I run this code. My f1-micro of PPI is 0.491,while in paper it is 0.976. And during the process, the following appears many times: Iterations not converging! tensor(32., device='cuda:0') max itrs . Waiting for you answer!

Hi, thank you for your code, I've learned a lot from it. Now I met a problem("ValueError: row, column, and data arrays must be 1-D") when I'm running train_IGNN.py in your graphclassification folder. The problem occurs because data.edge_attr is not None and it's a N*4 tensor, so adj_sp cannot be constructed correctly. Could you please tell me how to fix this?

Thank you and look forward to your reply:)

Dear authors,

Thanks for your impressive work. I have a few questions about the synthetic datasets.

1. In Figure 1, the x-axis starts with 0. Does it mean that the length is 0? How can we create chains without nodes? I think it would be totally a none graph?
2. You mentioned in the paper that 20 nodes for training. But the number of nodes is not fixed for datasets with chains having different lengths, right? So, the setting 20/100/200 for train/val/test is only for the dataset with 20 chains for each class and 10-length chains? Because the number of nodes is decided by c*n*l as you stated in the code.

Thanks in advance.

Hi there @SwiftieH , thank you for your great work on implicit GNN.

Since I am running the train_IGNN_PPI.py for training IGNN on PPI dataset, I found it's hard to get the micro F1-score, which is around 95% as reported in the paper, instead I get test set F1-score only around 50% after 1000 epochs, the training loss could be larger (over 10,000,000) after 250+ epochs, I think there is something wrong with the training. The experiment setting is all default (lr=0.01, kappa=0.98, no weight decay), the forward and backward iteration in implicit function is 300 respectively.

I am wondering if there is anything needs to be extra cared, for example increasing the forward/backward iterations to achieve the F1-score results above 95%. Or if I missed some important things in the training, please remind me. Thanks!

Hi there @SwiftieH , thanks for your impressive work!

On looking at the code in nodeclassification/layers.py. I am a little confused about the weight projection. It's certain that during training process, every time the forward() is called, self.W is normalized (||W||_inf < kappa/rho(A))

    def forward(self, X_0, A, U, phi, A_rho=1.0, fw_mitr=500, bw_mitr=500, A_orig=None):
        """Allow one to use a different A matrix for convolution operation in equilibrium equ"""
        if self.k is not None: # when self.k = 0, A_rho is not required
            self.W = projection_norm_inf(self.W, kappa=self.k/A_rho)
        support_1 = torch.spmm(torch.transpose(U, 0, 1), `self.Omega_1.T).T`
        support_1 = torch.spmm(torch.transpose(A, 0, 1), support_1.T).T
        support_2 = torch.spmm(torch.transpose(U, 0, 1), self.Omega_2.T).T
        b_Omega = support_1 #+ support_2
        return ImplicitFunction.apply(self.W, X_0, A if A_orig is None else A_orig, b_Omega, phi, fw_mitr, bw_mitr)

It seems like in the testing process, the weight (learned from training) is normalized again before being passed into the implicit function, is it what supposed to do? Or should we just use the 'original' weight from the training process and not normalize again in the testing process, which can be achieved by using e.g.

if self.training:
  if self.k is not None: # when self.k = 0, A_rho is not required
            self.W = projection_norm_inf(self.W, kappa=self.k/A_rho)

Thanks!