├── .gitignore
├── README.md
├── figures
    ├── all_path_cluster.png
    └── ws-flex_path_cluster.png
├── nx_ops
    ├── __init__.py
    ├── create.py
    ├── sample.py
    └── ws_flex.py
└── requirements.txt


/.gitignore:
--------------------------------------------------------------------------------
  1 | save
  2 | *.onnx
  3 | .*/
  4 | 
  5 | # Created by https://www.gitignore.io/api/vim,macos<Plug>PeepOpenython,visualstudiocode
  6 | # Edit at https://www.gitignore.io/?templates=vim,macos<Plug>PeepOpenython,visualstudiocode
  7 | 
  8 | ### Compressed ###
  9 | *.7z
 10 | *.deb
 11 | *.gz
 12 | *.pkg
 13 | *.rar
 14 | *.rpm
 15 | *.sit
 16 | *.sitx
 17 | *.tar
 18 | *.zip
 19 | *.zipx
 20 | *.tgz
 21 | 
 22 | ### macOS ###
 23 | # General
 24 | .DS_Store
 25 | .AppleDouble
 26 | .LSOverride
 27 | 
 28 | # Icon must end with two \r
 29 | Icon
 30 | 
 31 | # Thumbnails
 32 | ._*
 33 | 
 34 | # Files that might appear in the root of a volume
 35 | .DocumentRevisions-V100
 36 | .fseventsd
 37 | .Spotlight-V100
 38 | .TemporaryItems
 39 | .Trashes
 40 | .VolumeIcon.icns
 41 | .com.apple.timemachine.donotpresent
 42 | 
 43 | # Directories potentially created on remote AFP share
 44 | .AppleDB
 45 | .AppleDesktop
 46 | Network Trash Folder
 47 | Temporary Items
 48 | .apdisk
 49 | 
 50 | ### Python ###
 51 | # Byte-compiled / optimized / DLL files
 52 | __pycache__/
 53 | *.py[cod]
 54 | *$py.class
 55 | 
 56 | # C extensions
 57 | *.so
 58 | 
 59 | # Distribution / packaging
 60 | .Python
 61 | build/
 62 | develop-eggs/
 63 | dist/
 64 | downloads/
 65 | eggs/
 66 | .eggs/
 67 | lib/
 68 | lib64/
 69 | parts/
 70 | sdist/
 71 | var/
 72 | wheels/
 73 | pip-wheel-metadata/
 74 | share/python-wheels/
 75 | *.egg-info/
 76 | .installed.cfg
 77 | *.egg
 78 | MANIFEST
 79 | 
 80 | # PyInstaller
 81 | #  Usually these files are written by a python script from a template
 82 | #  before PyInstaller builds the exe, so as to inject date/other infos into it.
 83 | *.manifest
 84 | *.spec
 85 | 
 86 | # Installer logs
 87 | pip-log.txt
 88 | pip-delete-this-directory.txt
 89 | 
 90 | # Unit test / coverage reports
 91 | htmlcov/
 92 | .tox/
 93 | .nox/
 94 | .coverage
 95 | .coverage.*
 96 | .cache
 97 | nosetests.xml
 98 | coverage.xml
 99 | *.cover
100 | .hypothesis/
101 | .pytest_cache/
102 | 
103 | # Translations
104 | *.mo
105 | *.pot
106 | 
107 | # Scrapy stuff:
108 | .scrapy
109 | 
110 | # Sphinx documentation
111 | docs/_build/
112 | 
113 | # PyBuilder
114 | target/
115 | 
116 | # pyenv
117 | .python-version
118 | 
119 | # pipenv
120 | #   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
121 | #   However, in case of collaboration, if having platform-specific dependencies or dependencies
122 | #   having no cross-platform support, pipenv may install dependencies that don't work, or not
123 | #   install all needed dependencies.
124 | #Pipfile.lock
125 | 
126 | # celery beat schedule file
127 | celerybeat-schedule
128 | 
129 | # SageMath parsed files
