IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an email to s dot adaszewski at gmail dot com. User accounts are meant only to report issues and/or generate pull requests. This is a purpose-specific Git hosting for ADARED projects. Thank you for your understanding!
Sfoglia il codice sorgente

Add support for multiple input_dim in SparseMultiDGCA.

master
Stanislaw Adaszewski 4 anni fa
parent
commit
e371a83352
2 ha cambiato i file con 24 aggiunte e 8 eliminazioni
  1. +18
    -6
      src/decagon_pytorch/convolve.py
  2. +6
    -2
      tests/decagon_pytorch/test_convolve.py

+ 18
- 6
src/decagon_pytorch/convolve.py Vedi File

@@ -212,13 +212,25 @@ class SparseMultiDGCA(torch.nn.Module):
activation: Callable[[torch.Tensor], torch.Tensor]=torch.nn.functional.relu,
**kwargs) -> None:
super().__init__(**kwargs)
self.input_dim = input_dim
self.output_dim = output_dim
self.sparse_dgca = [ SparseDropoutGraphConvActivation(input_dim, output_dim, adj_mat, keep_prob, activation) for adj_mat in adjacency_matrices ]

def forward(self, x: List[torch.Tensor]) -> List[torch.Tensor]:
out = torch.zeros(len(x), self.output_dim, dtype=x.dtype)
for f in self.sparse_dgca:
out += f(x)
self.adjacency_matrices = adjacency_matrices
self.keep_prob = keep_prob
self.activation = activation
self.sparse_dgca = None
self.build()

def build(self):
if len(self.input_dim) != len(self.adjacency_matrices):
raise ValueError('input_dim must have the same length as adjacency_matrices')
self.sparse_dgca = []
for input_dim, adj_mat in zip(self.input_dim, self.adjacency_matrices):
self.sparse_dgca.append(SparseDropoutGraphConvActivation(input_dim, self.output_dim, adj_mat, self.keep_prob, self.activation))

def forward(self, x: List[torch.Tensor]) -> torch.Tensor:
out = torch.zeros(len(x[0]), self.output_dim, dtype=x[0].dtype)
for i, f in enumerate(self.sparse_dgca):
out += f(x[i])
out = torch.nn.functional.normalize(out, p=2, dim=1)
return out



+ 6
- 2
tests/decagon_pytorch/test_convolve.py Vedi File

@@ -236,16 +236,20 @@ def test_multi_dgca():
assert np.all(adjacency_matrices[i].numpy() == adjacency_matrices_sparse[i].to_dense().numpy())
torch.random.manual_seed(0)
multi_sparse = decagon_pytorch.convolve.SparseMultiDGCA(10, 10, adjacency_matrices_sparse, keep_prob=keep_prob)
multi_sparse = decagon_pytorch.convolve.SparseMultiDGCA([10,]*len(adjacency_matrices), 10, adjacency_matrices_sparse, keep_prob=keep_prob)
torch.random.manual_seed(0)
multi = decagon_pytorch.convolve.MultiDGCA(10, 10, adjacency_matrices, keep_prob=keep_prob)
print('len(adjacency_matrices):', len(adjacency_matrices))
print('len(multi_sparse.sparse_dgca):', len(multi_sparse.sparse_dgca))
print('len(multi.dgca):', len(multi.dgca))
for i in range(len(adjacency_matrices)):
assert np.all(multi_sparse.sparse_dgca[i].sparse_graph_conv.weight.detach().numpy() == multi.dgca[i].graph_conv.weight.detach().numpy())
# torch.random.manual_seed(0)
latent_sparse = multi_sparse(latent_sparse)
latent_sparse = multi_sparse([latent_sparse,] * len(adjacency_matrices))
# torch.random.manual_seed(0)
latent = multi(latent)


Loading…
Annulla
Salva