retworkx for networkx users¶

This is an introductory guide for existing networkx users on how to use retworkx, how it differs from networkx, and key differences to keep in mind.

Some Key Differences¶

retworkx (as the name implies) was inspired by networkx and the goal of the project is to provide a similar level of functionality and utility to what networkx offers but with much faster performance. However, because of limitations in the boundary between rust and python, different design decisions, and other differences the libraries are not identical.

The biggest difference to keep in mind is networkx is a very dynamic in how it can be used. It allows you to treat a graph object associatively (like a python dictionary) and interact with the graph using the objects you’re putting on the graph. For example:

import networkx as nx

graph = nx.MultiDiGraph()
graph.add_node('my_node_a')
graph.add_node('my_node_b')
graph.add_edge('my_node_a', 'my_node_b')

While retworkx being written in Rust puts more constraints on how you interact with graph objects. With retworkx you can still attach any Python object on the a graph but each node and edge is assigned an integer index. That index must be used for accessing nodes and edges on the graph. In retworkx the above example would be something like:

import retworkx as rx

graph = rx.PyDiGraph()
node_a = graph.add_node('my_node_a')
node_b = graph.add_node('my_node_b')
graph.add_edge(node_a, node_b, None)

where node_a == 0 and node_b == 1. These node indices can be used with a graph object to access the objects set as the payload object via the python mapping protocol (not the sequence protocol because the indices are not guaranteed to be a sequence after nodes or edges are removed from a graph). Continuing from the above retworkx example:

assert 'my_node_a' == graph[node_a]
assert 'my_node_b' == graph[node_b]

The use of integer indexes for everything is normally the biggest difference that existing networkx users have to adapt to when migrating to retworkx.

Similarly when there are algorithm functions that operate on a node or edge data, callback functions are used in retworkx to return statically typed data from node or edge payloads to use for various algorithms. In networkx this is typically done using named attributes of nodes or edges (the typical example of a node or edge attribute named weight is used by default for functions that need a numerical weight).

For example, in networkx:

import networkx as nx

graph = nx.MultiDiGraph()
graph.add_edges_from([(0, 1, {'weight': 1}), (0, 2, {'weight': 2}),
                      (1, 3, {'weight': 2}), (3, 0, {'weight': 3})])
dist_matrix = nx.floyd_warshall_numpy(graph, weight='weight')

while in retworkx you would use:

import retworkx as rx

graph = rx.PyDiGraph()
graph.extend_from weighted_edge_list(
    [(0, 1, {'weight': 1}), (0, 2, {'weight': 2}),
     (1, 3, {'weight': 2}), (3, 0, {'weight': 3})])
dist_matrix = rx.digraph_floyd_warshall_numpy(
    graph, weight_fn=lambda edge: edge[weight])

or more concisely:

import retworkx as rx

graph = rx.PyDiGraph()
graph.extend_from weighted_edge_list(
    [(0, 1, 1), (0, 2, 2),
     (1, 3, 2), (3, 0, 3)])
dist_matrix = rx.digraph_floyd_warshall_numpy(graph,
                                              weight_fn=lambda edge: edge)

The other large difference to keep in mind is that many functions in retworkx are explicitly typed. This means that they either always return or accept either a PyDiGraph or a PyGraph but not both. Generally, functions prefixed with graph_* and digraph_* explicitly typed. Explicitly typed functions also indicate their type on the docstrings. This is different from networkx where everything is pretty much dynamically typed and you can pass a graph object to any function and it will work as expected (unless it isn’t supported and then it will raise an exception).

Graph Data and Attributes¶

Nodes¶

In networkx a node can be any hashable python object. That object is then used to access or refer to a node. Additionally, you can set optional attributes on a node. This is described in more detail below.

In retworkx any python object (hashable or not) can be used as a node, however nodes can only be accessed by an integer node index (which is returned by any function adding a node). There are no optional attributes for nodes. If this is required that is expected to be added to the node’s data payload.

Edges¶

Edges in networkx are accessible by the tuple of the nodes the edge is between. Edges only have optional attributes (as described below) and no other object payload.

