Networks

This module contains the classes to define and create neural network objects. The abstract class NeuralNetwork represents a generic computational graph with layers as nodes and connections as edges. The children of NeuralNetwork are SequentialNetwork and AcyclicNetwork, to represent standard feed-forward NNs and residual NNs, respectively.

class pynever.networks.NeuralNetwork(identifier, input_ids)[source]

Bases: ABC

An abstract class used for our internal representation of a generic Neural Network. It consists of a graph of LayerNode objects and edges connecting them. It should be noted that this data structure is not able to compute the input-output relation defined by the network. The properties of the computational graph are specialized in the concrete classes.

nodes[source]

Dictionary containing string keys and LayerNode values. It contains the nodes of the graph, the identifier of the node is used as a key in the nodes’ dictionary.

Type:

dict[str, LayerNode]

edges[source]

Dictionary of LayerNode identifiers. It contains for each node identified by the keys, the list of nodes connected to it.

Type:

dict[str, list[str]]

identifier[source]

Identifier of the Neural Network.

Type:

str

input_ids[source]

Dictionary containing the inputs of the networks as keys and the corresponding layer identifier of the node of which they are the input.

Type:

dict[str, str | None]

is_empty()[source]

Procedure to check whether the network is empty.

Returns:

True if there are no nodes in the network, False otherwise.

Return type:

bool

is_acyclic()[source]

Procedure to check whether the network is acyclic.

Returns:

True if the network is acyclic, False otherwise.

Return type:

bool

has_children(node)[source]

Procedure to check if a node has children.

Parameters:

node (ConcreteLayerNode) – The node of which the existence of its children should be checked.

Returns:

True if the node has children, False otherwise.

Return type:

bool

get_children(node)[source]

Procedure to return the children of a node as a list of ConcreteLayerNodes.

Parameters:

node (ConcreteLayerNode) – The node whose children should be returned.

Returns:

The children of the node passed as argument.

Return type:

list[ConcreteLayerNode]

has_parents(node)[source]

Procedure to check if a node has parents.

Parameters:

node (ConcreteLayerNode) – The node of which the existence of its parents should be checked.

Returns:

True if the node has parents, False otherwise.

Return type:

bool

get_parents(node)[source]

Procedure to return the parents of a node as a list of ConcreteLayerNodes.

Parameters:

node (ConcreteLayerNode) – The node whose parents should be returned

Returns:

The parents of the node passed as argument.

Return type:

list[ConcreteLayerNode]

get_input_id()[source]

Procedure to return the input_id of the network, assuming there is a single input layer.

Returns:

The input_id of the network.

Return type:

str

get_roots()[source]

Procedure to return the roots of the network as a list of ConcreteLayerNodes.

Returns:

The roots of the network as a list of ConcreteLayerNode objects.

Return type:

list[ConcreteLayerNode]

get_leaves()[source]

Procedure to return the leaves of the network as a list of ConcreteLayerNodes.

Returns:

The leaves of the network as a list of ConcreteLayerNode objects.

Return type:

list[ConcreteLayerNode]

get_topological_order(reverse=False)[source]

Procedure to walk the network with a DFS and build the topological sort.

Parameters:

reverse (bool) – Flag to reverse the order

Returns:

The topological sort of the network nodes identifiers as a stack.

Return type:

list[str]

layers_iterator(offset=0)[source]

Procedure to build a generator for the layers of the network in sequential order. It allows having an iterable interface when needed

Parameters:

offset (int) – Offset to start the generation

Returns:

The generator object

Return type:

Generator[ConcreteLayerNode | None, None]

get_first_node()[source]

Procedure to get the first ConcreteLayerNode of the network.

Returns:

The first node of the network.

Return type:

ConcreteLayerNode

get_next_node(node)[source]

Procedure to get the next ConcreteLayerNode of the network given an input ConcreteLayerNode.

Returns:

The next node of the network.

Return type:

ConcreteLayerNode

get_previous_node(node)[source]

Procedure to get the previous ConcreteLayerNode of the network given an input ConcreteLayerNode.

Returns:

The previous node in the network.

Return type:

ConcreteLaterNode

get_last_node()[source]

Procedure to get the last ConcreteLayerNode of the network.

Returns:

The last node of the network.

Return type:

ConcreteLayerNode

get_input_len()[source]

Procedure to count the number of inputs in in_dim

Returns:

The number of single inputs

Return type:

int

get_output_len()[source]

Procedure to count the number of outputs in out_dim

Returns:

The number of single outputs

Return type:

int

get_id_from_index(index)[source]

This method returns the identifier of the layer at the given index

Parameters:

index (int) – Index of the layer to return

Returns:

The identifier of the layer at the given index

Return type:

str

get_index_from_id(identifier)[source]

This method returns the index of the layer with the given identifier

Parameters:

identifier (str) – Identifier of the layer to return

Returns:

The index of the layer with the given identifier

Return type:

int

get_previous_id(identifier)[source]

Procedure to get the identifier of the previous layer given another layer identifier

Parameters:

identifier (str) – Identifier of the layer

Returns:

The identifier of the previous layer

Return type:

str

layer_precedes(layer_id1, layer_id2)[source]

Procedure to check whether a given layer precedes another or not

Parameters:

  • layer_id1 (str) – Identifier of the first layer

  • layer_id2 (str) – Identifier of the second layer

Returns:

True if layer1 precedes layer2, False otherwise

Return type:

bool

generic_add_node(node, parents=None, children=None, input_ids=None)[source]

Procedure to add a node to the network. A node cannot have both parents and inputs.

Parameters:
remove_node(node)[source]

Procedure to remove a node from the network.

Parameters:

node (ConcreteLayerNode) – The node to be removed.

delete_last_node()[source]

Procedure to remove the last ConcreteLayerNode from the network.

Returns:

The last node of the network.

Return type:

ConcreteLayerNode

count_relu_layers()[source]

Procedure to count the number of ReLU layers.

Returns:

The number of ReLU layers

Return type:

int

class pynever.networks.SequentialNetwork(identifier, input_id)[source]

Bases: NeuralNetwork

Concrete child of NeuralNetwork representing a sequential NeuralNetwork. It consists of a graph of ConcreteLayerNodes. It should be noted that this data structure is not able to compute the input-output relation defined by the network. The computational graph of a SequentialNetwork must correspond to a standard list.

append_node(node)[source]

Procedure to add a new ConcreteLayerNode. In a sequential network the new node must be connected directly to the previous node forming a list.

Parameters:

node (ConcreteLayerNode) – New node to add to the SequentialNetwork.

class pynever.networks.AcyclicNetwork(identifier, input_ids)[source]

Bases: NeuralNetwork

Concrete child of NeuralNetwork representing an acyclic NeuralNetwork. The computational graph of an AcyclicNetwork must correspond to a standard list.

add_node(node, parents=None, children=None)[source]

Procedure to add a node to the network only if it preserves the acyclic property

Parameters: