-
Notifications
You must be signed in to change notification settings - Fork 0
/
nets.py
355 lines (281 loc) · 14.6 KB
/
nets.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
import time
import torch
import torch.nn.functional as F
from torch.nn import Linear
from torch_geometric.nn import DenseGraphConv
from torch_geometric.utils import to_dense_batch, to_dense_adj
from layers.CT_layer import dense_CT_rewiring
from layers.MinCut_Layer import dense_mincut_pool
from layers.GAP_layer import dense_mincut_rewiring
class GAPNet(torch.nn.Module):
def __init__(self, in_channels, out_channels, hidden_channels=32, derivative=None, EPS=1e-15, device=None):
super(GAPNet, self).__init__()
self.device = device
self.derivative = derivative
self.EPS = EPS
# GCN Layer - MLP - Dense GCN Layer
#self.conv1 = GCNConv(in_channels, hidden_channels)
self.conv1 = DenseGraphConv(hidden_channels, hidden_channels)
self.conv2 = DenseGraphConv(hidden_channels, hidden_channels)
num_of_centers2 = 16 # k2
#num_of_centers2 = 10 # k2
#num_of_centers2 = 5 # k2
num_of_centers1 = 2 # k1 #Fiedler vector
# The degree of the node belonging to any of the centers
self.pool1 = Linear(hidden_channels, num_of_centers1)
self.pool2 = Linear(hidden_channels, num_of_centers2)
# MLPs towards out
self.lin1 = Linear(in_channels, hidden_channels)
self.lin2 = Linear(hidden_channels, hidden_channels)
self.lin3 = Linear(hidden_channels, out_channels)
# Input: Batch of 20 graphs, each node F=3 features
# N1 + N2 + ... + N2 = 661
# TSNE here?
def forward(self, x, edge_index, batch): # x torch.Size([N, N]), data.batch torch.Size([661])
# Make all adjacencies of size NxN
adj = to_dense_adj(edge_index, batch) # adj torch.Size(B, N, N])
#print("adj_size", adj.size())
#print("adj",adj)
# Make all x_i of size N=MAX(N1,...,N20), e.g. N=40:
#print("x size", x.size())
x, mask = to_dense_batch(x, batch) # x torch.Size([20, N, 32]) ; mask torch.Size([20, N]) batch_size=20
#print("x size", x.size())
x = self.lin1(x)
# First mincut pool for computing Fiedler adn rewire
s1 = self.pool1(x)
#s1 = torch.variable()#s1 torch.Size([20, N, k1=2)
#s1 = Variable(torch.randn(D_in, H).type(float16), requires_grad=True)
#print("s 1st pool",s1)
#print("s 1st pool size", s1.size())
if torch.isnan(adj).any():
print("adj nan")
if torch.isnan(x).any():
print("x nan")
# REWIRING
#start = time.time()
adj, mincut_loss1, ortho_loss1 = dense_mincut_rewiring(x, adj, s1, mask, derivative = self.derivative, EPS=self.EPS, device=self.device) # out: x torch.Size([20, N, F'=32]), adj torch.Size([20, N, N])
#print('\t\tdense_mincut_rewiring: {:.6f}s'.format(time.time()- start))
#print("x",x)
#print("adj",adj)
#print("x and adj sizes", x.size(), adj.size())
#adj = torch.softmax(adj, dim=-1)
#print("adj softmaxed", adj)
# CONV1: Now on x and rewired adj:
x = self.conv1(x, adj) #out: x torch.Size([20, N, F'=32])
#print("x_1 ", x)
#print("x_1 size", x.size())
# MLP of k=16 outputs s
#print("adj_size", adj.size())
s2 = self.pool2(x) # s torch.Size([20, N, k])
#print("s 2nd pool", s2)
#print("s 2nd pool size", s2.size())
#adj = torch.softmax(adj, dim=-1)
