def transition_matrix(self, edge_index, edge_weight, num_nodes,
normalization):
r"""Calculate the approximate, sparse diffusion on a given sparse
matrix.
Args:
edge_index (LongTensor): The edge indices.
edge_weight (Tensor): One-dimensional edge weights.
num_nodes (int): Number of nodes.
normalization (str): Normalization scheme. Options:
1. `"sym"`: Symmetric normalization:
:math:`\mathbf{T} = \mathbf{D}^{-1/2} \mathbf{A}
\mathbf{D}^{-1/2}`
2. `"col"`: Column-wise normalization:
:math:`\mathbf{T} = \mathbf{A} \mathbf{D}^{-1}`
3. `"row"`: Row-wise normalization:
:math:`\mathbf{T} = \mathbf{D}^{-1} \mathbf{A}`
4. `None`: No normalization.
:rtype: (:class:`LongTensor`, :class:`Tensor`)
"""
diag_idx = torch.arange(0, num_nodes, dtype=torch.long,
device=edge_index.device)
diag_idx = diag_idx.unsqueeze(0).repeat(2, 1)
if normalization == 'sym':
_, col = edge_index
D_vec = scatter_add(edge_weight, col, dim=0, dim_size=num_nodes)
D_vec_invsqrt = 1 / torch.sqrt(D_vec)
edge_index, edge_weight = spspmm(diag_idx, D_vec_invsqrt,
edge_index, edge_weight,
num_nodes, num_nodes, num_nodes)
edge_index, edge_weight = spspmm(edge_index, edge_weight, diag_idx,
D_vec_invsqrt, num_nodes,
num_nodes, num_nodes)
elif normalization == 'col':
_, col = edge_index
D_vec = scatter_add(edge_weight, col, dim=0, dim_size=num_nodes)
D_vec_inv = 1 / D_vec
edge_index, edge_weight = spspmm(edge_index, edge_weight, diag_idx,
D_vec_inv, num_nodes, num_nodes,
num_nodes)
elif normalization == 'row':
row, _ = edge_index
D_vec = scatter_add(edge_weight, row, dim=0, dim_size=num_nodes)
D_vec_inv = 1 / D_vec
edge_index, edge_weight = spspmm(diag_idx, D_vec_inv, edge_index,
edge_weight, num_nodes, num_nodes,
num_nodes)
elif normalization is None:
pass
else:
raise ValueError(
'Transition matrix normalization {} unknown.'.format(
normalization))
return edge_index, edge_weight
def forward(self, phi_indices, phi_values, phi_inverse_indices,
phi_inverse_values, features):
"""
Forward propagation pass.
:param phi_indices: Sparse wavelet matrix index pairs.
:param phi_values: Sparse wavelet matrix values.
:param phi_inverse_indices: Inverse wavelet matrix index pairs.
:param phi_inverse_values: Inverse wavelet matrix values.
:param features: Feature matrix.
:return localized_features: Filtered feature matrix extracted.
"""
rescaled_phi_indices, rescaled_phi_values = spspmm(
phi_indices, phi_values, self.diagonal_weight_indices,
self.diagonal_weight_filter.view(-1), self.ncount, self.ncount,
self.ncount)
phi_product_indices, phi_product_values = spspmm(
rescaled_phi_indices, rescaled_phi_values, phi_inverse_indices,
phi_inverse_values, self.ncount, self.ncount, self.ncount)
filtered_features = torch.mm(features, self.weight_matrix)
localized_features = spmm(phi_product_indices, phi_product_values,
self.ncount, filtered_features)
return localized_features
def panentropy_sparse(self, edge_index, num_nodes, AFTERDROP,
edge_mask_list):
edge_value = torch.ones(edge_index.size(1), device=edge_index.device)
edge_index, edge_value = coalesce(edge_index, edge_value, num_nodes,
num_nodes)
# iteratively add weighted matrix power
pan_index, pan_value = eye(num_nodes, device=edge_index.device)
indextmp = pan_index.clone().to(edge_index.device)
valuetmp = pan_value.clone().to(edge_index.device)
pan_value = self.panconv_filter_weight[0] * pan_value
for i in range(self.filter_size - 1):
if AFTERDROP:
