-
Notifications
You must be signed in to change notification settings - Fork 1
/
GNN.py
executable file
·329 lines (270 loc) · 13.1 KB
/
GNN.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import Sequential, Linear, ReLU
from torch_geometric.nn import NNConv
import torch_geometric.utils as utils
import config
import helper
from torch.nn import init
# GCN basic operation
class GraphConv(nn.Module):
def __init__(self, input_dim, output_dim, add_self=False, normalize_embedding=False,
dropout=0.0, bias=True):
super(GraphConv, self).__init__()
self.add_self = add_self
self.dropout = dropout
if dropout > 0.001:
self.dropout_layer = nn.Dropout(p=dropout)
self.normalize_embedding = normalize_embedding
self.input_dim = input_dim
self.output_dim = output_dim
self.weight = nn.Parameter(torch.FloatTensor(input_dim, output_dim).cpu())
if bias:
self.bias = nn.Parameter(torch.FloatTensor(output_dim).cpu())
else:
self.bias = None
def forward(self, x, adj):
if self.dropout > 0.001:
x = self.dropout_layer(x)
y = torch.matmul(adj, x)
if self.add_self:
y += x
y = torch.matmul(y, self.weight)
if self.bias is not None:
y = y + self.bias
if self.normalize_embedding:
y = F.normalize(y, p=2, dim=2)
return y
class GcnEncoderGraph(nn.Module):
def __init__(self, input_dim, hidden_dim, embedding_dim, num_layers, view_dim,
concat=True, bn=True, args=None):
super(GcnEncoderGraph, self).__init__()
self.concat = concat
self.bn = bn
self.num_layers = num_layers
self.num_aggs = 1
self.bias = True
if args is not None:
self.bias = args.bias
self.conv_first, self.conv_block, self.conv_last = self.build_conv_layers(
input_dim, hidden_dim, embedding_dim, view_dim, num_layers)
if concat:
self.pred_input_dim = hidden_dim * (num_layers - 1) + embedding_dim
else:
self.pred_input_dim = embedding_dim
for m in self.modules():
if isinstance(m, GraphConv):
m.weight.data = init.xavier_uniform_(m.weight.data, gain=nn.init.calculate_gain('relu'))
if m.bias is not None:
m.bias.data = init.constant_(m.bias.data, 0.0)
def build_conv_layers(self, input_dim, hidden_dim, embedding_dim, view_dim, num_layers):
# For only 3 layers:
seq_nn = Sequential(Linear(view_dim, input_dim * hidden_dim), ReLU())
conv_first = NNConv(input_dim, hidden_dim, seq_nn, aggr='mean')
seq_nn = Sequential(Linear(view_dim, hidden_dim * hidden_dim), ReLU())
conv_block = nn.ModuleList([NNConv(hidden_dim, hidden_dim, seq_nn, aggr='mean') for i in range(num_layers - 2)])
seq_nn = Sequential(Linear(view_dim, hidden_dim * embedding_dim), ReLU())
conv_last = NNConv(hidden_dim, embedding_dim, seq_nn, aggr='mean')
return conv_first, conv_block, conv_last
def build_pred_layers(self, pred_input_dim, pred_hidden_dims, label_dim, num_aggs=1):
pred_input_dim = pred_input_dim * num_aggs
if len(pred_hidden_dims) == 0:
pred_model = nn.Linear(pred_input_dim, label_dim)
else:
pred_layers = []
for pred_dim in pred_hidden_dims:
pred_layers.append(nn.Linear(pred_input_dim, pred_dim))
pred_layers.append(self.act)
pred_input_dim = pred_dim
pred_layers.append(nn.Linear(pred_dim, label_dim)) # Put into loop? FSD 02.06.22
pred_model = nn.Sequential(*pred_layers)
return pred_model
def construct_mask(self, max_nodes, batch_num_nodes):
''' For each num_nodes in batch_num_nodes, the first num_nodes entries of the
corresponding column are 1's, and the rest are 0's (to be masked out).