# MINCUT_POOL
# Call to dense_cut_mincut_pool to get coarsened x, adj and the losses: k=16
#x, adj, mincut_loss1, ortho_loss1 = dense_mincut_rewiring(x, adj, s1, mask) # x torch.Size([20, k=16, F'=32]), adj torch.Size([20, k2=16, k2=16])
x, adj, mincut_loss2, ortho_loss2 = dense_mincut_pool(x, adj, s2, mask, EPS=self.EPS) # out x torch.Size([20, k=16, F'=32]), adj torch.Size([20, k2=16, k2=16])
#print("lossses2",mincut_loss2, ortho_loss2)
#print("mincut pool x", x)
#print("mincut pool adj", adj)
#print("mincut pool x size", x.size())
#print("mincut pool adj size", adj.size()) # Some nan in adjacency: maybe comming from the rewiring-> dissapear after clipping
# CONV2: Now on coarsened x and adj:
x = self.conv2(x, adj) #out x torch.Size([20, 16, 32])
#print("x_2", x)
#print("x_2 size", x.size())
# Readout for each of the 20 graphs
#x = x.mean(dim=1) # x torch.Size([20, 32])
x = x.sum(dim=1) # x torch.Size([20, 32])
#print("mean x_2 size", x.size())
# Final MLP for graph classification: hidden channels = 32
x = F.relu(self.lin2(x)) # x torch.Size([20, 32])
#print("final x1 size", x.size())
x = self.lin3(x) #x torch.Size([20, 2])
#print("final x2 size", x.size())
#print("losses: ", mincut_loss1, mincut_loss2, ortho_loss2, mincut_loss2)
mincut_loss = mincut_loss1 + mincut_loss2
ortho_loss = ortho_loss1 + ortho_loss2
#print("x", x)
return F.log_softmax(x, dim=-1), mincut_loss, ortho_loss
class CTNet(torch.nn.Module):
def __init__(self, in_channels, out_channels, k_centers, hidden_channels=32, EPS=1e-15):
super(CTNet, self).__init__()
self.EPS=EPS
# GCN Layer - MLP - Dense GCN Layer
#self.conv1 = GCNConv(in_channels, hidden_channels)
self.conv1 = DenseGraphConv(hidden_channels, hidden_channels)
self.conv2 = DenseGraphConv(hidden_channels, hidden_channels)
# The degree of the node belonging to any of the centers
num_of_centers1 = k_centers # k1 #order of number of nodes
self.pool1 = Linear(hidden_channels, num_of_centers1)
num_of_centers2 = 16 # k2 #mincut
self.pool2 = Linear(hidden_channels, num_of_centers2)
# MLPs towards out
self.lin1 = Linear(in_channels, hidden_channels)
self.lin2 = Linear(hidden_channels, hidden_channels)
self.lin3 = Linear(hidden_channels, out_channels)
def forward(self, x, edge_index, batch): # x torch.Size([N, N]), data.batch torch.Size([661])
# Make all adjacencies of size NxN
adj = to_dense_adj(edge_index, batch) # adj torch.Size(B, N, N])
#print("adj_size", adj.size())
#print("adj",adj)
# Make all x_i of size N=MAX(N1,...,N20), e.g. N=40:
#print("x size", x.size())
x, mask = to_dense_batch(x, batch) # x torch.Size([20, N, 32]) ; mask torch.Size([20, N]) batch_size=20
#print("x size", x.size())
x = self.lin1(x)
# First mincut pool for computing Fiedler adn rewire
s1 = self.pool1(x)
#s1 = torch.variable()#s1 torch.Size([20, N, k1=2)
#s1 = Variable(torch.randn(D_in, H).type(float16), requires_grad=True)
#print("s 1st pool",s1)
#print("s 1st pool size", s1.size())
if torch.isnan(adj).any():
print("adj nan")
if torch.isnan(x).any():
print("x nan")
# CT REWIRING
adj, CT_loss, ortho_loss1 = dense_CT_rewiring(x, adj, s1, mask, EPS = self.EPS) # out: x torch.Size([20, N, F'=32]), adj torch.Size([20, N, N])
