-
Notifications
You must be signed in to change notification settings - Fork 0
/
concat_search.py
168 lines (142 loc) · 6.47 KB
/
concat_search.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
import optuna
import numpy as np
import math
import random
import scipy.sparse as sp
from spectral_propagation import get_embedding_dense, propagate
from utils import load_embedding, load_adjacency_mx, sigmoid
class opConcatSearch(object):
def __init__(self, args):
self.prop_types = args.prop_types
self.max_evals = args.max_evals
self.svd = args.svd
self.loss_type = args.loss
self.n_workers = args.workers
# load adjacency matrix and raw embedding
self.emb = load_embedding(args.emb)
self.adj, self.num_nodes, self.num_edges = load_adjacency_mx(args.adj)
self.laplacian = None
self.dim = self.emb.shape[1]
assert self.num_nodes == self.emb.shape[0]
# negative pairs
self.negative_pairs = int(math.sqrt(self.num_edges))
if self.loss_type == "infonce":
self.batch_size = 64
neg_index = []
for i in range(self.num_nodes):
select = np.random.choice(self.num_nodes, self.batch_size, replace=False)
while i in select:
select = np.random.choice(self.num_nodes, self.batch_size, replace=False)
neg_index.append(select)
self.neg_index = np.array(neg_index)
self.neg_emb = self.emb[self.neg_index]
elif self.loss_type == "infomax":
# self.permutation = np.random.permutation(np.arange(self.num_nodes))
pass
def build_search_space(self, trial):
space = {}
for f in self.prop_types:
space[f] = trial.suggest_categorical(f, [0, 1])
if space.get("heat", 0) == 1:
space["t"] = trial.suggest_uniform("t", 0.1, 0.9)
if space.get("gaussian", 0) == 1:
space["mu"] = trial.suggest_uniform("mu", 0.1, 2)
space["theta"] = trial.suggest_uniform("theta", 0.2, 1.5)
if space.get("ppr", 0) == 1:
space["alpha"] = trial.suggest_uniform("alpha", 0.2, 0.8)
return space
def prop(self, params):
prop_types = [key for key, value in params.items() if value == 1 and key in self.prop_types]
if not prop_types:
print(" -- dropped -- ")
return None, None
prop_result_list = []
for selected_prop in prop_types:
prop_result = propagate(self.adj, self.emb, selected_prop, params)
prop_result_list.append(prop_result)
if self.loss_type == "infomax":
neg_prop_result = []
pmt = self.permutation
for s_prop in prop_types:
neg_prop = propagate(self.adj, self.emb[pmt], s_prop, params)
neg_prop[pmt] = neg_prop
neg_prop_result.append(neg_prop)
return np.array(prop_result_list), np.array(neg_prop_result)
elif self.loss_type == "infonce":
return np.array(prop_result_list), None
elif self.loss_type == "sparse":
return np.array(prop_result_list), None
else:
raise ValueError("use 'infonce', 'infomax' or 'sparse' loss, currently using {}".format(self.loss_type))
def target_func(self, trial):
params = self.build_search_space(trial)
self.permutation = np.random.permutation(np.arange(self.num_nodes))
prop_result_emb, neg_prop_result_emb = self.prop(params)
if prop_result_emb is None:
return 100
if self.loss_type == "infomax":
loss = self.infomax_loss(prop_result_emb, neg_prop_result_emb)
elif self.loss_type == "infonce":
loss = self.infonce_loss(prop_result_emb)
else:
loss = self.sparse_cut_loss(prop_result_emb)
return loss
def sparse_cut_loss(self, prop_emb_list):
if not self.laplacian:
degree = self.adj.sum(1).todense()
degree_matrix = sp.csc_matrix(
(degree, (np.arange(self.num_nodes), np.arange(self.num_nodes))),
shape=(self.num_nodes, self.num_nodes))
self.laplacian = degree_matrix - self.adj
# Sparest Cut Loss
prop_emb = np.concatenate(prop_emb_list, axis=1)
if self.svd:
prop_emb = get_embedding_dense(prop_emb, self.dim)
pairs = np.random.choice(self.num_nodes, (2, self.negative_pairs))
neg_emb = prop_emb[pairs]
loss = np.sum((neg_emb[0] - neg_emb[1]) ** 2, axis=1).mean()
return 1. / loss
def infonce_loss(self, prop_emb_list, *args, **kwargs):
T = 0.07
pos_infos = []
for smoothed in prop_emb_list:
pos_info = np.exp(np.sum(smoothed * self.emb, -1) / T)
assert pos_info.shape == (self.num_nodes,)
pos_infos.append(pos_info)
neg_infos = []
for idx, smoothed in enumerate(prop_emb_list):
neg_info = np.exp(np.sum(np.tile(smoothed[:, np.newaxis, :], (1, self.batch_size, 1)) * self.neg_emb, -1) / T).sum(-1)
assert neg_info.shape == (self.num_nodes,)
neg_infos.append(neg_info + pos_infos[idx])
pos_neg = np.array(pos_infos) / np.array(neg_infos)
if np.isnan(pos_neg).any():
pos_neg = np.nan_to_num(pos_neg)
loss = -np.log(pos_neg).mean()
return loss/10
def infomax_loss(self, prop_emb_list, neg_prop_emb_list):
prop_result = np.concatenate(prop_emb_list, axis=1)
if self.svd:
prop_result = get_embedding_dense(prop_result, self.dim)
pos_glb = prop_result.mean(0)
pos_info = sigmoid(pos_glb.dot(prop_result.T))
pos_loss = np.mean(np.log(pos_info)).mean()
neg_loss = 0
neg_step = 1
for _ in range(neg_step):
neg_prop_result = np.concatenate(neg_prop_emb_list, axis=1)
if self.svd:
neg_prop_result = get_embedding_dense(neg_prop_result, self.dim)
neg_info = sigmoid(pos_glb.dot(neg_prop_result.T))
neg_loss += np.mean(np.log(1 - neg_info)).mean()
random.shuffle(neg_prop_emb_list)
return -(pos_loss + neg_loss) / (1 + neg_step)
def __call__(self):
study = optuna.create_study()
study.optimize(self.target_func, n_jobs=self.n_workers, n_trials=self.max_evals)
best_params = study.best_params
best_result = self.prop(best_params)[0]
best_result = np.concatenate(best_result, axis=1)
print(f"best parameters: {best_params}")
if self.svd:
best_result = get_embedding_dense(best_result, self.dim)
return best_result