Comparing Traditional and GAN-based Approaches for the Synthesis of Wide Area Network Topologies

This repository contains the source code and (generated) datasets used in the CNSM2022 paper Comparing Traditional and GAN-based Approaches for the Synthesis of Wide Area Network Topologies. Below is an overview of the contained folders and files. [More info below to come :)]

adj_matrices

Adjancecy matrices of the networks extracted from the OOS. See code for more info.

plot_data and plots

As the name suggests, the plots and corresponding that from the paper, for all generators (2K, BA, WS, ER, for GAN: RGB and BW). For the hierarchical case the endings "_2", "_3", "_4" illustrate the number of clusters.

synthetic_networks

Contains archives that contain the generated networks for all of the investigated generators for the four real topologies, including one hierarchical use case.

For the traditional generations (2K, ER, BA, WS), it contains 1000 files each in the following format for the naive approach:

synth_sample_<generator>_ i.pkl, which are NetworkX graphs (nx.Graph()).

For the GAN, the contained data is more complex. First, there are files named image_at_epoch_1000_s_(BW|RGB).pdf, which showcase samples of the generated networks for each of the then seeds s, just for some visualization, other than that, they have no further use.

Furthermore, for the GAN, there are files named sample_at_epoch_1000_s_i_(BW|RGB).pkl, which are the raw files that the GAN produced with seed s. These are NumPy Arrays (np.array()) with dimensions n x n x 1 for BW and n x n x 3 for RGB.

Next and last, there are files named synth_sample_i_s_weightssampled?_(BW|RGB).pkl. These are the postprocessed samples for the GAN, for seed s and weightssampled? is a boolean (so, 0 or 1), which specifies if the weights were sampled onto the graph. Like for the traditional generators, these are NetworkX graphs (nx.Graph()).

For the single hierarchical example, it contains multiple subfolders for local and global views.

code

Contains the source code, including some helper functions for transforming the .ned files from the OOS into the matrices of the adj_matrices folder.

wangan.py is the main functionality that trains the GANs and saves in the end the generated samples.

WANGAN-simple.py and WANGAN-hierarchical call the previous script to either perform the naive approach or the hierarchical approach.

^{Note that since we trained the GANs on GPU, results may differ even though seeds are set (e.g., due to parallelization or non-deterministic GPU ops), especially for the RGB-based GANs; but general trends should remain the same. For TensorFlow >= 2.9.0 there are now options via tf.config.experimental.enable_op_determinism, so you might want to check this out, since research on our work started prior to that release.}

There are four eval scripts, two for the naive approach for GAN-based and traditional approaches, and two for the hierarchical approach. Note that the traditionally generated network are created in an ad-hoc manner by calling the traditional.py script due to their simplicity here (can be controlled via save and load parameters, if this is not wanted).

Lastly, there is the read_zoo_revised.py script, which takes the .ned files from the OOS (https://github.com/lsinfo3/OpenFlowOMNeTSuite/tree/master/openflow/networks) and transforms them into the adjacency matrices, weighted and unweighted ones.