-
Notifications
You must be signed in to change notification settings - Fork 4
/
utils.py
222 lines (172 loc) · 7.35 KB
/
utils.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
import torch
from torch import nn
from torch.nn import functional as F
import numpy as np
from torch_geometric.utils import get_laplacian, to_dense_adj, to_dense_batch
from models import device
def gradient_gcn(regularizer, fnet, data, input, grad):
gradient = []
reg = []
net = []
ad1 = get_laplacian(edge_index=data.edge_index, edge_weight=data.edge_attr, normalization='sym')
ad1 = to_dense_adj(edge_index=ad1[0], edge_attr=ad1[1], max_num_nodes=data.x.shape[0])[0, :, :]
ad1.requires_grad_(False)
ad1.to(device)
for params in regularizer.parameters():
reg.append(params.data.requires_grad_(True))
alpha = torch.nn.ReLU()(torch.matmul(input, reg[0]) + reg[1])
alpha = torch.matmul(alpha, reg[2].T) + reg[3]
alpha = torch.nn.Softmax(dim=0)(alpha)
alpha = alpha / (alpha.max() * 1200.0)
for param in fnet.parameters():
net.append(param.data.requires_grad_(False))
for i in range(len(alpha)):
net[i] = net[i] - alpha[i] * (grad[i].data + 0.0001 * net[i])
for j in range(0, len(net), 2):
if j == 0:
out1 = torch.nn.ReLU()(torch.spmm(ad1, torch.matmul(data.x, net[j+1].T)) + net[j])
else:
out1 = torch.nn.ReLU()(torch.spmm(ad1, torch.matmul(out1, net[j+1].T)) + net[j])
loss = sub_loss(out1, data)
torch.autograd.set_detect_anomaly(True)
loss.backward()
for k in range(len((reg))):
gradient.append(reg[k].grad.data)
return gradient
def gradient_gat(regularizer, fnet, data, input, grad):
adj1 = to_dense_adj(data.edge_index, edge_attr=data.edge_attr, max_num_nodes=data.x.shape[0])[0, :, :].to(device)
gradient = []
reg = []
net = []
for params in regularizer.parameters():
reg.append(params.data.requires_grad_(True))
alpha = torch.nn.ReLU()(torch.matmul(input, reg[0]) + reg[1])
alpha = torch.matmul(alpha, reg[2].T) + reg[3]
alpha = torch.nn.Softmax(dim=0)(alpha)
alpha = alpha / (alpha.max() * 1200.0)
for param in fnet.parameters():
net.append(param.data.requires_grad_(False))
for i in range(len(alpha)):
net[i] = net[i] - alpha[i] * (grad[i].data + 0.0001 * net[i])
for j in range(0, len(net), 4):
if j == 0:
Wh = torch.mm(data.x, net[j + 3].T)
Wh1 = torch.mm(Wh, net[j][0].T)
Wh2 = torch.mm(Wh, net[j + 1][0].T)
e = Wh1 + Wh2.T
e = nn.LeakyReLU(0.3)(e)
zero_vec = -9e15 * torch.ones_like(e)
attention = torch.where(adj1 > 0, e, zero_vec)
attention = F.softmax(attention, dim=1)
attention = F.dropout(attention, 0.2)
out1 = nn.ReLU()(torch.matmul(attention, Wh) + net[j + 2])
else:
Wh = torch.mm(out1, net[j + 3].T)
Wh1 = torch.mm(Wh, net[j][0].T)
Wh2 = torch.mm(Wh, net[j + 1][0].T)
e = Wh1 + Wh2.T
e = nn.LeakyReLU(0.3)(e)
zero_vec = -9e15 * torch.ones_like(e)
attention = torch.where(adj1 > 0, e, zero_vec)
attention = F.softmax(attention, dim=1)
attention = F.dropout(attention, 0.2)
out1 = nn.ReLU()(torch.matmul(attention, Wh) + net[j + 2])
loss = sub_loss(out1, data)
torch.autograd.set_detect_anomaly(True)
loss.backward()
for k in range(len((reg))):
gradient.append(reg[k].grad.data)
return gradient
def gradient_gin(regularizer, fnet, data, input, grad):
gradient = []
reg = []