indextmp, valuetmp = spspmm(indextmp, valuetmp, edge_index,
edge_value * edge_mask_list[i],
num_nodes, num_nodes, num_nodes)
else:
indextmp, valuetmp = spspmm(indextmp, valuetmp, edge_index,
edge_value, num_nodes, num_nodes,
num_nodes)
valuetmp = valuetmp * self.panconv_filter_weight[i + 1]
indextmp, valuetmp = coalesce(indextmp, valuetmp, num_nodes,
num_nodes)
pan_index = torch.cat((pan_index, indextmp), 1)
pan_value = torch.cat((pan_value, valuetmp))
return coalesce(pan_index, pan_value, num_nodes, num_nodes, op='add')
def test_spspmm(dtype, device):
indexA = torch.tensor([[0, 0, 1, 2, 2], [1, 2, 0, 0, 1]], device=device)
valueA = tensor([1, 2, 3, 4, 5], dtype, device)
sizeA = torch.Size([3, 3])
indexB = torch.tensor([[0, 2], [1, 0]], device=device)
valueB = tensor([2, 4], dtype, device)
sizeB = torch.Size([3, 2])
indexC, valueC = spspmm(indexA, valueA, indexB, valueB, 3, 3, 2)
assert indexC.tolist() == [[0, 1, 2], [0, 1, 1]]
assert valueC.tolist() == [8, 6, 8]
A = torch.sparse_coo_tensor(indexA, valueA, sizeA, device=device)
A = A.to_dense().requires_grad_()
B = torch.sparse_coo_tensor(indexB, valueB, sizeB, device=device)
B = B.to_dense().requires_grad_()
torch.matmul(A, B).sum().backward()
valueA = valueA.requires_grad_()
valueB = valueB.requires_grad_()
indexC, valueC = spspmm(indexA, valueA, indexB, valueB, 3, 3, 2)
valueC.sum().backward()
assert valueA.grad.tolist() == A.grad[indexA[0], indexA[1]].tolist()
assert valueB.grad.tolist() == B.grad[indexB[0], indexB[1]].tolist()
def power_adj(adj, dim, p):
val = torch.ones(adj.shape[1])
ic, vc = spspmm(adj, val, adj, val, dim, dim, dim)
if p > 2:
for i in range(p - 2):
ic, vc = spspmm(ic, vc, adj, val, dim, dim, dim)
return ic
def forward(self, phi_indices, phi_values, phi_inverse_indices,
phi_inverse_values, feature_indices, feature_values, dropout):
"""
Forward propagation pass.
:param phi_indices: Sparse wavelet matrix index pairs.
:param phi_values: Sparse wavelet matrix values.
:param phi_inverse_indices: Inverse wavelet matrix index pairs.
:param phi_inverse_values: Inverse wavelet matrix values.
:param feature_indices: Feature matrix index pairs.
:param feature_values: Feature matrix values.
:param dropout: Dropout rate.
:return dropout_features: Filtered feature matrix extracted.
"""
rescaled_phi_indices, rescaled_phi_values = spspmm(
phi_indices, phi_values, self.diagonal_weight_indices,
self.diagonal_weight_filter.view(-1), self.ncount, self.ncount,
self.ncount)
phi_product_indices, phi_product_values = spspmm(
rescaled_phi_indices, rescaled_phi_values, phi_inverse_indices,
phi_inverse_values, self.ncount, self.ncount, self.ncount)
filtered_features = spmm(feature_indices, feature_values, self.ncount,
self.weight_matrix)
localized_features = spmm(phi_product_indices, phi_product_values,
self.ncount, filtered_features)
dropout_features = torch.nn.functional.dropout(
torch.nn.functional.relu(localized_features),
training=self.training,
p=dropout)
return dropout_features
def scale_elements(adj_matrix, adj_part, node_count, row_vtx, col_vtx):
if not normalization:
return adj_part
# Scale each edge (u, v) by 1 / (sqrt(u) * sqrt(v))
# indices = adj_part._indices()
# values = adj_part._values()
# deg_map = dict()
# for i in range(adj_part._nnz()):
# u = indices[0][i] + row_vtx
# v = indices[1][i] + col_vtx
# if u.item() in deg_map:
# degu = deg_map[u.item()]
# else:
# degu = (adj_matrix[0] == u).sum().item()
# deg_map[u.item()] = degu
# if v.item() in deg_map:
# degv = deg_map[v.item()]
# else:
# degv = (adj_matrix[0] == v).sum().item()
# deg_map[v.item()] = degv
# values[i] = values[i] / (math.sqrt(degu) * math.sqrt(degv))
adj_part = adj_part.coalesce()
