def _build(self):
self.layers.append(
layers.GraphConvolution(
input_dim=self.input_dim
if not self.precalc else self.input_dim * 2,
output_dim=FLAGS.hidden1,
placeholders=self.placeholders,
act=tf.nn.relu,
dropout=True,
sparse_inputs=self.sparse_inputs, # False
logging=self.logging,
norm=self.norm,
precalc=self.precalc,
residual=self.residual))
for _ in range(self.num_layers - 2):
self.layers.append(
layers.GraphConvolution(
input_dim=FLAGS.hidden1
if not self.residual else FLAGS.hidden1 * 2,
output_dim=FLAGS.hidden1,
placeholders=self.placeholders,
act=tf.nn.relu,
dropout=True,
sparse_inputs=False,
logging=self.logging,
norm=self.norm,
precalc=False,
residual=self.residual))
self.layers.append(
layers.GraphConvolution(input_dim=FLAGS.hidden1 if
not self.residual else FLAGS.hidden1 * 2,
output_dim=self.output_dim,
placeholders=self.placeholders,
act=lambda x: x,
dropout=True,
logging=self.logging,
norm=False,
precalc=False,
residual=self.residual))
def __init__(self, args, adj, node_num, re_hidden, dis_hidden, vi_feature,
kernel_size, alphi1, alphi2, alphi3):
super(MTGC, self).__init__()
# Mask Temporal Graph Convolution
self.adj = adj
self.node_num = node_num
self.re_hidden = re_hidden
self.dis_hidden = dis_hidden
self.vi_feature = vi_feature
self.kernel_size = kernel_size
self.alphi1 = alphi1
self.alphi2 = alphi2
self.alphi3 = alphi3
# feature extraction
self.s_conv = layers.Convolutinon1D(in_features=node_num,
kernal_size=self.kernel_size)
self.s_gcn_e = layers.GraphConvolution(in_features=self.node_num,
out_features=self.node_num)
self.x_conv = layers.Convolutinon1D(in_features=node_num,
kernal_size=self.kernel_size)
self.x_gcn_e = layers.GraphConvolution(in_features=self.node_num,
out_features=self.node_num)
self.x_vae_tras = nn.Linear(self.kernel_size * self.node_num,
self.node_num)
self.s_vae_tras = nn.Linear(self.kernel_size * self.node_num,
self.node_num)
self.fusion = nn.Linear(self.node_num * 2, self.node_num)
# reconstruction network
self.vae = layers.VariationalAutoencoder(
node_num=self.node_num,
hidden_features=self.re_hidden,
vi_feature=self.vi_feature)
self.x_vae_traspose = nn.Linear(self.node_num,
self.kernel_size * self.node_num)
# feature extraction
self.gcn_d = layers.GraphConvolution(in_features=self.node_num,
out_features=self.node_num)
self.transpose_conv = layers.ConvolutinonTranspose1D(
in_features=node_num, kernal_size=self.kernel_size)
self.gcn_d_std = layers.GraphConvolution(in_features=self.node_num,
out_features=self.node_num)
self.transpose_conv_std = layers.ConvolutinonTranspose1D(
in_features=node_num, kernal_size=self.kernel_size)
# Discriminator
self.mask_discriminator = layers.Discriminator(
node_num=self.node_num,
kernel_size=kernel_size,
vi_feature=self.vi_feature,
hidden_features=self.dis_hidden)
self.args = args