net = []
z1 = []
for params in regularizer.parameters():
reg.append(params.data.requires_grad_(True))
alpha = torch.nn.ReLU()(torch.matmul(input, reg[0]) + reg[1])
alpha = torch.matmul(alpha, reg[2].T) + reg[3]
alpha = torch.nn.Softmax(dim=0)(alpha)
alpha = alpha / (alpha.max() * 1200.0)
for param in fnet.parameters():
net.append(param.data.requires_grad_(False))
for i in range(len(alpha)):
net[i] = net[i] - alpha[i] * (grad[i].data + 0.0001 * net[i])
for j in range(0, len(net), 6):
if j == 0:
out1 = data.x + fnet.encoder.layers[int(j / 6)].propagate(data.edge_index, x=data.x,
edge_attr=data.edge_attr)
out1 = torch.nn.ReLU()(torch.matmul(out1, net[j].T) + net[j + 1])
out1 = torch.nn.ReLU()(torch.matmul(out1, net[j + 2].T) + net[j + 3])
out1 = (out1 * net[j + 4]) + net[j + 5]
z1.append(out1)
else:
out1 = out1 + fnet.encoder.layers[int(j / 6)].propagate(data.edge_index, x=out1, edge_attr=data.edge_attr)
out1 = torch.nn.ReLU()(torch.matmul(out1, net[j].T) + net[j + 1])
out1 = torch.nn.ReLU()(torch.matmul(out1, net[j + 2].T) + net[j + 3])
out1 = (out1 * net[j + 4]) + net[j + 5]
z1.append(out1)
out1 = torch.cat(z1, dim=1)
loss = sub_loss(out1, data)
torch.autograd.set_detect_anomaly(True)
loss.backward()
for k in range(len((reg))):
gradient.append(reg[k].grad.data)
return gradient
def Adam(param, grad, alpha):
beta_1 = 0.9
beta_2 = 0.999
epsilon = 1e-8
m_t = 0
v_t = 0
t = 0
for _ in range(1):
t += 1
m_t = beta_1 * m_t + (1 - beta_1) * grad
v_t = beta_2 * v_t + (1 - beta_2) * (torch.pow(grad, 2))
m_cap = m_t / (1 - (beta_1 ** t))
v_cap = v_t / (1 - (beta_2 ** t))
param = param - (alpha * m_cap) / (torch.sqrt(v_cap) + epsilon)
return param
def sub_loss(z, data):
batch_z = to_dense_batch(z, data.batch)[0].to(device)
batch_z = F.normalize(batch_z, dim=2)
pos_mask = to_dense_adj(edge_index=data.edge_index, edge_attr=data.edge_attr, batch=data.batch)
num_nodes = batch_z.shape[1]
num_graphs = batch_z.shape[0]
pos_mask += torch.linalg.matrix_power(pos_mask, 2)
pos_mask[pos_mask != 0.0] = 1.0
neg_mask = 1. - pos_mask
batch_idx = np.arange(0, batch_z.shape[0])
loss = 0.0
for j in range(batch_z.shape[0]):
if j == 0:
loss += intra_jsd(batch_z, batch_z, pos_mask, neg_mask)
else:
idx = np.roll(batch_idx, j).tolist()
loss += inter_jsd(batch_z, batch_z, pos_mask, idx)
loss /= num_nodes
return loss
def intra_jsd(cand1, cand2, pos_mask, neg_mask):
num_neg = neg_mask.int().sum()
num_pos = pos_mask.int().sum()
similarity = torch.bmm(cand1, torch.transpose(cand2, 1, 2))
E_pos = (np.log(2) - F.softplus(- similarity * pos_mask)).sum()
E_pos /= num_pos
neg_sim = similarity * neg_mask
E_neg = (F.softplus(- neg_sim) + neg_sim - np.log(2)).sum()
E_neg /= num_neg
return E_neg - E_pos
def inter_jsd(cand1, cand2, pos_mask, idx):
pos_mask = pos_mask + pos_mask[idx]
pos_mask[pos_mask != 0.0] = 1.0
neg_mask = 1.0 - pos_mask
num_neg = neg_mask.int().sum()
num_pos = pos_mask.int().sum()
similarity = torch.bmm(cand1, torch.transpose(cand2[idx], 1, 2))
E_pos = (np.log(2) - F.softplus(- similarity * pos_mask)).sum()
E_pos /= num_pos
neg_sim = similarity * neg_mask
E_neg = (F.softplus(- neg_sim) + neg_sim - np.log(2)).sum()
E_neg /= num_neg
return E_neg - E_pos