130 | *.sage.py
131 | 
132 | # Spyder project settings
133 | .spyderproject
134 | .spyproject
135 | 
136 | # Rope project settings
137 | .ropeproject
138 | 
139 | # Mr Developer
140 | .mr.developer.cfg
141 | .project
142 | .pydevproject
143 | 
144 | # mkdocs documentation
145 | /site
146 | 
147 | # mypy
148 | .mypy_cache/
149 | .dmypy.json
150 | dmypy.json
151 | 
152 | # Pyre type checker
153 | .pyre/
154 | 
155 | ### Vim ###
156 | # Swap
157 | [._]*.s[a-v][a-z]
158 | [._]*.sw[a-p]
159 | [._]s[a-rt-v][a-z]
160 | [._]ss[a-gi-z]
161 | [._]sw[a-p]
162 | 
163 | # Session
164 | Session.vim
165 | Sessionx.vim
166 | 
167 | # Temporary
168 | .netrwhist
169 | *~
170 | 
171 | # Auto-generated tag files
172 | tags
173 | 
174 | # Persistent undo
175 | [._]*.un~
176 | 
177 | # Coc configuration directory
178 | .vim
179 | 
180 | ### VisualStudioCode ###
181 | .vscode/*
182 | !.vscode/settings.json
183 | !.vscode/tasks.json
184 | !.vscode/launch.json
185 | !.vscode/extensions.json
186 | 
187 | ### VisualStudioCode Patch ###
188 | # Ignore all local history of files
189 | .history
190 | 
191 | # End of https://www.gitignore.io/api/vim,macos<Plug>PeepOpenython,visualstudiocode
192 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # Graph-Structure-of-Neural-Networks
 2 | An unofficial re-implementation of Graph Structure of Neural Networks (Jiaxuan You · Kaiming He · Jure Leskovec · Saining Xie) ICML 2020
 3 | https://arxiv.org/abs/2007.06559
 4 | 
 5 | # TODO
 6 | - [ ] Graph
 7 |     - [x] Graph Generator
 8 |       - [x] WS-flex
 9 |       - [x] sample
10 |       - [x] analyse **clustering coeffient** and **average length** of graphs generated by this generator
11 |     - [x] Sample Graphs
12 |     - [ ] remove isomorphic graphs by Hash ( Not implemented in paper, but I think it should be done)
13 |     - [ ] Graph to Neural Network Converter
14 |         - [ ] Mask out Linear
15 |         - [ ] Mask out Conv2d
16 |         - [ ] Mask out SeperateConvo2d
17 | - [ ] Evaluate
18 |   - [ ] Train/Eval on cifar10
19 |   - [ ] Train/Eval on ImageNet
20 | - [ ] Analyze
21 | 
22 | ## Graph
23 | ### Graph Generator
24 | All | WS-flex
25 | :---:|:---:
26 | ![](./figures/all_path_cluster.png) | ![](./figures/ws-flex_path_cluster.png) 
27 | 1. Generate Graphs
28 | ```bash
29 | python -m nx_ops.create create_all
30 | ```
31 | 2. Analyse Clustering coeffient and average length.
32 | 
33 | Use pandas and draw graphs by ipython, like:
34 | ```python
35 | import pandas
36 | from matplotlib import pyplot as plt
37 | df_ws = pd.read_csv("save/csv/ws-paper.csv.gz")
38 | df_ws.dropna().sample(4000).plot.scatter("cluster_coefficient", "avg_path_length",alpha=0.25, color="yellow", ax=ax, label="WS")
39 | ```
40 | 


--------------------------------------------------------------------------------
/figures/all_path_cluster.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/CuriousCat-7/Graph-Structure-of-Neural-Networks/342bc4be2de730278e99c0a43cdf3708b0cbd025/figures/all_path_cluster.png


