multi-graph multi-adjacency matrix

[DuanhaoranCC]

import torch
from torch_geometric.datasets import PPI
from torch_geometric.loader import DataLoader
from torch_geometric.nn import GraphSAGE

path = "./"
train_dataset = PPI(path, split='train')
val_dataset = PPI(path, split='val')
test_dataset = PPI(path, split='test')
train_data = Batch.from_data_list(train_dataset)
train_loader = DataLoader(train_dataset, batch_size=3, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=2, shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=2, shuffle=False)

 for data in train_loader:
      data = data.to(device)
      
      edge_index = data.edge_index[:, torch.nonzero(data.batch == 0).squeeze()]
      p1 = to_dense_adj(edge_index).cpu().detach().numpy()[0]
      edge_index = data.edge_index[:, torch.nonzero(data.batch == 1).squeeze()]
      p2 = to_dense_adj(edge_index).cpu().detach().numpy()[0]
      edge_index = data.edge_index[:, torch.nonzero(data.batch == 2).squeeze()]
      p3 = to_dense_adj(edge_index).cpu().detach().numpy()[0]
      adj = torch.block_diag(p1, p2, p3)

I currently have a requirement to get the adjacency matrix of each graph, and then form it into a large matrix, but I found that the dimensions are not uniform, where is the error?(7452 not = 6860)

data :  DataBatch(x=[6860, 50], edge_index=[2, 191870], y=[6860, 121], batch=[6860], ptr=[4])
adj:    (7452,7452)

[rusty1s]

I think the issue is that you are not decreasing the values of the edge_index going into p2 and p3 (such that they are starting from zero):

edge_index = data.edge_index[:, data.batch[data.edge_index[0]] == 0] - data.ptr[1]
edge_index = data.edge_index[:, data.batch[data.edge_index[0]] == 1] - data.ptr[2]

Let me know if this fixes your issues :)

[DuanhaoranCC]

I found out that it doesn't work, when I changed the batch to 1, I printed the following output"

path = "./"
train_dataset = PPI(path, split='train')
val_dataset = PPI(path, split='val')
test_dataset = PPI(path, split='test')
train_loader = DataLoader(train_dataset, batch_size=1, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=2, shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=2, shuffle=False)

 for data in train_loader:
      data = data.to(device)
      print(data)
      #DataBatch(x=[1767, 50], edge_index=[2, 32318], y=[1767, 121], batch=[1767], ptr=[2])
      edge_index = data.edge_index[:, torch.nonzero(data.batch == 0).squeeze()]  
      print(edge_index)
      #(2,1767)tensor([[   0,    0,    0,  ...,   88,   88,   88],
      #                         [ 372, 1101,  766,  ...,  495,  646, 1193]], device='cuda:0')
      print(edge_index.max())
      #tensor(1551, device='cuda:0')

The dimension of data.x corresponds to the dimension of the edge, but it seems wrong, because each node has many repeated edges, the dimension of the edge must be larger than that of data.x, which is intuitive.

[rusty1s]

Can you clarify what you mean? Each node has a number of edges (that's why x.size(0) < edge_index.size(1)), but there are no repeat edges since each pair in edge_index is different.

I noticed that there is a mistake in my previous script. Here is the corrected version:

from torch_geometric.datasets import PPI
from torch_geometric.loader import DataLoader

path = "./"
train_dataset = PPI(path, split='train')
train_loader = DataLoader(train_dataset, batch_size=2, shuffle=True)

for data in train_loader:
    print(data)

    row = data.edge_index[0]

    edge_index = data.edge_index[:, data.batch[row] == 0] - data.ptr[0]
    print(edge_index.min())
    print(edge_index.max())

    edge_index = data.edge_index[:, data.batch[row] == 1] - data.ptr[1]
    print(edge_index.min())
    print(edge_index.max())

[DuanhaoranCC]

I apologize for my unclear statement here, thank you for your patience and guidance, the problem has been solved, thank you!