def call(self, inputs):
X = inputs[0] # (batch_size, N, F)
A = inputs[1] # (batch_size, N, N)
E = inputs[2] # (n_edges, S) or (batch_size, N, N, S)
mode = ops.autodetect_mode(A, X)
if mode == modes.SINGLE:
return self._call_single(inputs)
# Parameters
N = K.shape(X)[-2]
F = K.int_shape(X)[-1]
F_ = self.channels
# Normalize adjacency matrix
A = ops.normalize_A(A)
# Filter network
kernel_network = E
for l in self.kernel_network_layers:
kernel_network = l(kernel_network)
# Convolution
target_shape = (-1, N, N, F_, F) if mode == modes.BATCH else (N, N, F_, F)
kernel = K.reshape(kernel_network, target_shape)
output = kernel * A[..., None, None]
output = tf.einsum('abicf,aif->abc', output, X)
if self.use_bias:
output = K.bias_add(output, self.bias)
if self.activation is not None:
output = self.activation(output)
return output
def call(self, inputs):
X, A = inputs
N = K.shape(A)[-1]
# Check if the layer is operating in mixed or batch mode
mode = ops.autodetect_mode(X, A)
self.reduce_loss = mode in (modes.MIXED, modes.BATCH)
# Get normalized adjacency
if K.is_sparse(A):
I_ = tf.sparse.eye(N, dtype=A.dtype)
A_ = tf.sparse.add(A, I_)
else:
I_ = tf.eye(N, dtype=A.dtype)
A_ = A + I_
fltr = ops.normalize_A(A_)
# Node embeddings
Z = K.dot(X, self.kernel_emb)
Z = ops.modal_dot(fltr, Z)
if self.activation is not None:
Z = self.activation(Z)
# Compute cluster assignment matrix
S = K.dot(X, self.kernel_pool)
S = ops.modal_dot(fltr, S)
S = activations.softmax(S, axis=-1) # softmax applied row-wise
# Link prediction loss
S_gram = ops.modal_dot(S, S, transpose_b=True)
if mode == modes.MIXED:
A = tf.sparse.to_dense(A)[None, ...]
if K.is_sparse(A):
LP_loss = tf.sparse.add(A, -S_gram) # A/tf.norm(A) - S_gram/tf.norm(S_gram)
else:
LP_loss = A - S_gram
LP_loss = tf.norm(LP_loss, axis=(-1, -2))
if self.reduce_loss:
LP_loss = K.mean(LP_loss)
self.add_loss(LP_loss)
# Entropy loss
entr = tf.negative(
tf.reduce_sum(tf.multiply(S, K.log(S + K.epsilon())), axis=-1)
)
entr_loss = K.mean(entr, axis=-1)
if self.reduce_loss:
entr_loss = K.mean(entr_loss)
self.add_loss(entr_loss)
# Pooling
X_pooled = ops.modal_dot(S, Z, transpose_a=True)
A_pooled = ops.matmul_at_b_a(S, A)
output = [X_pooled, A_pooled]
if self.return_mask:
output.append(S)
return output
def connect(self, a, s, **kwargs):
a_pool = ops.matmul_at_b_a(s, a)
# Post-processing of A
a_pool = tf.linalg.set_diag(
a_pool, tf.zeros(K.shape(a_pool)[:-1], dtype=a_pool.dtype))
a_pool = ops.normalize_A(a_pool)
return a_pool
def call(self, inputs):
if len(inputs) == 3:
X, A, I = inputs
if K.ndim(I) == 2:
I = I[:, 0]
else:
X, A = inputs
I = None
# Check if the layer is operating in batch mode (X and A have rank 3)
batch_mode = K.ndim(X) == 3
# Compute cluster assignment matrix
S = self.mlp(X)
# MinCut regularization
A_pooled = ops.matmul_at_b_a(S, A)
num = tf.linalg.trace(A_pooled)
D = ops.degree_matrix(A)
den = tf.linalg.trace(ops.matmul_at_b_a(S, D)) + K.epsilon()
cut_loss = -(num / den)
if batch_mode:
cut_loss = K.mean(cut_loss)
self.add_loss(cut_loss)
# Orthogonality regularization
SS = ops.modal_dot(S, S, transpose_a=True)
I_S = tf.eye(self.k, dtype=SS.dtype)
ortho_loss = tf.norm(
SS / tf.norm(SS, axis=(-1, -2), keepdims=True) -
I_S / tf.norm(I_S),
axis=(-1, -2),
)
if batch_mode:
ortho_loss = K.mean(ortho_loss)
self.add_loss(ortho_loss)
# Pooling
X_pooled = ops.modal_dot(S, X, transpose_a=True)
A_pooled = tf.linalg.set_diag(
A_pooled, tf.zeros(K.shape(A_pooled)[:-1],
dtype=A_pooled.dtype)) # Remove diagonal
A_pooled = ops.normalize_A(A_pooled)
output = [X_pooled, A_pooled]
if I is not None:
I_mean = tf.math.segment_mean(I, I)
I_pooled = ops.repeat(I_mean, tf.ones_like(I_mean) * self.k)
output.append(I_pooled)
if self.return_mask:
output.append(S)
return output
def call(self, inputs, mask=None):
x, a, i = self.get_inputs(inputs)
# Graph filter for GNNs
if K.is_sparse(a):
i_n = tf.sparse.eye(self.n_nodes, dtype=a.dtype)
a_ = tf.sparse.add(a, i_n)
else:
i_n = tf.eye(self.n_nodes, dtype=a.dtype)
a_ = a + i_n
fltr = ops.normalize_A(a_)
output = self.pool(x, a, i, fltr=fltr, mask=mask)
return output
def connect(self, a, s, **kwargs):
a_pool = ops.matmul_at_b_a(s, a)
# MinCut loss
cut_loss = self.mincut_loss(a, s, a_pool)
if K.ndim(a) == 3:
cut_loss = K.mean(cut_loss)
self.add_loss(cut_loss)
# Post-processing of A
a_pool = tf.linalg.set_diag(
a_pool, tf.zeros(K.shape(a_pool)[:-1], dtype=a_pool.dtype))
a_pool = ops.normalize_A(a_pool)
return a_pool
def call(self, inputs):
x, a, i = self.get_inputs(inputs)
# Graph filter for GNN
if K.is_sparse(a):
i_n = tf.sparse.eye(self.n_nodes, dtype=a.dtype)
a_ = tf.sparse.add(a, i_n)
else:
i_n = tf.eye(self.n_nodes, dtype=a.dtype)
a_ = a + i_n
fltr = ops.normalize_A(a_)
y = ops.modal_dot(fltr, K.dot(x, self.kernel))
output = self.pool(x, a, i, y=y)
if self.return_score:
output.append(y)
return output
def call(self, inputs):