#print("CT_loss, ortho_loss1", CT_loss, ortho_loss1)
#print("x",x)
#print("adj",adj)
#print("x and adj sizes", x.size(), adj.size())
#adj = torch.softmax(adj, dim=-1)
#print("adj softmaxed", adj)
# CONV1: Now on x and rewired adj:
x = self.conv1(x, adj) #out: x torch.Size([20, N, F'=32])
#print("x_1 ", x)
#print("x_1 size", x.size())
# MLP of k=16 outputs s
#print("adj_size", adj.size())
s2 = self.pool2(x) # s torch.Size([20, N, k])
#print("s 2nd pool", s2)
#print("s 2nd pool size", s2.size())
#adj = torch.softmax(adj, dim=-1)
# MINCUT_POOL
# Call to dense_cut_mincut_pool to get coarsened x, adj and the losses: k=16
#x, adj, mincut_loss1, ortho_loss1 = dense_mincut_rewiring(x, adj, s1, mask) # x torch.Size([20, k=16, F'=32]), adj torch.Size([20, k2=16, k2=16])
x, adj, mincut_loss2, ortho_loss2 = dense_mincut_pool(x, adj, s2, mask, EPS=self.EPS) # out x torch.Size([20, k=16, F'=32]), adj torch.Size([20, k2=16, k2=16])
#print("lossses2",mincut_loss2, ortho_loss2)
#print("mincut pool x", x)
#print("mincut pool adj", adj)
#print("mincut pool x size", x.size())
#print("mincut pool adj size", adj.size()) # Some nan in adjacency: maybe comming from the rewiring-> dissapear after clipping
# CONV2: Now on coarsened x and adj:
x = self.conv2(x, adj) #out x torch.Size([20, 16, 32])
#print("x_2", x)
#print("x_2 size", x.size())
# Readout for each of the 20 graphs
#x = x.mean(dim=1) # x torch.Size([20, 32])
x = x.sum(dim=1) # x torch.Size([20, 32])
#print("mean x_2 size", x.size())
# Final MLP for graph classification: hidden channels = 32
x = F.relu(self.lin2(x)) # x torch.Size([20, 32])
#print("final x1 size", x.size())
x = self.lin3(x) #x torch.Size([20, 2])
#print("final x2 size", x.size())
CT_loss = CT_loss + ortho_loss1
mincut_loss = mincut_loss2 + ortho_loss2
#print("x", x)
return F.log_softmax(x, dim=-1), CT_loss, mincut_loss
class MinCutNet(torch.nn.Module):
def __init__(self, in_channels, out_channels, hidden_channels=32, EPS=1e-15):
super(MinCutNet, self).__init__()
self.EPS=EPS
# GCN Layer - MLP - Dense GCN Layer
self.conv1 = DenseGraphConv(hidden_channels, hidden_channels)
self.conv2 = DenseGraphConv(hidden_channels, hidden_channels)
# The degree of the node belonging to any of the centers
num_of_centers2 = 16 # k2 #mincut
self.pool2 = Linear(hidden_channels, num_of_centers2)
# MLPs towards out
self.lin1 = Linear(in_channels, hidden_channels)
self.lin2 = Linear(hidden_channels, hidden_channels)
self.lin3 = Linear(hidden_channels, out_channels)
def forward(self, x, edge_index, batch): # x torch.Size([N, N]), data.batch torch.Size([661])
# Make all adjacencies of size NxN
adj = to_dense_adj(edge_index, batch) # adj torch.Size(B, N, N])
# Make all x_i of size N=MAX(N1,...,N20), e.g. N=40:
x, mask = to_dense_batch(x, batch) # x torch.Size([20, N, 32]) ; mask torch.Size([20, N]) batch_size=20
x = self.lin1(x)
if torch.isnan(adj).any():
print("adj nan")
if torch.isnan(x).any():
print("x nan")
# CONV1: Now on x and rewired adj:
x = self.conv1(x, adj) #out: x torch.Size([20, N, F'=32])
# MLP of k=16 outputs s
s2 = self.pool2(x) # s torch.Size([20, N, k])
