-
Notifications
You must be signed in to change notification settings - Fork 0
/
SHIG_trainer.py
194 lines (161 loc) · 8.13 KB
/
SHIG_trainer.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
# -*- coding: utf-8 -*-
"""
Created on Fri Apr 2 10:37:11 2021
@author: sxp
"""
import time
import torch
import torch.nn as nn
import random
import numpy as np
from tqdm import trange
from utils import setup_features
from sklearn.model_selection import train_test_split
from SHIG import SHIG_Model
from tensorboardX import SummaryWriter
import datetime
from optimizers.radam import RiemannianAdam
class SHIGNetwork(torch.nn.Module):
def __init__(self, device, args, trial, X):
super(SHIGNetwork, self).__init__()
"""
SGCN Initialization.
:param device: Device for calculations.
:param args: Arguments object.
:param X: Node features.
"""
self.args = args
self.trial = trial
torch.manual_seed(self.args.seed)
self.device = device
self.X = X
self.setup_layers()
def setup_layers(self):
"""
Adding Base Layers, Deep Signed GraphSAGE layers.
Assing Regression Parameters if the model is not a single layer model.
"""
self.nodes = range(self.X.shape[0])
self.neurons = self.args.layers
self.layers = len(self.neurons)
self.aggregator = SHIG_Model(self.X.shape[1], self.neurons[-1], num_layers=self.args.num_layers,
trial=self.trial, args=self.args).cuda()
def forward(self, positive_edges, negative_edges, target):
"""
Model forward propagation pass. Can fit deep and single layer SGCN models.
:param positive_edges: Positive edges.
:param negative_edges: Negative edges.
:param target: Target vectors.
:return loss: Loss value.
:return self.z: Hidden vertex representations.
"""
self.z = self.aggregator.forward(self.X, positive_edges, negative_edges)
loss = self.aggregator.loss(self.z, positive_edges, negative_edges, self.device)
return loss, self.z
class SHIGTrainer(object):
def __init__(self, args, edges):
self.args = args
self.edges = edges
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.setup_logs()
def setup_logs(self):
"""
Creating a log dictionary.
"""
self.logs = {}
self.logs["parameters"] = vars(self.args)
self.logs["loss"] = []
self.logs["performance"] = [["Epoch", "AUC", "F1"]]
self.logs["training_time"] = [["Epoch", "Seconds"]]
# tensorboard
self.writer = SummaryWriter(self.args.log_path + self.args.dataset + '_Layer_{}/'.format(self.args.num_layers)+'_{}'.
format((datetime.datetime.now()).strftime("%Y%m%d%H%M%S")))
def setup_dataset(self):
"""
Creating train and test split.
"""
self.positive_edges, self.test_positive_edges = train_test_split(self.edges["positive_edges"],
test_size=self.args.test_size,
random_state=self.args.seed)
self.negative_edges, self.test_negative_edges = train_test_split(self.edges["negative_edges"],
test_size=self.args.test_size,
random_state=self.args.seed)
self.ecount = len(self.positive_edges + self.negative_edges)
self.neg_ratio = len(self.negative_edges) / self.ecount
self.X = setup_features(self.args,
self.positive_edges,
self.negative_edges,
self.edges["ncount"])
self.positive_edges = torch.from_numpy(np.array(self.positive_edges,
dtype=np.int64).T).type(torch.long).cuda()
self.negative_edges = torch.from_numpy(np.array(self.negative_edges,
dtype=np.int64).T).type(torch.long).cuda()
self.y = np.array([0 if i < int(self.ecount/2) else 1 for i in range(self.ecount)]+[2]*(self.ecount*2))
self.y = torch.from_numpy(self.y).type(torch.LongTensor).cuda()
self.X = self.X.cuda()
def score_model(self, epoch, last=False):
"""
Score the model on the test set edges in each epoch.
:param epoch: Epoch number.
"""
self.model.eval()
score_positive_edges = torch.from_numpy(np.array(self.test_positive_edges, dtype=np.int64).T).type(torch.long).cuda()
score_negative_edges = torch.from_numpy(np.array(self.test_negative_edges, dtype=np.int64).T).type(torch.long).cuda()
loss, self.z = self.model(self.positive_edges, self.negative_edges, self.y)
auc, f1, f1_macro, f1_micro = self.model.aggregator.test(self.z, score_positive_edges, score_negative_edges, self.neg_ratio, last)
# self.trial.report(auc, epoch+1)
self.logs["performance"].append([epoch+1, auc, f1_micro, f1, f1_macro])
self.writer.add_scalar('AUC', auc, epoch)
self.writer.add_scalar('F1', f1, epoch)
self.writer.add_scalar('F1_macro', f1_macro, epoch)
self.writer.add_scalar('F1_micro', f1_micro, epoch)
if last:
embedding_pos = torch.cat([self.z[score_positive_edges[0]], self.z[score_positive_edges[1]]], dim=1).cpu()
embedding_neg = torch.cat([self.z[score_negative_edges[0]], self.z[score_negative_edges[1]]], dim=1).cpu()
embedding = torch.cat([embedding_pos, embedding_neg], dim=0)
y = torch.cat(
[embedding.new_ones((score_positive_edges.size(1))),
-1*embedding.new_ones(score_negative_edges.size(1))])
embedding, y = embedding.detach().numpy(), y.int().numpy()
self.writer.add_embedding(embedding, metadata=y, global_step=epoch)
self.writer.close()
torch.save(self.z,"output/z"+".pt")
if self.args.verbose:
print('{}{} Val(auc,f1,f1_macro,f1_micro):{} {} {} {}'.format("#" * 10, "BEST EPOCH",
self.logs["performance"][-1][1],
self.logs["performance"][-1][3],
self.logs["performance"][-1][4],
self.logs["performance"][-1][2]))
self.model.train()
def create_and_train_model(self, trial):
"""
Model training and scoring.
"""
print("\nTraining started.\n")
self.trial = trial
# self.model = SHIGNetwork(self.device, self.args, trial, self.X).cuda()
self.model = SHIGNetwork(self.device, self.args, trial, self.X)
# self.optimizer = torch.optim.Adam(self.model.parameters(),
# lr=self.args.learning_rate,
# weight_decay=self.args.weight_decay)
self.optimizer = RiemannianAdam(self.model.parameters(),
lr=self.args.learning_rate,
weight_decay=self.args.weight_decay)
self.lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(self.optimizer, self.args.epochs)
last = False
self.model.train()
self.epochs = trange(self.args.epochs, desc="Loss")
for epoch in self.epochs:
start_time = time.time()
self.optimizer.zero_grad()
loss, _ = self.model(self.positive_edges, self.negative_edges, self.y)
self.logs["loss"].append(loss.item())
loss.backward()
self.epochs.set_description("SHIG (Loss=%g)" % round(loss.item(), 4))
self.optimizer.step()
self.lr_scheduler.step()
self.logs["training_time"].append([epoch+1, time.time()-start_time])
if self.args.test_size > 0:
if epoch == self.args.epochs -1:
last = True
self.score_model(epoch, last)