def p_builder(p_input):
# CANDO: change activation
p_layer1 = keras.layers.Dense(dim_l1, use_bias=use_bias, activation='relu',
kernel_regularizer='l2', bias_regularizer='l2',
name='p_layer1')(p_input)
p_adj = layers.Bilinear(0, use_bias=use_bias,
kernel_regularizer='l2', bias_regularizer='l2',
name='p_adj')([p_layer1, p_layer1])
p_v = keras.layers.Dense(dim_data, use_bias=use_bias,
kernel_regularizer='l2', bias_regularizer='l2',
name='p_v')(p_layer1)
return ([p_adj, p_v], ('SigmoidBernoulliScaledAdjacency', 'SigmoidBernoulli'))
def p_builder(p_input):
# Get slices of the embeddings for each prediction
p_input_adj = layers.InnerSlice(adj_embedding_slice)(p_input)
p_input_v = layers.InnerSlice(v_embedding_slice)(p_input)
# Build layer1 for adj and features, if requested, and have
# it shared between adjacency and features if requested.
if share_l1:
# Shared intermediate layer.
assert with_l1, "Can't share l1 if l1 is not requested"
p_layer1 = keras.layers.Dense(dim_l1,
use_bias=use_bias,
activation='relu',
kernel_regularizer='l2',
bias_regularizer='l2',
name='p_layer1')
p_penultimate_adj = p_layer1(p_input_adj)
p_penultimate_v = p_layer1(p_input_v)
else:
if with_l1:
# Unshared intermediate layer.
p_penultimate_adj = keras.layers.Dense(
dim_l1,
use_bias=use_bias,
activation='relu',
kernel_regularizer='l2',
bias_regularizer='l2',
name='p_layer1_adj')(p_input_adj)
p_penultimate_v = keras.layers.Dense(
dim_l1,
use_bias=use_bias,
activation='relu',
kernel_regularizer='l2',
bias_regularizer='l2',
name='p_layer1_v')(p_input_v)
else:
# No intermediate layer.
p_penultimate_adj = p_input_adj
p_penultimate_v = p_input_v
# Prepare kwargs for the Bilinear adj decoder, then build it.
adj_kwargs = {}
if adj_kernel is not None:
adj_kwargs['fixed_kernel'] = adj_kernel
else:
adj_kwargs['kernel_regularizer'] = 'l2'
p_adj = layers.Bilinear(
0,
use_bias=use_bias,
name='p_adj',
bias_regularizer='l2',
**adj_kwargs)([p_penultimate_adj, p_penultimate_adj])
# Finally build the feature decoder according to the requested codec.
if feature_codec in ['SigmoidBernoulli', 'SoftmaxMultinomial']:
p_v = keras.layers.Dense(dim_data,
use_bias=use_bias,
kernel_regularizer='l2',
bias_regularizer='l2',
name='p_v')(p_penultimate_v)
else:
assert feature_codec == 'OrthogonalGaussian'
p_v_μ_flat = keras.layers.Dense(
dim_data,
use_bias=use_bias,
kernel_regularizer='l2',
bias_regularizer='l2',
name='p_v_mu_flat')(p_penultimate_v)
p_v_logS_flat = keras.layers.Dense(
dim_data,
use_bias=use_bias,
kernel_regularizer='l2',
bias_regularizer='l2',
name='p_v_logS_flat')(p_penultimate_v)
p_v = keras.layers.Concatenate(name='p_v_mulogS_flat')(
[p_v_μ_flat, p_v_logS_flat])
return ([p_adj,
p_v], ('SigmoidBernoulliScaledAdjacency', feature_codec))