Custom Dataset One Hot Encoding

[Sfonzie98]

Hello everyone!
I apologize for the banality of the question, but I am a rookie in the world of geometric deep learning and I am learning to use PyG. I would like to run the Weisfeiler-Lehman kernel on a dataset that I generated, where the data looks like:
Data(x=[9, 9], edge_index=[2, 18], edge_attr=[18, 2], y=[1])
And the node feature matrix (x) is a torch FloatTensor who looks like:

`tensor([[6., 1., 0., 4., 0., 3., 0., 0., 0.],

[8., 2., 0., 3., 0., 0., 0., 0., 0.],

[6., 3., 0., 3., 0., 0., 0., 0., 0.],

[8., 1., 0., 3., 0., 0., 0., 0., 0.],

[6., 2., 0., 3., 0., 1., 0., 0., 0.],

[6., 2., 0., 3., 0., 1., 0., 0., 0.],

[7., 3., 0., 3., 0., 0., 0., 1., 0.],

[6., 2., 0., 4., 0., 2., 0., 1., 0.],

[6., 2., 0., 4., 0., 2., 0., 1., 0.]])`

Now my question is: is there any way to transform this into a matrix consisting of just 1's and 0's, like the x-matrix of the ENZYMES dataset used in example (https://github.com/pyg-team/pytorch_geometric/blob/master/examples/wl_kernel.py)? Without carrying out this transformation the code returns the AssertionError (x.sum(dim=-1) == 1).sum() == x.size(0). Thanks you all!

[rusty1s]

As you correctly pointed out, you will need unique one-hot encoded features to apply WLConv. For this, you would need to convert each row into a unique integer, for example via (not tested):

offset = 1
for i in range(x.size(-1)):
    x[:, i] *= offset
    offset = int(x[:, i].max()) + 1
x = x.sum(dim=-1)
_, x = torch.unique(x, return_inverse=True)