Pytorch_geometric: Fixed, static graph with varying Node samples

[emanuelazcona]

I'm trying to get started on using this package for an experiment I'm running and I'd like to use the ChebConv class.

My problem uses a fixed, static graph of shape:

A: _N_ x _N_ (Nodes x Nodes)

and an input of shape:

X: _S_ x _N_ x _F_ (Samples x Nodes x Features)

where what "changes" here is different instances/samples of the same graph (same number of Nodes & Features) for each sample.

How can I use the Data functions from utils library here?

I don't want to use a Data object for each Sample because every sample has the same edges and edge features, where the number of nodes is ~= 600k. If I use a Data object for every sample I'd be replicating the same graph _S_-times (one for each sample of _X_)

I just want one object that has my adjacency matrix (2D) and another for my input tensor (3D) and batch over the _Samples_ dimension of the input tensor

[emanuelazcona]

I'm not sure if I'm answering my own question here, but:

Should I just create a single Data object of dimensions: _Nodes_ x _Samples_ x _Features_ ?

How would I batch over a single Data object where my batch dimension is in the middle though..I'm very confused

[emanuelazcona]

close was an accident

[rusty1s]

This is now supported with the latest commit via the node_dim attribute of MessagePassing [1, 2]. If you have batched node features x with shape [B, N, F] and a sparse single adjacency matrix edge_index with shape [2, E], all you need to do is:

conv = GCNConv(in_channels, out_channels, node_dim=1)
out = conv(x, edge_index)

[1] https://github.com/rusty1s/pytorch_geometric/blob/master/torch_geometric/nn/conv/message_passing.py#L17-L40
[2] https://github.com/rusty1s/pytorch_geometric/blob/master/test/nn/conv/test_static_graph.py

[rusty1s]

All you need to test is to pull PyG from GitHub and run python setup.py install. If you encounter any issues with this, please let me know.

[emanuelazcona]

This is now supported with the latest commit via the node_dim attribute of MessagePassing [1, 2]. If you have batched node features x with shape [B, N, F] and a sparse single adjacency matrix edge_index with shape [2, E], all you need to do is:

conv = GCNConv(in_channels, out_channels, node_dim=1)
out = conv(x, edge_index)

[1] https://github.com/rusty1s/pytorch_geometric/blob/master/torch_geometric/nn/conv/message_passing.py#L17-L40
[2] https://github.com/rusty1s/pytorch_geometric/blob/master/test/nn/conv/test_static_graph.py

This didn't fix the issue. As it stands, the code works regardless of whether

node_dim = 0 or node_dim=1 .

By using the Batch operation, my data, x, of dimensions B x N x F, is now just converted to B*N x F

[emanuelazcona]

[emanuelazcona]

`---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
in
13 out = conv_layer(x = batch2.x,
14 edge_index = A_test.edge_index,
---> 15 edge_weight = A_test.edge_attr)

/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in __call__(self, input, *kwargs)
539 result = self._slow_forward(input, *kwargs)
540 else:
--> 541 result = self.forward(input, *kwargs)
542 for hook in self._forward_hooks.values():
543 hook_result = hook(self, input, result)

/usr/local/lib/python3.6/dist-packages/torch_geometric/nn/conv/cheb_conv.py in forward(self, x, edge_index, edge_weight, batch, lambda_max)
113 edge_index, norm = self.norm(edge_index, x.size(0), edge_weight,
114 self.normalization, lambda_max,
--> 115 dtype=x.dtype, batch=batch)
116
117 Tx_0 = x

/usr/local/lib/python3.6/dist-packages/torch_geometric/nn/conv/cheb_conv.py in norm(edge_index, num_nodes, edge_weight, normalization, lambda_max, dtype, batch)
90
91 edge_index, edge_weight = get_laplacian(
---> 92 edge_index, edge_weight, normalization, dtype, num_nodes)
93
94 if batch is not None and torch.is_tensor(lambda_max):

/usr/local/lib/python3.6/dist-packages/torch_geometric/utils/get_laplacian.py in get_laplacian(edge_index, edge_weight, normalization, dtype, num_nodes)
42
43 row, col = edge_index
---> 44 deg = scatter_add(edge_weight, row, dim=0, dim_size=num_nodes)
45
46 if normalization is None:

/usr/local/lib/python3.6/dist-packages/torch_scatter/add.py in scatter_add(src, index, dim, out, dim_size, fill_value)
73 """
74 src, out, index, dim = gen(src, index, dim, out, dim_size, fill_value)
---> 75 return out.scatter_add_(dim, index, src)

RuntimeError: Invalid index in scatterAdd at /pytorch/aten/src/TH/generic/THTensorEvenMoreMath.cpp:721`

[emanuelazcona]

This appears to be the same issue as #611

[rusty1s]

Are you sure you are working on PyTorch Geometric master? What happens if you run the following test code?

import torch
from torch_geometric.data import Data, Batch
from torch_geometric.nn import MessagePassing, GCNConv, ChebConv


class MyConv(MessagePassing):
    def forward(self, x, edge_index):
        return self.propagate(edge_index, x=x)


def test_static_graph():
    edge_index = torch.tensor([[0, 1, 1, 2], [1, 0, 2, 1]])
    x1, x2 = torch.randn(3, 8), torch.randn(3, 8)

    data1 = Data(edge_index=edge_index, x=x1)
    data2 = Data(edge_index=edge_index, x=x2)
    batch = Batch.from_data_list([data1, data2])

    x = torch.stack([x1, x2], dim=0)
    for conv in [MyConv(), GCNConv(8, 16), ChebConv(8, 16, K=2)]:
        out1 = conv(batch.x, batch.edge_index)
        assert out1.size(0) == 6
        conv.node_dim = 1
        out2 = conv(x, edge_index)
        assert out2.size()[:2] == (2, 3)
        assert torch.allclose(out1, out2.view(-1, out2.size(-1)))

test_static_graph()

[emanuelazcona]

**On another note, I'm now getting an error after reinstalling PyTorch Geometric (to ensure it's from master even though I'm sure it is).

I'll open this as a new issue and keep the batch-question issue here.

Once this is remedied I'll re-test the batch test you suggested.

**

see #843