Graph-Structure-of-Neural-Networks

An unofficial re-implementation of Graph Structure of Neural Networks (Jiaxuan You · Kaiming He · Jure Leskovec · Saining Xie) ICML 2020 https://arxiv.org/abs/2007.06559

TODO

• Graph
  • Graph Generator
    • WS-flex
    • sample
    • analyse clustering coeffient and average length of graphs generated by this generator
  • Sample Graphs
  • remove isomorphic graphs by Hash ( Not implemented in paper, but I think it should be done)
  • Graph to Neural Network Converter
    • Mask out Linear
    • Mask out Conv2d
    • Mask out SeperateConvo2d
• Evaluate
  • Train/Eval on cifar10
  • Train/Eval on ImageNet
• Analyze

Graph

Graph Generator

All	WS-flex

1. Generate Graphs

python -m nx_ops.create create_all

2. Analyse Clustering coeffient and average length.

Use pandas and draw graphs by ipython, like:

import pandas
from matplotlib import pyplot as plt
df_ws = pd.read_csv("save/csv/ws-paper.csv.gz")
df_ws.dropna().sample(4000).plot.scatter("cluster_coefficient", "avg_path_length",alpha=0.25, color="yellow", ax=ax, label="WS")