# Note that I is useless, because thee layer cannot be used in graph
# batch mode.
if len(inputs) == 3:
X, A, I = inputs
else:
X, A = inputs
I = None
N = K.shape(A)[-1]
# Check if the layer is operating in batch mode (X and A have rank 3)
mode = ops.autodetect_mode(A, X)
self.reduce_loss = mode in (ops._modes['M'], ops._modes['B'])
# Get normalized adjacency
if K.is_sparse(A):
I_ = tf.sparse.eye(N, dtype=A.dtype)
A_ = tf.sparse.add(A, I_)
else:
I_ = tf.eye(N, dtype=A.dtype)
A_ = A + I_
fltr = ops.normalize_A(A_)
# Node embeddings
Z = K.dot(X, self.kernel_emb)
Z = ops.filter_dot(fltr, Z)
if self.activation is not None:
Z = self.activation(Z)
# Compute cluster assignment matrix
S = K.dot(X, self.kernel_pool)
S = ops.filter_dot(fltr, S)
S = activations.softmax(S, axis=-1) # softmax applied row-wise
# Link prediction loss
S_gram = ops.matmul_A_BT(S, S)
if K.is_sparse(A):
LP_loss = tf.sparse.add(A, -S_gram) # A/tf.norm(A) - S_gram/tf.norm(S_gram)
else:
LP_loss = A - S_gram
LP_loss = tf.norm(LP_loss, axis=(-1, -2))
if self.reduce_loss:
LP_loss = K.mean(LP_loss)
self.add_loss(LP_loss)
# Entropy loss
entr = tf.negative(tf.reduce_sum(tf.multiply(S, K.log(S + K.epsilon())), axis=-1))
entr_loss = K.mean(entr, axis=-1)
if self.reduce_loss:
entr_loss = K.mean(entr_loss)
self.add_loss(entr_loss)
# Pooling
X_pooled = ops.matmul_AT_B(S, Z)
A_pooled = ops.matmul_AT_B_A(S, A)
if K.ndim(A_pooled) == 3:
self.mixed_mode = True
output = [X_pooled, A_pooled]
if I is not None:
I_mean = tf.segment_mean(I, I)
I_pooled = ops.repeat(I_mean, tf.ones_like(I_mean) * self.k)
output.append(I_pooled)
if self.return_mask:
output.append(S)
return output
def call(self, inputs):
# Note that I is useless, because thee layer cannot be used in graph
# batch mode.
if len(inputs) == 3:
X, A, I = inputs
else:
X, A = inputs
I = None
# Check if the layer is operating in batch mode (X and A have rank 3)
batch_mode = K.ndim(A) == 3
# Optionally compute hidden layer
if self.h is None:
Hid = X
else:
Hid = K.dot(X, self.kernel_in)
if self.use_bias:
Hid = K.bias_add(Hid, self.bias_in)
if self.activation is not None:
Hid = self.activation(Hid)
# Compute cluster assignment matrix
S = K.dot(Hid, self.kernel_out)
if self.use_bias:
S = K.bias_add(S, self.bias_out)
S = activations.softmax(S, axis=-1) # Apply softmax to get cluster assignments
# MinCut regularization
A_pooled = ops.matmul_AT_B_A(S, A)
num = tf.trace(A_pooled)
D = ops.degree_matrix(A)
den = tf.trace(ops.matmul_AT_B_A(S, D))
cut_loss = -(num / den)
if batch_mode:
cut_loss = K.mean(cut_loss)
self.add_loss(cut_loss)
# Orthogonality regularization
SS = ops.matmul_AT_B(S, S)
I_S = tf.eye(self.k)
ortho_loss = tf.norm(
SS / tf.norm(SS, axis=(-1, -2)) - I_S / tf.norm(I_S), axis=(-1, -2)
)
if batch_mode:
ortho_loss = K.mean(cut_loss)
self.add_loss(ortho_loss)
# Pooling
X_pooled = ops.matmul_AT_B(S, X)
A_pooled = tf.linalg.set_diag(A_pooled, tf.zeros(K.shape(A_pooled)[:-1])) # Remove diagonal
A_pooled = ops.normalize_A(A_pooled)
output = [X_pooled, A_pooled]
if I is not None:
I_mean = tf.segment_mean(I, I)
I_pooled = ops.repeat(I_mean, tf.ones_like(I_mean) * self.k)
output.append(I_pooled)
if self.return_mask:
output.append(S)
return output