In retworkx any python object can be an edge and have a unique integer index assigned to it, just like nodes. However, edges are in most functions/methods referenced by the tuple of the indices of the nodes the edge is between instead of the edge’s index.

Attributes¶

networkx has a concept of graph, node, and edge attributes in addition to the hashable object used for a node’s payload. Retworkx has no analogous concept. Instead, the payloads for nodes and edges are any python object (hashable or not). This enables you to build similar structures to the attributes concept, but also use alternative structures specific to your use case.

For example, something like:

import networkx as nx

graph = nx.Graph()
graph.add_node(1, time='5pm')
graph.add_nodes_from([3], time='2pm')
graph.nodes[1]['room'] = 714

can be accomplished by using a dict for node weights:

import retworkx as rx

graph = rx.PyGraph()
node_a = graph.add_node({'time': '5pm'})
node_b = graph.add_nodes_from([{'time': '2pm'}])
graph[node_a]['room'] = 714

Examining elements of a graph¶

networkx provides 4 attributes on graph objects nodes, edges, adj, and degree which act as set like views for the nodes, edges, neighbors, and degrees of nodes respectively. These properties provide a real time view into the different properties of the graphs and provide additional methods on those attributes for looking at graph properties in different ways.

retworkx doesn’t offer views, but instead provides different accessor methods that return copies of the analogous data. There are different functions/methods that offer different views on that data. For example, edge_list() is analogous to networkx’s edges view and weighted_edge_list() is equivalent to networkx’s edges(data=True).

Additionally, since everything in retworkx is integer indexed, to access node data the PyDiGraph and PyGraph classes implement the python mapping protocol so you can access node’s data using a node’s index.

API Equivalents¶

Class Constructors¶

networkx	retworkx	Notes
`Graph()`	`PyGraph(multigraph=False)`	Only in multigraph flag added in retworkx>= 0.8.0 prior releases always allow multiple edges
`DiGraph()`	`PyDiGraph(multigraph=False)`	Only in multigraph flag added in retworkx>= 0.8.0 prior releases always allow multiple edges
`MultiGraph()`	`PyGraph()`
`MultiDiGraph()`	`PyDiGraph()`

The other thing to note here is that retworkx does not allow initialization of a graph when the constructor is called. You will need to call an appropriate method of the object to add nodes or edges or use an alternative constructor method:

networkx

retworkx

Notes

Graph([(0, 1), (1, 0)])

graph = PyGraph()
graph.extend_from_edge_list([(0, 1), (1, 0)])

retworkx input must be a list of 2-tuples, while networkx can be an iterator

Graph([(0, 1, {'weight': 2}), (1, 0, {'weight': 1})])

graph = PyGraph()
graph.extend_from_edge_list([(0, 1, 2), (1, 0, 1)])

retworkx input must be a list of 3-tuples, while networkx can be an iterator

Graph(np.array([[0, 1, 1], [1, 0, 1], [1, 0, 1]]))

PyGraph.from_adjacency_matrix(np.array([[0, 1, 1], [1, 0, 1], [1, 0, 1]], dtype=np.float64))

retworkx from_adjacency_matrix() can only take a float dtype numpy array, you can use .astype(np.float64, copy=False) to adapt a non-float array.

Graph Modifiers¶

networkx	retworkx	Notes
`add_node()`	`add_node()`	retworkx returns a node index for the newly created node
`add_nodes_from`	`add_nodes_from()`	retworkx requires the input to be a list of objects and will return a list of node indices for the newly created nodes
`add_edge`	`add_edge()`	retworkx requires 3 parameters be used, the 2 node indices and the payload (networkx works with either 2 or 3)
`add_edges_from`	`add_edges_from()`, `add_edges_from_no_data()`, `extend_from_edge_list()`, `extend_from_weighted_edge_list()`	retworkx requires a list of either a 3 or 2 tuple (depending on whether weights/data are expected or not). The difference between the retworkx `extend_from` and `add_edges_from` methods are that the `extend_from*` will create new nodes with a weight/data payload of `None` if any node indices are missing.

