dropout_adj

[DuanhaoranCC]

Hello, I suggest that from torch_geometric.utils import dropout_adj, to_dense_adj support SparseTensor(data.adj_t),it only support data.edge_index now, I want to get the augmentation adjacency matrix. Do you have any suggestions? Thank you very much for your answer! In addition, it is desirable to increase the conversion from adjacency matrix to edge_index to better create Data.
I found that when I process PPI datasets on Linux, the data is:

DataBatch(x=[4814, 50], edge_index=[2, 130252], y=[4814, 121], batch=[4814], ptr=[3])

I found that when I process PPI datasets on windows, the data is:

DataBatch(x=[3865, 50], y=[3865, 121], adj_t=[3865, 3865, nnz=103959], batch=[3865], ptr=[3])

This makes it difficult for me to migrate on different machines. What is the reason?(torch-geometric 2.0.4)

import torch
from torch_geometric.utils import dropout_adj, to_dense_adj, dense_to_sparse
import torch_geometric.transforms as T
from torch_geometric.datasets import PPI
from torch_geometric.loader import DataLoader
from torch_geometric.nn import GCN2Conv

path = "./"
pre_transform = T.Compose([T.GCNNorm()])
train_dataset = PPI(path, split='train')
train_loader = DataLoader(train_dataset, batch_size=2, shuffle=True)
for data in train_loader:
      data = data.to(device)
      edge_index_1 = dropout_adj(data.adj_t, p=0.3)[0]
'''

[Padarn]

The difference between Windows and Linux is quite strange, could you check the source of PPI on the windows machine, and maybe provide a full sample?

[DuanhaoranCC]

import torch

import torch_geometric.transforms as T
from torch_geometric.datasets import PPI
from torch_geometric.loader import DataLoader

path = "./"
pre_transform = T.Compose([T.GCNNorm(), T.ToSparseTensor(remove_edge_index=False)])
train_dataset = PPI(path, split='train', pre_transform=pre_transform)
train_loader = DataLoader(train_dataset, batch_size=2, shuffle=True)
for data in train_loader:
    print(data)

Windows:

DataBatch(x=[4145, 50], y=[4145, 121], adj_t=[4145, 4145, nnz=109775], batch=[4145], ptr=[3])

Linux:

DataBatch(x=[5489, 50], edge_index=[2, 154020], y=[5489, 121], batch=[5489], ptr=[3])

Unable to use data.edge_index.

[Padarn]

Oh, this is likely because the pre-processed data is stored on your machine. You can try removing the data that was downloaded to ./ and loading it again

[Padarn]

It should be printing a warning in this case, something like

UserWarning: The `pre_transform` argument differs from the one used in the pre-processed version of this dataset. If you want to make use of another pre-processing technique, make sure to sure to delete 'test/processed' first
  f"The `pre_transform` argument differs from the one used in "```

[DuanhaoranCC]

Thank you, solved the problem, thank you very much for your patience!

[Padarn]

No worries 👍

As for your original suggestion.. I'm not sure if we want to support this explicitly or not, but maybe yo can raise an issue if you think its still something needed?