# MINCUT_POOL
# Call to dense_cut_mincut_pool to get coarsened x, adj and the losses: k=16
x, adj, mincut_loss2, ortho_loss2 = dense_mincut_pool(x, adj, s2, mask, EPS=self.EPS) # out x torch.Size([20, k=16, F'=32]), adj torch.Size([20, k2=16, k2=16])
# CONV2: Now on coarsened x and adj:
x = self.conv2(x, adj) #out x torch.Size([20, 16, 32])
# Readout for each of the 20 graphs
#x = x.mean(dim=1) # x torch.Size([20, 32])
x = x.sum(dim=1) # x torch.Size([20, 32])
# Final MLP for graph classification: hidden channels = 32
x = F.relu(self.lin2(x)) # x torch.Size([20, 32])
x = self.lin3(x) #x torch.Size([20, 2])
mincut_loss = mincut_loss2 + ortho_loss2
#print("x", x)
return F.log_softmax(x, dim=-1), mincut_loss2, ortho_loss2
class DiffWire(torch.nn.Module):
def __init__(self, in_channels, out_channels, k_centers, derivative=None, hidden_channels=32, EPS=1e-15, device=None):
super(DiffWire, self).__init__()
self.EPS=EPS
self.derivative = derivative
self.device=device
# First X transformation
self.lin1 = Linear(in_channels, hidden_channels)
#Fiedler vector -- Pool previous to GAP-Layer
self.pool_rw = Linear(hidden_channels, 2)
#CT Embedding -- Pool previous to CT-Layer
self.num_of_centers1 = k_centers # k1 - order of number of nodes
self.pool_ct = Linear(hidden_channels, self.num_of_centers1) #CT
#Conv1
self.conv1 = DenseGraphConv(hidden_channels, hidden_channels)
#MinCutPooling
self.pool_mc = Linear(hidden_channels, 16) #MC
#Conv2
self.conv2 = DenseGraphConv(hidden_channels, hidden_channels)
# MLPs towards out
self.lin2 = Linear(hidden_channels, hidden_channels)
self.lin3 = Linear(hidden_channels, out_channels)
def forward(self, x, edge_index, batch):
# Make all adjacencies of size NxN
adj = to_dense_adj(edge_index, batch)
# Make all x_i of size N=MAX(N1,...,N20), e.g. N=40:
x, mask = to_dense_batch(x, batch)
x = self.lin1(x)
if torch.isnan(adj).any():
print("adj nan")
if torch.isnan(x).any():
print("x nan")
#Gap Layer RW
s0 = self.pool_rw(x)
adj, mincut_loss_rw, ortho_loss_rw = dense_mincut_rewiring(x, adj, s0, mask,
derivative = self.derivative, EPS=self.EPS, device=self.device)
# CT REWIRING
# First mincut pool for computing Fiedler adn rewire
s1 = self.pool_ct(x)
adj, CT_loss, ortho_loss_ct = dense_CT_rewiring(x, adj, s1, mask, EPS = self.EPS) # out: x torch.Size([20, N, F'=32]), adj torch.Size([20, N, N])
# CONV1: Now on x and rewired adj:
x = self.conv1(x, adj)
# MINCUT_POOL
# MLP of k=16 outputs s
s2 = self.pool_mc(x)
# Call to dense_cut_mincut_pool to get coarsened x, adj and the losses: k=16
x, adj, mincut_loss, ortho_loss_mc = dense_mincut_pool(x, adj, s2, mask, EPS=self.EPS) # out x torch.Size([20, k=16, F'=32]), adj torch.Size([20, k2=16, k2=16])
# CONV2: Now on coarsened x and adj:
x = self.conv2(x, adj)
# Readout for each of the 20 graphs
x = x.sum(dim=1)
# Final MLP for graph classification: hidden channels = 32
x = F.relu(self.lin2(x))
x = self.lin3(x)
main_loss = mincut_loss_rw + CT_loss + mincut_loss
ortho_loss = ortho_loss_rw + ortho_loss_ct + ortho_loss_mc
#ortho_loss_rw/2 + (1/self.num_of_centers1)*ortho_loss_ct + ortho_loss_mc/16
#print("x", x)
return F.log_softmax(x, dim=-1), main_loss, ortho_loss