--------------------------------------------------------------------------------
/figures/ws-flex_path_cluster.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/CuriousCat-7/Graph-Structure-of-Neural-Networks/342bc4be2de730278e99c0a43cdf3708b0cbd025/figures/ws-flex_path_cluster.png


--------------------------------------------------------------------------------
/nx_ops/__init__.py:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/CuriousCat-7/Graph-Structure-of-Neural-Networks/342bc4be2de730278e99c0a43cdf3708b0cbd025/nx_ops/__init__.py


--------------------------------------------------------------------------------
/nx_ops/create.py:
--------------------------------------------------------------------------------
 1 | import os
 2 | from .sample import *
 3 | from loguru import logger
 4 | from glob import glob
 5 | import json
 6 | 
 7 | 
 8 | def _dump(f, path, overwrite):
 9 |     if os.path.exists(path) and not overwrite:
10 |         logger.warning("{} already exists", path)
11 |         return
12 |     logger.info(path)
13 |     f().to_csv(path, compression="gzip")
14 | 
15 | 
16 | def create_all(save_root="save/csv/", overwrite=False):
17 |     os.makedirs(save_root, exist_ok=True)
18 |     _dump(sample_harary, os.path.join(save_root, "harary-paper.csv.gz"), overwrite)
19 |     _dump(sample_ring, os.path.join(save_root, "ring-paper.csv.gz"), overwrite)
20 |     _dump(sample_er, os.path.join(save_root, "er-paper.csv.gz"), overwrite)
21 |     _dump(sample_ba, os.path.join(save_root, "ba-paper.csv.gz"), overwrite)
22 |     _dump(sample_ws, os.path.join(save_root, "ws-paper.csv.gz"), overwrite)
23 |     _dump(sample_ws_flex, os.path.join(save_root, "ws_flex-paper.csv.gz"), overwrite)
24 | 
25 | 
26 | def calculate_avg_cluster_path(save_root="save/csv/"):
27 |     csv_paths = glob(f"{save_root}/*.csv.gz")
28 |     for csv_path in csv_paths:
29 |         df = pd.read_csv(csv_path)
30 |         ccs = []
31 |         apls = []
32 |         logger.info(csv_path)
33 |         for i in tqdm(range(len(df))):
34 |             g = nx.node_link_graph(eval(df.iloc[i].graph))
35 |             d = get_avg_cluater_path(g)
36 |             ccs.append(d["cluster_coefficient"])
37 |             apls.append(d["avg_path_length"])
38 | 
39 |         df["cluster_coefficient"] = ccs
40 |         df["avg_path_length"] = apls
41 |         df.to_csv(csv_path, compression="gzip")
42 | 
43 | 
44 | if __name__ == "__main__":
45 |     fire.Fire()
46 | 