deg = torch.histc(adj_matrix[0].double(), bins=node_count)
deg = deg.pow(-0.5)
row_len = adj_part.size(0)
col_len = adj_part.size(1)
dleft = torch.sparse_coo_tensor([np.arange(0, row_len).tolist(),
np.arange(0, row_len).tolist()],
deg[row_vtx:(row_vtx + row_len)].float(),
size=(row_len, row_len),
requires_grad=False, device=torch.device("cpu"))
dright = torch.sparse_coo_tensor([np.arange(0, col_len).tolist(),
np.arange(0, col_len).tolist()],
deg[col_vtx:(col_vtx + col_len)].float(),
size=(col_len, col_len),
requires_grad=False, device=torch.device("cpu"))
# adj_part = torch.sparse.mm(torch.sparse.mm(dleft, adj_part), dright)
ad_ind, ad_val = torch_sparse.spspmm(adj_part._indices(), adj_part._values(),
dright._indices(), dright._values(),
adj_part.size(0), adj_part.size(1), dright.size(1))
adj_part_ind, adj_part_val = torch_sparse.spspmm(dleft._indices(), dleft._values(),
ad_ind, ad_val,
dleft.size(0), dleft.size(1), adj_part.size(1))
adj_part = torch.sparse_coo_tensor(adj_part_ind, adj_part_val,
size=(adj_part.size(0), adj_part.size(1)),
requires_grad=False, device=torch.device("cpu"))
return adj_part
def StAS(index_A, value_A, index_S, value_S, device, N, kN):
r"""StAS: a function which returns new edge weights for the pooled graph using the formula S^{T}AS"""
index_A, value_A = coalesce(index_A, value_A, m=N, n=N)
index_S, value_S = coalesce(index_S, value_S, m=N, n=kN)
index_B, value_B = spspmm(index_A, value_A, index_S, value_S, N, N, kN)
index_St, value_St = transpose(index_S, value_S, N, kN)
index_B, value_B = coalesce(index_B, value_B, m=N, n=kN)
index_E, value_E = spspmm(index_St.cpu(), value_St.cpu(), index_B.cpu(), value_B.cpu(), kN, N, kN)
return index_E.to(device), value_E.to(device)
def forward(self, phi_indices, phi_values, phi_inverse_indices,
phi_inverse_values, features):
rescaled_phi_indices, rescaled_phi_values = spspmm(
phi_indices, phi_values, self.diagonal_weight_indices,
self.diagonal_weight_filter.view(-1), self.ncount, self.ncount,
self.ncount)
phi_product_indices, phi_product_values = spspmm(
rescaled_phi_indices, rescaled_phi_values, phi_inverse_indices,
phi_inverse_values, self.ncount, self.ncount, self.ncount)
filtered_features = torch.mm(features, self.weight_matrix)
localized_features = spmm(phi_product_indices, phi_product_values,
self.ncount, filtered_features)
return localized_features
def k_hop_edges(edge_index, num_nodes, K):
# edge_index, _ = remove_self_loops(edge_index, None)
n = num_nodes
edge_index, _ = coalesce(edge_index, None, n, n)
value = edge_index.new_ones((edge_index.size(1), ), dtype=torch.float)
edges = []
new_edge_index = edge_index.clone()
new_edge_value = value.clone()
useful_edge_index, edge_weight = add_self_loops(new_edge_index,
None,
fill_value=-1.,
num_nodes=num_nodes)
edges.append(useful_edge_index)
for i in range(K - 1):
new_edge_index, new_edge_value = spspmm(new_edge_index, new_edge_value,
edge_index, value, n, n, n)
new_edge_index, new_edge_value = coalesce(new_edge_index,
new_edge_value, n, n)
useful_edge_index, edge_weight = add_self_loops(new_edge_index,
None,
fill_value=-1.,
num_nodes=num_nodes)
