Two hop node info

[shrimonmuke0202]

How to add three-hop away information in a edge_index using the class TwoHop ?

[rusty1s]

You should be able to do this by stacking the TwoHop transform twice:

transform = T.Compose([T.TwoHop(), T.TwoHop()])

[wsad1]

Correct me if I am wrong , but looking at the TwoHop() code it looks like it computes A^2 (square of adjacency matrix). So wouldn't computing two hop twice give return 4 hop neighbors.

In the below example the node 4 is 4-hops hops away from node 0.

data = Data(edge_index = torch.tensor([[0,1,2,3],[1,2,3,4]]))
from torch_geometric.transforms import TwoHop
th = TwoHop()
print(th(th(data)).edge_index) # this returns 4 hop neighbors

tensor([[0, 0, 0, 0, 1, 1, 1, 2, 2, 3],
[1, 2, 3, 4, 2, 3, 4, 3, 4, 4]])

[rusty1s]

Yes, you are right. My bad. So it seems that currently one needs to call spspmm in TwoHop twice to achieve the desired effect. A custom transform should do the job. Alternatively, we can think about a KHop transform.

[shrimonmuke0202]

When I applied coellased coalesced=False
AssertionError Traceback (most recent call last)
in ()
90 for epoch in range(NUM_EPOCHS):
91 hidden,cell = model.init_hidden(batch_size=TRAIN_BATCH_SIZE)
---> 92 train(model, device, train_loader, optimizer, epoch+1,hidden,cell)
93 G,P = predicting(model, device, test_loader,hidden,cell)
94 ret = [rmse(G,P),mse(G,P),pearson(G,P),spearman(G,P)]

8 frames
/usr/local/lib/python3.7/dist-packages/torch_sparse/storage.py in init(self, row, rowptr, col, value, sparse_sizes, rowcount, colptr, colcount, csr2csc, csc2csr, is_sorted)
68 assert row.max().item() < sparse_sizes[0]
69 if col.numel() > 0:
---> 70 assert col.max().item() < sparse_sizes[1]
71
72 if row is not None:

AssertionError:

This type of error ocurrus . How can I recover from this error?

[rusty1s]

Can you give some more information about when this error happens? Do you have an example to reproduce?

[shrimonmuke0202]

AssertionError Traceback (most recent call last)
in ()
90 for epoch in range(NUM_EPOCHS):
91 hidden,cell = model.init_hidden(batch_size=TRAIN_BATCH_SIZE)
---> 92 train(model, device, train_loader, optimizer, epoch+1,hidden,cell)
93 G,P = predicting(model, device, test_loader,hidden,cell)
94 ret = [rmse(G,P),mse(G,P),pearson(G,P),spearman(G,P)]

9 frames
in train(model, device, train_loader, optimizer, epoch, hidden, cell)
18 data = data.to(device)
19 optimizer.zero_grad()
---> 20 output = model(data,hidden,cell)
21 loss = loss_fn(output, data.y.view(-1, 1).float().to(device))
22 loss.backward()

/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1049 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1050 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051 return forward_call(*input, **kwargs)
1052 # Do not call functions when jit is used
1053 full_backward_hooks, non_full_backward_hooks = [], []

in forward(self, data, hidden, cell)
72 #edge_index = torch.cat([data.edge_index,index],dim=1)
73 #edge_index_square,_ = coalesce(edge_index, None, data.num_nodes, data.num_nodes)
---> 74 squared_data = self.twohop(data)
75 h4 = self.Conv1(x,squared_data.edge_index)
76 h4 = self.relu(h4)

/usr/local/lib/python3.7/dist-packages/torch_geometric/transforms/two_hop.py in call(self, data)
14 value = edge_index.new_ones((edge_index.size(1), ), dtype=torch.float)
15
---> 16 index, value = spspmm(edge_index, value, edge_index, value, N, N, N)
17 value.fill_(0)
18 index, value = remove_self_loops(index, value)

/usr/local/lib/python3.7/dist-packages/torch_sparse/spspmm.py in spspmm(indexA, valueA, indexB, valueB, m, k, n, coalesced)
28 sparse_sizes=(k, n), is_sorted=not coalesced)
29
---> 30 C = matmul(A, B)
31 row, col, value = C.coo()
32

/usr/local/lib/python3.7/dist-packages/torch_sparse/matmul.py in matmul(src, other, reduce)
137 return spmm(src, other, reduce)
138 elif isinstance(other, SparseTensor):
--> 139 return spspmm(src, other, reduce)
140 raise ValueError
141

/usr/local/lib/python3.7/dist-packages/torch_sparse/matmul.py in spspmm(src, other, reduce)
114 reduce: str = "sum") -> SparseTensor:
115 if reduce == 'sum' or reduce == 'add':
--> 116 return spspmm_sum(src, other)
117 elif reduce == 'mean' or reduce == 'min' or reduce == 'max':
118 raise NotImplementedError

/usr/local/lib/python3.7/dist-packages/torch_sparse/matmul.py in spspmm_sum(src, other)
104 valueC = valueC.to(value.dtype)
105 return SparseTensor(row=None, rowptr=rowptrC, col=colC, value=valueC,
--> 106 sparse_sizes=(M, K), is_sorted=True)
107
108

/usr/local/lib/python3.7/dist-packages/torch_sparse/tensor.py in init(self, row, rowptr, col, value, sparse_sizes, is_sorted)
23 rowcount=None, colptr=None, colcount=None,
24 csr2csc=None, csc2csr=None,
---> 25 is_sorted=is_sorted)
26
27 @classmethod

/usr/local/lib/python3.7/dist-packages/torch_sparse/storage.py in init(self, row, rowptr, col, value, sparse_sizes, rowcount, colptr, colcount, csr2csc, csc2csr, is_sorted)
68 assert row.max().item() < sparse_sizes[0]
69 if col.numel() > 0:
---> 70 assert col.max().item() < sparse_sizes[1]
71
72 if row is not None:

AssertionError:

when Train my model in a batch this type of error occurs but for batch=1 this type of error does not show to me.

[rusty1s]

This seems to be a problem with spspmm on CUDA. Can you give me some information about your CUDA version?

[shrimonmuke0202]

cuda version 11.2 nvidia tesla v100 gpu

[shrimonmuke0202]

I WANT three hop away information from a particular node.

…

On Mon, 20 Sep 2021, 2:41 pm Jinu Sunil, ***@***.***> wrote: Correct me if I am wrong here, but looking at the TwoHop() code it looks like it computes A^2 (square of adjacency matrix). So wouldn't computing two hop twice give you the 4 hop neighborhood. In the below example the node 4 is 4-hops hops away from node 0. data = Data(edge_index = torch.tensor([[0,1,2,3],[1,2,3,4]])) from torch_geometric.transforms import TwoHop th = TwoHop() print(th(th(data)).edge_index) tensor([[0, 0, 0, 0, 1, 1, 1, 2, 2, 3], [1, 2, 3, 4, 2, 3, 4, 3, 4, 4]]) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#3172 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQ7MTCROMYXMHPSUTMYASHLUC33FPANCNFSM5EIFCXKA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[rusty1s]

See my answer above.

[shrimonmuke0202]

How to multiply a sparse adjacency matrix with a dense adjacency matrix.

[rusty1s]

Either:

from torch_scatter import scatter

row, col = edge_index
out = scatter(x[col], row, dim=0, reduce='add')

or

from torch_sparse import SparseTensor

adj = SparseTensor(row=edge_index[0], col=edge_index[1])
out = adj @ x