Dimension of mask: [batch_size x max_nodes x 1]
'''
# masks
packed_masks = [torch.ones(int(num)) for num in batch_num_nodes]
batch_size = len(batch_num_nodes)
out_tensor = torch.zeros(batch_size, max_nodes)
for i, mask in enumerate(packed_masks):
out_tensor[i, :int(batch_num_nodes[i])] = mask # (there was no int(...)) FSD 02.01.22
return out_tensor.unsqueeze(2).cpu()
def apply_bn(self, x):
''' Batch normalization of 3D tensor x
'''
bn_module = nn.BatchNorm1d(x.size()[1]).cpu()
return bn_module(x)
def gcn_forward(self, x, edge_index, edge_attr, conv_first, conv_block, conv_last, embedding_mask=None):
''' Perform forward prop with graph convolution.
Returns:
Embedding matrix with dimension [batch_size x num_nodes x embedding]
'''
x = conv_first(x, edge_index, edge_attr)
if self.bn:
x = self.apply_bn(x)
x_all = [x]
for i in range(len(conv_block)):
x = conv_block[i](x, edge_index, edge_attr)
if self.bn:
x = self.apply_bn(x)
x_all.append(x)
x = conv_last(x, edge_index, edge_attr)
x_all.append(x)
x_tensor = torch.cat(x_all, dim=1)
if embedding_mask is not None:
x_tensor = x_tensor * embedding_mask
return x_tensor
def forward(self, x, adj, batch_num_nodes=None, **kwargs):
# mask
max_num_nodes = adj.size()[1]
if batch_num_nodes is not None:
self.embedding_mask = self.construct_mask(max_num_nodes, batch_num_nodes)
else:
self.embedding_mask = None
# conv
x = self.conv_first(x, adj)
if self.bn:
x = self.apply_bn(x)
out_all = []
out, _ = torch.max(x, dim=1)
out_all.append(out)
for i in range(self.num_layers - 2):
x = self.conv_block[i](x, adj)
if self.bn:
x = self.apply_bn(x)
out, _ = torch.max(x, dim=1)
out_all.append(out)
if self.num_aggs == 2:
out = torch.sum(x, dim=1)
out_all.append(out)
x = self.conv_last(x, adj)
out, _ = torch.max(x, dim=1)
out_all.append(out)
if self.num_aggs == 2:
out = torch.sum(x, dim=1)
out_all.append(out)
if self.concat:
output = torch.cat(out_all, dim=1)
else:
output = out
ypred = self.pred_model(output)
return ypred
class SoftPoolingGcnEncoder(GcnEncoderGraph):
def __init__(self, max_num_nodes, input_dim, hidden_dim, embedding_dim, num_layers,
assign_hidden_dim, view_dim, not_ablated=True, assign_ratio=0.25, assign_num_layers=-1, num_pooling=1,
concat=True, bn=True, assign_input_dim=-1, args=None): # FSD 02.06.22
super(SoftPoolingGcnEncoder, self).__init__(input_dim, hidden_dim, embedding_dim, num_layers,
view_dim, concat=concat, bn=bn, args=args)
self.num_pooling = num_pooling
self.assign_ent = True
self.not_ablated = not_ablated
# GC
self.conv_first_after_pool = nn.ModuleList()
self.conv_block_after_pool = nn.ModuleList()
self.conv_last_after_pool = nn.ModuleList()
for i in range(num_pooling):
# use self to register the modules in self.modules()
conv_first2, conv_block2, conv_last2 = self.build_conv_layers(
self.pred_input_dim, hidden_dim, embedding_dim, view_dim, num_layers)
self.conv_first_after_pool.append(conv_first2)
self.conv_block_after_pool.append(conv_block2)
self.conv_last_after_pool.append(conv_last2)
# assignment
assign_dims = []
if assign_num_layers == -1:
assign_num_layers = num_layers
if assign_input_dim == -1:
assign_input_dim = input_dim
self.assign_conv_first_modules = nn.ModuleList()
self.assign_conv_block_modules = nn.ModuleList()
self.assign_conv_last_modules = nn.ModuleList()
self.assign_pred_modules = nn.ModuleList()
assign_dim = round(max_num_nodes * assign_ratio)
for i in range(num_pooling):
assign_dims.append(assign_dim)