edges.append(useful_edge_index)
# edge_index, _ = add_self_loops(edge_index, None,fill_value=-1.,num_nodes=num_nodes)
return edges
def forward(self, x, edge_index, edge_attr):
N, dim = x.shape
# x = self.dropout(x)
# adj_mat_ind, adj_mat_val = add_self_loops(edge_index, num_nodes=N)[0], edge_attr.squeeze()
adj_mat_ind = add_remaining_self_loops(edge_index, num_nodes=N)[0]
adj_mat_val = torch.ones(adj_mat_ind.shape[1]).to(x.device)
h = torch.mm(x, self.weight)
h = F.dropout(h, p=self.dropout, training=self.training)
for _ in range(self.nhop - 1):
adj_mat_ind, adj_mat_val = spspmm(adj_mat_ind, adj_mat_val,
adj_mat_ind, adj_mat_val, N, N,
N, True)
adj_mat_ind, adj_mat_val = self.attention(h, adj_mat_ind, adj_mat_val)
# MATRIX_MUL
# laplacian matrix normalization
adj_mat_val = self.normalization(adj_mat_ind, adj_mat_val, N)
val_h = h
# N, dim = val_h.shape
# MATRIX_MUL
# val_h = spmm(adj_mat_ind, F.dropout(adj_mat_val, p=self.node_dropout, training=self.training), N, N, val_h)
val_h = spmm(adj_mat_ind, adj_mat_val, N, N, val_h)
val_h[val_h != val_h] = 0
val_h = val_h + self.bias
val_h = self.adaptive_enc(val_h)
val_h = F.dropout(val_h, p=self.dropout, training=self.training)
# val_h = self.activation(val_h)
return val_h
def augment_adj(self, edge_index, edge_weight, num_nodes):
edge_index, edge_weight = sort_edge_index(edge_index, edge_weight,
num_nodes)
edge_index, edge_weight = spspmm(edge_index, edge_weight, edge_index,
edge_weight, num_nodes, num_nodes,
num_nodes)
return edge_index.to(device)
def forward(self, x, edge_index, edge_attr):
row, col = edge_index
deg = degree(col, x.size(0), dtype=x.dtype)
deg_inv_sqrt = deg.pow(-0.5)
edge_attr_t = deg_inv_sqrt[row] * edge_attr * deg_inv_sqrt[col]
N = x.size(0)
adj = torch.sparse_coo_tensor(edge_index, edge_attr_t, size=(N, N))
theta = self.alpha * (1 - self.alpha)
result = [theta * adj]
for i in range(1, self.steps - 1):
theta = theta * (1 - self.alpha)
adj_ind, adj_val = spspmm(edge_index, edge_attr_t,
result[i - 1]._indices(),
result[i - 1]._values(), N, N, N, True)
result.append(
torch.sparse_coo_tensor(adj_ind, adj_val, size=(N, N)))
identity = torch.sparse_coo_tensor([range(N)] * 2,
torch.ones(N),
size=(N, N)).to(x.device)
result.append(self.alpha * identity)
def fn(x, y):
return x.add(y)
res = reduce(fn, result)
return res._indices(), res._values()
def __call__(self, data):
edge_index, edge_attr = data.edge_index, data.edge_attr
n = data.num_nodes
fill = 1e16
value = edge_index.new_full((edge_index.size(1), ),
fill,
dtype=torch.float)
index, value = spspmm(edge_index, value, edge_index, value, n, n, n,
True)
edge_index = torch.cat([edge_index, index], dim=1)
if edge_attr is None:
data.edge_index, _ = coalesce(edge_index, None, n, n)
else:
value = value.view(-1, *[1 for _ in range(edge_attr.dim() - 1)])
value = value.expand(-1, *list(edge_attr.size())[1:])
edge_attr = torch.cat([edge_attr, value], dim=0)
data.edge_index, edge_attr = coalesce(edge_index,
edge_attr,
n,
n,
op='min',
fill_value=fill)
edge_attr[edge_attr >= fill] = 0
data.edge_attr = edge_attr
return data
def spmmsp(sp1: torch.Tensor, sp2: torch.Tensor) -> torch.Tensor:
assert sp1.size(-1) == sp2.size(0) and sp1.is_sparse and sp2.is_sparse
m = sp1.size(0)
k = sp2.size(0)
n = sp2.size(-1)
indices, values = spspmm(sp1.indices(), sp1.values(), sp2.indices(),
sp2.values(), m, k, n)
