def test_get_number_nodes():
""" Get the number of nodes of a layer """
ii = K.constant(np.random.rand(2, 1))
nodes = 10
layer = Dense(units=nodes)
# Keras won't build the layer until it is called with some input
_ = layer(ii)
# Check that indeed the number of nodes is parsed correctly
assert nodes == get_number_nodes(layer)
def test_get_number_nodes():
""" Get the number of nodes of a layer """
nodes = 10
# Tensorflow won't build the layer until it is called in a model
input_layer = Input(shape=(1, ))
output_layer = Dense(units=nodes, name="test_layer")
modelito = EvolModel(input_layer, output_layer(input_layer))
# Check that indeed the number of nodes is parsed correctly
assert nodes == get_number_nodes(modelito)
def get_shape(self):
""" Study the model to get the shapes of all trainable weight as well
as the number of nodes. It also saves a reference to the non-trainable weights
in the system.
Returns
-------
`weight_shapes`: a list of the shapes of all trainable weights
"""
# Get trainable weight from the model and their shapes
trainable_weights = self.model.trainable_weights
weight_shapes = [
weight.shape.as_list() for weight in trainable_weights
]
weights = [weight for weight in trainable_weights]
self.n_nodes = get_number_nodes(self.model)
# check compatibility of the shape with the NGA optimizer
count_nodes = 0
for num, layer in enumerate(weights):
layer_shape = layer.shape.as_list()
num += 1
if num % 2 == 0:
count_nodes += np.array(layer_shape)
if np.array(layer_shape).size != 1:
raise ValueError(
f"The NGA optimizer expects a (weight-bias)\N{SUPERSCRIPT LATIN SMALL LETTER N} architecture, {layer.name} does not satisfy this condition"
)
else:
if np.array(layer_shape).size != 2:
raise ValueError(
f"The NGA optimizer expects a (weight-bias)\N{SUPERSCRIPT LATIN SMALL LETTER N} architecture, {layer.name} does not satisfy this condition"
)
if count_nodes != self.n_nodes:
raise ValueError(
"The number of nodes with a bias attribute differs from the number of trainable nodes found based on the architecture of the trainable weights."
)
return weight_shapes
def get_shape(self):
""" Study the model to get the shapes of all trainable weight as well
as the number of nodes. It also saves a reference to the non-trainable weights
in the system.
Returns
-------
`weight_shapes`: a list of the shapes of all trainable weights
"""
# Initialize number of nodes
self.n_nodes = 0
# Get trainable weight from the model and their shapes
trainable_weights = self.model.trainable_weights
weight_shapes = [
weight.shape.as_list() for weight in trainable_weights
]
# TODO: eventually we should save here a reference to the layer and their
# corresponding weights, since the nodes are the output of the layer
# and the weights the corresponding to that layer
for layer in self.model.layers:
self.n_nodes += get_number_nodes(layer)
# TODO related to previous TODO: non trianable weights should not be important
self.non_training_weights = self.model.non_trainable_weights
return weight_shapes