def target_func(batch_adj, required_nodes, final_nodes):
"""Create training targets given the batch's adjacency matrix, the subset of nodes it used as input,
and the set of nodes for which output values are computed.
In the minibatch setting, `required_nodes` is usually the 2-hop neighbourhood of `final_nodes`.
In the fullbatch setting all this is constant and equal to the values of the full dataset.
"""
# Keras requires a target value for each output of the model trained,
# so we have one entry per output/loss in the model. Each entry must also have
# a shape which is compatible with the corresponding output, which is why we have
# the `utils.expand_dims_tile()` calls: the target value is the same for each of the
# values generated from the `n_ξ_samples` embeddings.
return [
# Keras feeds this to the embedding-gaussian divergence loss function,
# but we ignore this input in the actual loss function.
np.zeros(1), # ignored
# Target value for adjacency reconstruction
utils.expand_dims_tile(utils.expand_dims_tile(batch_adj + np.eye(batch_adj.shape[0]),
0, n_ξ_samples),
0, 1),
# Target value for feature reconstruction
utils.expand_dims_tile(labels[final_nodes], 1, n_ξ_samples),
]
def stochastic_value(self, n_samples):
"""TODOC"""
μ_shape = tf.shape(self.μ)
ε = tf.random_normal(tf.concat([μ_shape[:-2], [n_samples], μ_shape[-2:]], 0),
dtype=K.floatx())
return K.squeeze(expand_dims_tile(self.μ, -3, n_samples)
+ expand_dims_tile(self.R, -3, n_samples) @ ε,
-1)
def target_func(batch_adj, required_nodes, final_nodes):
return [
np.zeros(1), # ignored
utils.expand_dims_tile(
utils.expand_dims_tile(batch_adj + np.eye(batch_adj.shape[0]), 0,
n_ξ_samples), 0, 1),
utils.expand_dims_tile(labels[final_nodes], 1, n_ξ_samples),
]