return torch.sparse_coo_tensor(indices, values, torch.Size([m, n]))
def augment_adj(self, edge_index, edge_weight, num_nodes):
edge_index, edge_weight = coalesce(edge_index, edge_weight, num_nodes, num_nodes)
edge_index, edge_weight = sort_edge_index(edge_index, edge_weight,
num_nodes)
edge_index, edge_weight = spspmm(edge_index, edge_weight, edge_index,
edge_weight, num_nodes, num_nodes,
num_nodes)
return edge_index, edge_weight
def test_spspmm(dtype, device):
indexA = torch.tensor([[0, 0, 1, 2, 2], [1, 2, 0, 0, 1]], device=device)
valueA = tensor([1, 2, 3, 4, 5], dtype, device)
indexB = torch.tensor([[0, 2], [1, 0]], device=device)
valueB = tensor([2, 4], dtype, device)
indexC, valueC = spspmm(indexA, valueA, indexB, valueB, 3, 3, 2)
assert indexC.tolist() == [[0, 1, 2], [0, 1, 1]]
assert valueC.tolist() == [8, 6, 8]
def augment_adj(self, edge_index, edge_weight, num_nodes):
edge_index, edge_weight = add_self_loops(edge_index, edge_weight,
num_nodes=num_nodes)
edge_index, edge_weight = sort_edge_index(edge_index, edge_weight,
num_nodes)
edge_index, edge_weight = spspmm(edge_index, edge_weight, edge_index,
edge_weight, num_nodes, num_nodes,
num_nodes)
edge_index, edge_weight = remove_self_loops(edge_index, edge_weight)
return edge_index, edge_weight
def forward(self, phi_indices, phi_values, phi_inverse_indices,
phi_inverse_values, feature_indices, feature_values, dropout):
rescaled_phi_indices, rescaled_phi_values = spspmm(
phi_indices, phi_values, self.diagonal_weight_indices,
self.diagonal_weight_filter.view(-1), self.ncount, self.ncount,
self.ncount)
phi_product_indices, phi_product_values = spspmm(
rescaled_phi_indices, rescaled_phi_values, phi_inverse_indices,
phi_inverse_values, self.ncount, self.ncount, self.ncount)
filtered_features = spmm(feature_indices, feature_values, self.ncount,
self.weight_matrix)
localized_features = spmm(phi_product_indices, phi_product_values,
self.ncount, filtered_features)
dropout_features = torch.nn.functional.dropout(
torch.nn.functional.relu(localized_features),
training=self.training,
p=dropout)
return dropout_features
def test_spmm_spspmm(dtype, device):
row = torch.tensor([0, 0, 1, 2, 2], device=device)
col = torch.tensor([0, 2, 1, 0, 1], device=device)
index = torch.stack([row, col], dim=0)
value = tensor([1, 2, 4, 1, 3], dtype, device)
x = tensor([[1, 4], [2, 5], [3, 6]], dtype, device)
value = value.requires_grad_(True)
out_index, out_value = spspmm(index, value, index, value, 3, 3, 3)
out = spmm(out_index, out_value, 3, x)
assert out.size() == (3, 2)
def forward(self, x, edge_index, edge_attr):
device = x.device
N, dim = x.shape
row, col = edge_index
deg = degree(col, x.size(0), dtype=x.dtype)
deg_inv_sqrt = deg.pow(-0.5)
edge_attr_t = deg_inv_sqrt[row] * edge_attr * deg_inv_sqrt[col]
p_val = F.relu(self.p(x))
q_val = F.relu(self.q(x))
# add dropout
p_val = self.dropout(p_val)
q_val = self.dropout(q_val)
# --------------------
diag_ind = torch.LongTensor([range(N)] * 2).to(device)
_, p_adj_mat_val = spspmm(edge_index, edge_attr_t, diag_ind, p_val.view(-1), N, N, N, True)
_, q_adj_mat_val = spspmm(diag_ind, q_val.view(-1), edge_index, edge_attr, N, N, N, True)
return edge_index, p_adj_mat_val + q_adj_mat_val
def nth_deg_adjacency(adj_mat, n=1, sparse=False):
""" Calculates the n-th degree adjacency matrix.
Performs mm of adj_mat and adds the newly added.
Default is dense. Mods for sparse version are done when needed.