--------------------------------------------------------------------------------
/nx_ops/sample.py:
--------------------------------------------------------------------------------
  1 | import networkx as nx
  2 | from networkx.generators.harary_graph import hnm_harary_graph
  3 | from .ws_flex import watts_strogatz_flexible_graph
  4 | import numpy as np
  5 | import pandas as pd
  6 | from tqdm import tqdm
  7 | import fire
  8 | from loguru import logger
  9 | 
 10 | 
 11 | def get_avg_cluater_path(g: nx.Graph) -> dict:
 12 |     try:
 13 |         cc = nx.average_clustering(g)
 14 |     except nx.NetworkXError as e:
 15 |         logger.warning(e)
 16 |         cc = np.nan
 17 |     try:
 18 |         apl = nx.average_shortest_path_length(g)
 19 |     except nx.NetworkXError as e:
 20 |         logger.warning(e)
 21 |         apl = np.nan
 22 |     return dict(
 23 |         cluster_coefficient=cc,
 24 |         avg_path_length=apl,
 25 |     )
 26 | 
 27 | 
 28 | def sample_ws(n=64, k_max=62, k_min=8, p_num=300, seed_num=30) -> pd.DataFrame:
 29 |     rows = []
 30 |     assert k_min < k_max <= n
 31 |     pbar = tqdm(total=int((k_max-k_min)*p_num*seed_num))
 32 |     for k in np.arange(k_min, k_max):
 33 |         for p in np.linspace(0, 1, p_num)**2:
 34 |             for seed in range(seed_num):
 35 |                 g = nx.generators.watts_strogatz_graph(n, k, p, seed=seed)
 36 |                 rows.append(dict(
 37 |                     method="ws", n=n, k=k, p=p,
 38 |                     graph=nx.node_link_data(g),
 39 |                 ))
 40 |                 pbar.update()
 41 |     return pd.DataFrame(rows)
 42 | 
 43 | 
 44 | def sample_ba(n=64, m_max=30, m_min=4, seed_num=300):
 45 |     rows = []
 46 |     pbar = tqdm(total=int(m_max-m_min)*seed_num)
 47 |     for m in np.arange(m_min, m_max):
 48 |         for seed in range(seed_num):
 49 |             g = nx.generators.barabasi_albert_graph(n, m, seed=seed)
 50 |             rows.append(dict(
 51 |                 method="ba", n=n, m=m,
 52 |                 graph=nx.node_link_data(g),
 53 |             ))
 54 |             pbar.update()
 55 |     return pd.DataFrame(rows)
 56 | 
 57 | 
 58 | def sample_ws_flex(n=64, k_max=62, k_min=8, p_num=300, seed_num=30) -> pd.DataFrame:
 59 |     rows = []
 60 |     assert k_min < k_max <= n
 61 |     pbar = tqdm(total=int((k_max-k_min)*p_num*seed_num))
 62 |     for k in np.arange(k_min, k_max):
 63 |         for p in np.linspace(0, 1, p_num)**2:
 64 |             for seed in range(seed_num):
 65 |                 g = watts_strogatz_flexible_graph(n, k, p, seed=seed)
 66 |                 rows.append(dict(
 67 |                     method="ws-flex", n=n, k=k, p=p,
 68 |                     graph=nx.node_link_data(g),
 69 |                 ))
 70 |                 pbar.update()
 71 |     return pd.DataFrame(rows)
 72 | 
 73 | 
 74 | def sample_er(n=64, m_max=int(64*63/2), m_min=int(64*4), seed_num=30):
 75 |     rows = []
 76 |     pbar = tqdm(total=int(m_max-m_min)*seed_num)
 77 |     e = n*(n-1)
 78 |     for m in np.arange(m_min, m_max):
 79 |         for seed in range(seed_num):
 80 |             g = nx.generators.erdos_renyi_graph(n, m/e)
 81 |             rows.append(dict(
 82 |                 method="er", n=n, m=m,
 83 |                 graph=nx.node_link_data(g),
 84 |             ))
 85 |             pbar.update()
 86 |     return pd.DataFrame(rows)
 87 | 
 88 | 
 89 | def sample_ring(n=64, k_max=62, k_min=8):
 90 |     rows = []
 91 |     pbar = tqdm(total=k_max-k_min)
 92 |     for k in np.arange(k_min, k_max):
 93 |         g = nx.generators.watts_strogatz_graph(n, k, 0.0)
 94 |         rows.append(dict(
 95 |             method="ring", n=n, k=k,
 96 |             graph=nx.node_link_data(g),
 97 |         ))
 98 |         pbar.update()
 99 |     return pd.DataFrame(rows)
100 | 
101 | 
102 | def sample_harary(n=64, m_max=int(64*63/2), m_min=int(64*4)):
103 |     rows = []
104 |     pbar = tqdm(total=m_max-m_min)
105 |     for m in np.arange(m_min, m_max):
106 |         g = hnm_harary_graph(n, m)
107 |         rows.append(dict(
108 |             method="harary", n=n, m=m,
109 |             graph=nx.node_link_data(g),
110 |         ))
111 |         pbar.update()
112 |     return pd.DataFrame(rows)
113 | 
114 | 
115 | if __name__ == "__main__":
116 |     fire.Fire()
117 | 


