me with. As I said in the title, why in the Meta-Train stage, when constructing the dataset, do you add the target data?

After doing this, when Meta-train reads the data, the target data will be randomly obtained, as follows:

I think it leads to the leakage of information of target data.
Is my point correct? Or maybe I'm missing something and misunderstood your approach. Looking forward to your answer, thank you very much.

What are the meanings represented by the 2 feature dimensions in the dataset

Hello author, in your public dataset, the feature dimension of each node in each dataset is 2. It can be seen that the 0th dimension represents speed, so what data does the 1st dimension represent? Its range of values is [0,1.] .
When predicting, only the speed was predicted, without considering dimension 1.