Inputs:
* adj_mat: (N, N) adjacency tensor
* n: int. degree of the output adjacency
* sparse: bool. whether to use torch-sparse module
Outputs:
* edge_idxs: ij positions of the adjacency matrix
* edge_attrs: degree of connectivity (1 for neighs, 2 for neighs^2, ... )
"""
adj_mat = adj_mat.float()
attr_mat = torch.zeros_like(adj_mat)
new_adj_mat = adj_mat.clone()
for i in range(n):
if i == 0:
attr_mat += adj_mat
continue
if i == 1 and sparse:
idxs = adj_mat.nonzero().t()
vals = adj_mat[idxs[0], idxs[1]]
new_idxs = idxs.clone()
new_vals = vals.clone()
m, k, n = 3 * [
adj_mat.shape[0]
] # (m, n) * (n, k) , but adj_mats are squared: m=n=k
if sparse:
new_idxs, new_vals = torch_sparse.spspmm(new_idxs,
new_vals,
idxs,
vals,
m=m,
k=k,
n=n)
new_vals = new_vals.bool().float()
new_adj_mat = torch.zeros_like(attr_mat)
new_adj_mat[new_idxs[0], new_idxs[1]] = new_vals
# sparse to dense is slower
# torch.sparse.FloatTensor(idxs, vals).to_dense()
else:
new_adj_mat = (new_adj_mat @ adj_mat).bool().float()
attr_mat.masked_fill((new_adj_mat - attr_mat.bool().float()).bool(),
i + 1)
return new_adj_mat, attr_mat
def spspmm(a, b, separate=False):
n = a.size(0)
m = a.size(1)
assert m == b.size(0)
k = b.size(1)
ai, av, bi, bv = get_iv([a, b])
print('n,m,k=', n, m, k)
print('ai, av, bi, bv=', apply(lambda x: x.numel(), ai, av, bi, bv))
i, v = torch_sparse.spspmm(ai, av, bi, bv, n, m, k)
if separate:
nonzero_mask = v != 0.
return i[:, nonzero_mask], v[nonzero_mask], [n, k]
return torch.sparse_coo_tensor(i, v, [n, k])
def spspmm(a, b, separate=False):
n = a.size(0)
m = a.size(1)
assert m == b.size(0)
k = b.size(1)
a = a.coalesce()
b = b.coalesce()
ai, av = a._indices(), a._values()
bi, bv = b._indices(), b._values()
del a, b
i, v = torch_sparse.spspmm(ai, av, bi, bv, n, m, k)
if separate:
nonzero_mask = v != 0.
return i[:, nonzero_mask], v[nonzero_mask], [n, k]
return torch.sparse_coo_tensor(i, v, [n, k])
def forward(self, x, edge_index, edge_attr):
device = x.device
N, dim = x.shape
diag_val = self.p(x)
diag_val = F.sigmoid(diag_val)
self.dropout(diag_val)
row, col = edge_index
deg = degree(col, x.size(0), dtype=x.dtype)
deg_inv = deg.pow(-1)
edge_attr_t = deg_inv[row] * edge_attr
diag_ind = torch.LongTensor([range(N)] * 2).to(device)
_, adj_mat_val = spspmm(edge_index, edge_attr_t, diag_ind, diag_val.view(-1), N, N, N, True)
return edge_index, adj_mat_val
def forward(self, graph, x):
edge_index = graph.edge_index
N, dim = x.shape
# nl_adj_mat_ind, nl_adj_mat_val = add_self_loops(edge_index, num_nodes=N)[0], edge_attr.squeeze()
nl_adj_mat_ind = add_remaining_self_loops(edge_index, num_nodes=N)[0]
nl_adj_mat_ind = torch.stack(nl_adj_mat_ind)
nl_adj_mat_val = torch.ones(nl_adj_mat_ind.shape[1]).to(x.device)
for _ in range(self.nhop - 1):
nl_adj_mat_ind, nl_adj_mat_val = spspmm(nl_adj_mat_ind,
nl_adj_mat_val,
nl_adj_mat_ind,
nl_adj_mat_val, N, N, N,
True)
result = []
for i in range(self.subheads):
h = torch.mm(x, self.weight[i])
adj_mat_ind, adj_mat_val = nl_adj_mat_ind, nl_adj_mat_val
h = F.dropout(h, p=self.dropout, training=self.training)
adj_mat_ind, adj_mat_val = self.attention(h, adj_mat_ind,
adj_mat_val)
# laplacian matrix normalization
adj_mat_val = self.normalization(adj_mat_ind, adj_mat_val, N)
val_h = h
with graph.local_graph():
graph.edge_index = adj_mat_ind
graph.edge_weight = adj_mat_val
for _ in range(i + 1):
val_h = spmm(graph, val_h)
