def _call_single(self, X, A): # Reshape kernels for efficient message-passing kernel = tf.reshape(self.kernel, (-1, self.attn_heads * self.channels)) attn_kernel_self = ops.transpose(self.attn_kernel_self, (2, 1, 0)) attn_kernel_neighs = ops.transpose(self.attn_kernel_neighs, (2, 1, 0)) # Prepare message-passing indices = A.indices N = tf.shape(X, out_type=indices.dtype)[0] indices = ops.sparse_add_self_loops(indices, N) targets, sources = indices[:, -2], indices[:, -1] # Update node features X = ops.dot(X, kernel) X = tf.reshape(X, (-1, self.attn_heads, self.channels)) # Compute attention attn_for_self = tf.reduce_sum(X * attn_kernel_self, -1) attn_for_self = tf.gather(attn_for_self, targets) attn_for_neighs = tf.reduce_sum(X * attn_kernel_neighs, -1) attn_for_neighs = tf.gather(attn_for_neighs, sources) attn_coef = attn_for_self + attn_for_neighs attn_coef = tf.nn.leaky_relu(attn_coef, alpha=0.2) attn_coef = ops.unsorted_segment_softmax(attn_coef, targets, N) attn_coef = self.dropout(attn_coef) attn_coef = attn_coef[..., None] # Update representation output = attn_coef * tf.gather(X, sources) output = ops.scatter_sum(output, targets, N) return output, attn_coef
def call(self, inputs, mask=None): x, a, e = inputs # Parameters N = tf.shape(x)[-2] F = tf.shape(x)[-1] F_ = self.channels # Filter network kernel_network = e for layer in self.kernel_network_layers: kernel_network = layer(kernel_network) # Convolution mode = ops.autodetect_mode(x, a) if mode == modes.BATCH: kernel = K.reshape(kernel_network, (-1, N, N, F_, F)) output = kernel * a[..., None, None] output = tf.einsum("abcde,ace->abd", output, x) else: # Enforce sparse representation if not K.is_sparse(a): warnings.warn("Casting dense adjacency matrix to SparseTensor." "This can be an expensive operation. ") a = tf.sparse.from_dense(a) target_shape = (-1, F, F_) if mode == modes.MIXED: target_shape = (tf.shape(x)[0], ) + target_shape kernel = tf.reshape(kernel_network, target_shape) index_i = a.indices[:, 1] index_j = a.indices[:, 0] messages = tf.gather(x, index_j, axis=-2) messages = tf.einsum("...ab,...abc->...ac", messages, kernel) output = ops.scatter_sum(messages, index_i, N) if self.root: output += K.dot(x, self.root_kernel) if self.use_bias: output = K.bias_add(output, self.bias) if mask is not None: output *= mask[0] output = self.activation(output) return output
def call(self, inputs): features = inputs[0] fltr = inputs[1] # Enforce sparse representation if not K.is_sparse(fltr): fltr = ops.dense_to_sparse(fltr) # Propagation targets = fltr.indices[:, -2] sources = fltr.indices[:, -1] messages = tf.gather(features, sources) aggregated = ops.scatter_sum(targets, messages, N=tf.shape(features)[0]) hidden = (1.0 + self.eps) * features + aggregated # MLP output = self.mlp(hidden) return output
def _call_single(self, inputs): X = inputs[0] # (N, F) A = inputs[1] # (N, N) E = inputs[2] # (n_edges, S) assert K.ndim( E) == 2, 'In single mode, E must have shape (n_edges, S).' # Enforce sparse representation if not K.is_sparse(A): A = ops.dense_to_sparse(A) # Parameters N = tf.shape(X)[-2] F = K.int_shape(X)[-1] F_ = self.channels # Filter network kernel_network = E for l in self.kernel_network_layers: kernel_network = l(kernel_network) # (n_edges, F * F_) target_shape = (-1, F, F_) kernel = tf.reshape(kernel_network, target_shape) # Propagation index_i = A.indices[:, -2] index_j = A.indices[:, -1] messages = tf.gather(X, index_j) messages = ops.dot(messages[:, None, :], kernel)[:, 0, :] aggregated = ops.scatter_sum(messages, index_i, N) # Update output = aggregated if self.root: output += ops.dot(X, self.root_kernel) 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_single(self, inputs): x, a, e = inputs if K.ndim(e) != 2: raise ValueError('In single mode, E must have shape ' '(n_edges, n_edge_features).') # Enforce sparse representation if not K.is_sparse(a): a = ops.dense_to_sparse(a) # Parameters N = tf.shape(x)[-2] F = K.int_shape(x)[-1] F_ = self.channels # Filter network kernel_network = e for layer in self.kernel_network_layers: kernel_network = layer(kernel_network) # (n_edges, F * F_) target_shape = (-1, F, F_) kernel = tf.reshape(kernel_network, target_shape) # Propagation index_i = a.indices[:, -2] index_j = a.indices[:, -1] messages = tf.gather(x, index_j) messages = ops.dot(messages[:, None, :], kernel)[:, 0, :] aggregated = ops.scatter_sum(messages, index_i, N) # Update output = aggregated if self.root: output += ops.dot(x, self.root_kernel) if self.use_bias: output = K.bias_add(output, self.bias) if self.activation is not None: output = self.activation(output) return output