Search code examples
deep-learningpytorchpytorch-geometric

spectral_norm on GCNConv module

I want to call torch.nn.utils spectral_norm function on a GCNConv layer

gc1 = GCNConv(18, 16)
spectral_norm(gc1)

but I'm getting the following error:

KeyError: 'weight'

meaning gc1._parameters doesn't have weight (only bias):

gc1._parameters
OrderedDict([('bias', Parameter containing:
              tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
                     requires_grad=True))])

However, gc1.parameters() stores two objects and one of them is a 16 by 18 matrix (weight matrix).

for p in gc1.parameters():
  print('P: ', p.shape)
P:  torch.Size([16])
P:  torch.Size([16, 18])

How can I make spectral_norm function work on a GCNConv module?

Solution

  • According to the source code, the weight parameter is wrapped within a linear module contained in GCNConv objects as lin.

    I imagine that this should then work:

    gc1 = GCNConv(18, 16)
spectral_norm(gc1.lin)