# val_h = spmm(adj_mat_ind, F.dropout(adj_mat_val, p=self.node_dropout, training=self.training), N, N, val_h)
# val_h = val_h / norm
val_h[val_h != val_h] = 0
val_h = val_h + self.bias[i]
val_h = self.adaptive_enc[i](val_h)
val_h = self.activation(val_h)
val_h = F.dropout(val_h,
p=self.dropout,
training=self.training)
result.append(val_h)
h_res = torch.cat(result, dim=1)
return h_res
def top_k(self, graph, x: torch.Tensor, scores: torch.Tensor) -> Tuple[Graph, torch.Tensor]:
org_n_nodes = x.shape[0]
num = int(self.pooling_rate * x.shape[0])
values, indices = torch.topk(scores, max(2, num))
if self.aug_adj:
edge_index = graph.edge_index.cpu()
edge_attr = torch.ones(edge_index.shape[1])
edge_index, _ = spspmm(edge_index, edge_attr, edge_index, edge_attr, org_n_nodes, org_n_nodes, org_n_nodes)
edge_index = edge_index.to(x.device)
batch = Graph(x=x, edge_index=edge_index)
else:
batch = graph
new_batch = batch.subgraph(indices)
new_batch.row_norm()
return new_batch, indices
def nth_deg_adjacency(adj_mat, n=1, sparse=False):
""" Calculates the n-th degree adjacency matrix.
Performs mm of adj_mat and adds the newly added.
Default is dense. Mods for sparse version are done when needed.
Inputs:
* adj_mat: (N, N) adjacency tensor
* n: int. degree of the output adjacency
* sparse: bool. whether to use torch-sparse module
Outputs:
* edge_idxs: the ij positions of the adjacency matrix
* edge_attrs: the degree of connectivity (1 for neighs, 2 for neighs^2 )
"""
adj_mat = adj_mat.float()
attr_mat = torch.zeros_like(adj_mat)
for i in range(n):
if i == 0:
attr_mat += adj_mat
continue
if i == 1 and sparse:
# create sparse adj tensor
adj_mat = torch.sparse.FloatTensor(adj_mat.nonzero().t(),
adj_mat[adj_mat != 0]).to(
adj_mat.device).coalesce()
idxs, vals = adj_mat.indices(), adj_mat.values()
m, k, n = 3 * [
adj_mat.shape[0]
] # (m, n) * (n, k) , but adj_mats are squared: m=n=k
if sparse:
idxs, vals = torch_sparse.spspmm(idxs,
vals,
idxs,
vals,
m=m,
k=k,
n=n)
adj_mat = torch.zeros_like(attr_mat)
adj_mat[idxs[0], idxs[1]] = vals.bool().float()
else:
adj_mat = (adj_mat @ adj_mat).bool().float()
attr_mat[(adj_mat - attr_mat.bool().float()).bool()] += i + 1
return adj_mat, attr_mat
def forward(self, A, H_=None):
if self.first == True:
result_A = self.conv1(A)
result_B = self.conv2(A)
W = [(F.softmax(self.conv1.weight, dim=1)).detach(),(F.softmax(self.conv2.weight, dim=1)).detach()]
else:
result_A = H_
result_B = self.conv1(A)
W = [(F.softmax(self.conv1.weight, dim=1)).detach()]
H = []
for i in range(len(result_A)):
a_edge, a_value = result_A[i]
b_edge, b_value = result_B[i]
edges, values = torch_sparse.spspmm(a_edge, a_value, b_edge, b_value, self.num_nodes, self.num_nodes, self.num_nodes)
H.append((edges, values))
return H, W
def sgc_precompute(features, adj, degree):
t = perf_counter()
adj_index = adj.coalesce().indices()
adj_value = adj.coalesce().values()
features_index = features.coalesce().indices()
features_value = features.coalesce().values()
m = adj.shape[0]
n = adj.shape[1]
k = features.shape[1]
for i in range(degree):
#features = torch.spmm(adj, features)
features_index, features_value = torch_sparse.spspmm(
adj_index, adj_value, features_index, features_value, m, n, k)
precompute_time = perf_counter() - t
return torch.sparse.FloatTensor(features_index, features_value,
torch.Size(
features.shape)), precompute_time