--------------------------------------------------------------------------------
/nx_ops/ws_flex.py:
--------------------------------------------------------------------------------
 1 | import networkx as nx
 2 | from networkx.utils import py_random_state
 3 | 
 4 | 
 5 | @py_random_state(3)
 6 | def watts_strogatz_flexible_graph(n, k, p, seed=None):
 7 |     """Returns a Watts–Strogatz flexible small-world graph defined by
 8 |     https://arxiv.org/abs/2007.06559
 9 | 
10 |     Parameters
11 |     ----------
12 |     n : int
13 |         The number of nodes
14 |     k : float or int
15 |         Each node is joined with its `k` nearest neighbors in a ring
16 |         topology.
17 |     p : float
18 |         The probability of rewiring each edge
19 |     seed : integer, random_state, or None (default)
20 |         Indicator of random number generation state.
21 |         See :ref:`Randomness<randomness>`.
22 | 
23 |     See Also
24 |     --------
25 |     watts_strogatz_graph()
26 | 
27 |     References
28 |     ---------
29 |     .. [1] You, Jiaxuan et al. “Graph Structure of Neural Networks.” (2020).
30 |     """
31 |     if k > n:
32 |         raise nx.NetworkXError("k>n, choose smaller k or larger n")
33 | 
34 |     # If k == n, the graph is complete not Watts-Strogatz
35 |     if k == n:
36 |         return nx.complete_graph(n)
37 | 
38 |     G = nx.Graph()
39 |     nodes = list(range(n))  # nodes are labeled 0 to n-1
40 |     # connect each node to k/2 neighbors
41 |     for j in range(1, int(k//2) + 1):
42 |         targets = nodes[j:] + nodes[0:j]  # first j nodes are now last in list
43 |         G.add_edges_from(zip(nodes, targets))
44 | 
45 |     # picks e mod n nodes and connects each node to
46 |     # one closest neighboring node
47 |     potential_nodes = list(range(n))
48 |     seed.shuffle(potential_nodes)
49 |     total_add_edges = int(n*k//2) % n
50 |     add_edge_num = 0
51 |     for j in potential_nodes:
52 |         if add_edge_num == total_add_edges:
53 |             break
54 |         i = (j + int(k//2) + 1) % n
55 |         if G.has_edge(j, i):
56 |             continue  # when n = 4, k = 3, additional edges may crash
57 |         else:
58 |             G.add_edge(j, i)
59 |             add_edge_num += 1
60 | 
61 |     # rewire edges from each node
62 |     # loop over all nodes in order (label) and neighbors in order (distance)
63 |     # no self loops or multiple edges allowed
64 |     for j in range(1, int(k//2) + 1):  # outer loop is neighbors
65 |         targets = nodes[j:] + nodes[0:j]  # first j nodes are now last in list
66 |         # inner loop in node order
67 |         for u, v in zip(nodes, targets):
68 |             if seed.random() < p:
69 |                 w = seed.choice(nodes)
70 |                 # Enforce no self-loops or multiple edges
71 |                 while w == u or G.has_edge(u, w):
72 |                     w = seed.choice(nodes)
73 |                     if G.degree(u) >= n - 1:
74 |                         break  # skip this rewiring
75 |                 else:
76 |                     G.remove_edge(u, v)
77 |                     G.add_edge(u, w)
78 |     return G
79 | 


--------------------------------------------------------------------------------
/requirements.txt:
--------------------------------------------------------------------------------
1 | networkx==2.4
2 | pandas
3 | fire
4 | loguru
5 | numpy
6 | 


--------------------------------------------------------------------------------