(note the retworkx version links to the PyDiGraph version, but there are also equivalent PyGraph methods available)

Matrix Converter Functions¶

NetworkX has several functions for going back and forth between a NetworkX graph and matrices in other libraries. This includes to_numpy_matrix(), to_numpy_array(), to_numpy_recarray(), to_scipy_sparse_matrix(), to_pandas_adjacency(), and adjacency_matrix() (which is equivalent to to_scipy_sparse_matrix() and returns a scipy csr sparse matrix of the adjacency matrix).

However, in retworkx there is only a adjacency_matrix() function (and it’s per type variants digraph_adjacency_matrix() and graph_adjacency_matrix()) which will return a numpy array of the adjacency matrix (not a scipy csr sparse matrix like networkx’s function). This function is equivalent to networkx’s to_numpy_array() function.

This difference with retworkx is primarily because numpy exposes a public C interface which retworkx can interface with directly, while the other libraries and types only expose Python APIs.

Visualization Functions¶

NetworkX provides a native drawer with a matplotlib drawer (the networkx_drawer* functions) and then functions to interface with pygraphviz and pydot to enable visualization with graphviz via those libraries (in addition to functions to serialize graphs in formats other graph visualization tools can use). NetworkX also provides several functions layout functions for generating different layouts that can be used for visualizing the graph.

retworkx has drawer functions with 2 visualization backends, matplotlib (mpl_draw()) and graphviz (graphviz_draw()). Unlike networkx the graphviz_draw() will handle calling graphviz and generate an image file. For layout functions retworkx has a similar variety of Layout Functions, however it should be noted that retworkx’s functions are strictly 2 dimensional. The also return a Pos2DMapping custom return type which acts as read-only dictionary (which is different from networkx which returns a normal dictionary that can be modified).

Matplotlib Drawers¶

The retwork function mpl_draw() function is basically equivalent to the networkx function draw_networkx (it was actually originally forked from the networkx drawer). However, there are some key differences to keep in mind between the networkx and retworkx matplotlib drawer.

networkx.draw_networkx and retworkx.mpl_draw differences:

networkx	retworkx	Notes
`nodelist`	`node_list`
`edgelist`	`edge_list`
`arrowsize`	`arrow_size`
`labels`	`labels`	For `networkx_drawer` `labels` is a dict of nodes to their label, while retworkx’s `mpl_drawer` `labels` is a callback function that will be passed a node’s data payload and expected to return the node’s label
`networkx.draw_networkx_edge_labels()`	`edge_labels`	NetworkX’s `networkx_drawer` doesn’t have an option for edge labels and instead adding labels is only exposed via a separate function `draw_networkx_edge_labels()` which requires the `pos` dictionary from the original visualization to be used. retworkx’s `edge_labels` kwarg takes a callback function that will be passed an edge’s data payload and expected to return the label.

Converting from a networkx graph¶

If you’re using a function or an external library that is already generating a networkx graph then you can use retworkx.networkx_converter() to convert that networkx Graph object into an equivalent retworkx PyGraph or PyDiGraph object. Note that networkx is not a dependency for retworkx and you are responsible for installing networkx to use this function. Accordingly, there is not equivalent function provided to convert the reverse direction (because doing so would add an unwanted dependency on networkx, even an optional one) but writing such a function is straightforward, for example:

import networkx as nx
import retworkx as rx


def convert_retworkx_to_networkx(graph):
    """Convert a retworkx PyGraph or PyDiGraph to a networkx graph."""
    edge_list = [(
        graph[x[0]], graph[x[1]],
        {'weight': x[2]}) for x in graph.weighted_edge_list()]

    if isinstance(graph, rx.PyGraph):
        if graph.multigraph:
            return nx.MultiGraph(edge_list)
        else:
            return nx.Graph(edge_list)
    else:
        if graph.multigraph:
            return nx.MultiDiGraph(edge_list)
        else:
            return nx.DiGraph(edge_list)

Functionality Gaps¶

networkx is a mature library that has a wide user base and extensive feature set, while retworkx, by comparison, is a much younger library and is missing a lot of the features that networkx offers. If you encounter a feature that networkx offers which is missing from retworkx that you would like to use please open an “Enhancement request” issue at: https://github.com/Qiskit/retworkx/issues/new/choose Once an issue is opened we can prioritize working on adding an equivalent feature to retworkx.