assign_conv_first, assign_conv_block, assign_conv_last = self.build_conv_layers(
assign_input_dim, assign_hidden_dim, assign_dim, view_dim, assign_num_layers)
assign_pred_input_dim = assign_hidden_dim * (num_layers - 1) + assign_dim if concat else assign_dim
assign_pred = self.build_pred_layers(assign_pred_input_dim, [], assign_dim, num_aggs=1)
# next pooling layer
assign_input_dim = self.pred_input_dim
assign_dim = round(assign_dim * assign_ratio)
self.assign_conv_first_modules.append(assign_conv_first)
self.assign_conv_block_modules.append(assign_conv_block)
self.assign_conv_last_modules.append(assign_conv_last)
self.assign_pred_modules.append(assign_pred)
n_Encoded_Embeddings = (num_pooling + 1 * not_ablated) * (
hidden_dim * (num_layers - 1) + embedding_dim) if concat else (
hidden_dim * (num_layers - 1) + embedding_dim)
self.decode = nn.Linear(n_Encoded_Embeddings, config.N_Nodes) # n_Encoded_Embeddings -> n_ROIs. FSD 01.31.22
for m in self.modules():
if isinstance(m, GraphConv):
m.weight.data = init.xavier_uniform_(m.weight.data, gain=nn.init.calculate_gain('relu'))
if m.bias is not None:
m.bias.data = init.constant_(m.bias.data, 0.0)
def forward(self, data, batch_num_nodes=None):
x, edge_attr, edge_index, adj = data.x, data.edge_attr, data.edge_index, data.con_mat
x_a = x # (For the mean time) FSD 02.06.22
# mask
max_num_nodes = config.N_Nodes # adj.size()[1] FSD 02.06.22
if batch_num_nodes is not None:
embedding_mask = self.construct_mask(max_num_nodes, batch_num_nodes)
else:
embedding_mask = None
out_all = []
# GCN(X, A) = Z
embedding_tensor = self.gcn_forward(x, edge_index, edge_attr,
self.conv_first, self.conv_block, self.conv_last, embedding_mask)
if (self.not_ablated):
out, _ = torch.max(embedding_tensor, dim=0)
if self.num_aggs == 2:
out = torch.sum(embedding_tensor, dim=0)
out_all.append(out)
S_list = []
for i in range(self.num_pooling):
if batch_num_nodes is not None and i == 0:
embedding_mask = self.construct_mask(max_num_nodes, batch_num_nodes)
else:
embedding_mask = None
# softmax(GCN(X, A)) = S
self.assign_tensor = self.gcn_forward(x_a, edge_index, edge_attr,
self.assign_conv_first_modules[i], self.assign_conv_block_modules[i],
self.assign_conv_last_modules[i],
embedding_mask)
self.assign_tensor = nn.Softmax(dim=-1)(self.assign_pred_modules[i](self.assign_tensor))
S_list.append(self.assign_tensor)
# update pooled features and adj matrix
# S * Z = X
x = torch.matmul(torch.transpose(self.assign_tensor, 0, 1), embedding_tensor)
# S.T * A * S = A
if i > 0:
adj = utils.to_dense_adj(edge_index, edge_attr=edge_attr)
adj = torch.squeeze(adj)
repeated_S = self.assign_tensor.repeat(adj.size(2), 1, 1)
adj = torch.permute(adj, (2, 0, 1))
adj = torch.matmul(torch.transpose(repeated_S, 1, 2), adj) @ repeated_S
for j in range(adj.size(0)):
adj[j, :, :].fill_diagonal_(0)
adj = torch.permute(adj, (1, 2, 0))
edge_index, edge_attr = helper.dense_to_sparse(adj)
x_a = x # FSD 02.06.22
embedding_tensor = self.gcn_forward(x, edge_index, edge_attr,
self.conv_first_after_pool[i], self.conv_block_after_pool[i],
self.conv_last_after_pool[i])
out, _ = torch.max(embedding_tensor, dim=0)
out_all.append(out)
if self.num_aggs == 2:
out = torch.sum(embedding_tensor, dim=0)
out_all.append(out)
if self.concat:
output = torch.cat(out_all, dim=-1)
else:
output = out
D = self.decode(output) # N_encoded_embeddings -> N_ROIs
repeated_D = D.repeat(max_num_nodes, 1)
diff = torch.abs(repeated_D - torch.transpose(repeated_D, 0, 1))
cbt = diff
return cbt, S_list