def get_adj(self, n, edges):
adj = sp.coo_matrix((np.ones(len(edges)), (edges[:, 0], edges[:, 1])),
shape=(n, n),
dtype='float32')
adj_ = spm_to_tensor(normt_spm(adj, method='in')).cuda()
r_adj = spm_to_tensor(normt_spm(adj.transpose(), method='in')).cuda()
return adj_, r_adj
def __init__(self, n, edges_set, in_channels, out_channels, hidden_layers):
super().__init__()
self.n = n
self.d = len(edges_set)
self.a_adj_set = []
self.r_adj_set = []
for edges in edges_set:
edges = np.array(edges)
adj = sp.coo_matrix(
(np.ones(len(edges)), (edges[:, 0], edges[:, 1])),
shape=(n, n),
dtype='float32')
a_adj = spm_to_tensor(normt_spm(adj, method='in')).cuda()
r_adj = spm_to_tensor(normt_spm(adj.transpose(),
method='in')).cuda()
self.a_adj_set.append(a_adj)
self.r_adj_set.append(r_adj)
hl = hidden_layers.split(',')
if hl[-1] == 'd':
dropout_last = True
hl = hl[:-1]
else:
dropout_last = False
self.a_att = nn.Parameter(torch.ones(self.d))
self.r_att = nn.Parameter(torch.ones(self.d))
i = 0
layers = []
last_c = in_channels
for c in hl:
if c[0] == 'd':
dropout = True
c = c[1:]
else:
dropout = False
c = int(c)
i += 1
conv = GraphConv(last_c, c, dropout=dropout)
self.add_module('conv{}'.format(i), conv)
layers.append(conv)
last_c = c
conv = GraphConv(last_c,
out_channels,
relu=False,
dropout=dropout_last)
self.add_module('conv-last', conv)
layers.append(conv)
self.layers = layers
def get_adj(self, weights, edges, n):
adj = sp.coo_matrix((weights, (edges[:, 0], edges[:, 1])),
shape=(n, n),
dtype='float32')
adj = normt_spm(adj, method='in')
adj = spm_to_tensor(adj)
return adj.cuda()
def __init__(self,
n,
edges,
in_channels,
out_channels,
hidden_layers,
norm_method='in'):
super().__init__()
edges = np.array(edges)
adj = sp.coo_matrix((np.ones(len(edges)), (edges[:, 0], edges[:, 1])),
shape=(n, n),
dtype='float32')
adj = normt_spm(adj, method=norm_method)
adj = spm_to_tensor(adj)
self.adj = adj.cuda()
hl = hidden_layers.split(',')
if hl[-1] == 'd':
dropout_last = True
hl = hl[:-1]
else:
dropout_last = False
i = 0
layers = []
last_c = in_channels
for c in hl:
if c[0] == 'd':
dropout = True
c = c[1:]
else:
dropout = False
c = int(c)
i += 1
conv = GraphConv(last_c, c, dropout=dropout)
self.add_module('conv{}'.format(i), conv)
layers.append(conv)
last_c = c
conv = GraphConv(last_c,
out_channels,
relu=False,
dropout=dropout_last)
self.add_module('conv-last', conv)
layers.append(conv)
self.layers = layers
def __init__(self, n, edges_set, weights_set, in_channels, out_channels, hidden_layers):
super().__init__()
self.n = n
self.d = len(edges_set)
self.a_adj_set = []
for i, edges in enumerate(edges_set):
edges = np.array(edges)
adj = sp.coo_matrix((np.array(weights_set[i]), (edges[:, 0], edges[:, 1])),
shape=(n, n), dtype='float32')
a_adj = spm_to_tensor(normt_spm(adj, method='in')).cuda()
self.a_adj_set.append(a_adj)
print(self.a_adj_set)
hl = hidden_layers.split(',')
if hl[-1] == 'd':
dropout_last = True
hl = hl[:-1]
else:
dropout_last = False
i = 0
layers = []
last_c = in_channels
for c in hl:
if c[0] == 'd':
dropout = True
c = c[1:]
else:
dropout = False
c = int(c)
i += 1
conv = GraphConv(last_c, c, self.d, dropout=dropout)
self.add_module('conv{}'.format(i), conv)
layers.append(conv)
last_c = c
conv = GraphConv(last_c, out_channels, self.d, relu=False, dropout=dropout_last)
self.add_module('conv-last', conv)
layers.append(conv)
self.layers = layers
def __init__(self, n, edges, in_channels, out_channels, hidden_layers, k):
super().__init__()
edges = np.array(edges)
adj = sp.coo_matrix((np.ones(len(edges)), (edges[:, 0], edges[:, 1])),
shape=(n, n),
dtype='float32')
adj = normt_spm(adj, method='in')
adj = spm_to_tensor(adj)
self.adj = adj.cuda()
self.k = k
hl = hidden_layers.split(',')
if hl[-1] == 'd':
dropout_last = True
hl = hl[:-1]
else:
dropout_last = False
layers = []
last_c = in_channels
for i, c in enumerate(hl):
if c[0] == 'd':
dropout = True
c = c[1:]
else:
dropout = False
c = int(c)
layers.append(FullyConnect(last_c, c, dropout=dropout))
last_c = c
layers.append(
FullyConnect(last_c,
out_channels,
relu=False,
dropout=dropout_last))
self.mlp = nn.Sequential(*layers)
