-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathSGST_sentinet_model.py
247 lines (211 loc) · 13.8 KB
/
SGST_sentinet_model.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
'''
author:Donglin Zhou
time:2021.12.15
function:SGST-model
使用情感引导和价格序列引导的未来股市长期价格预测
'''
from new_attention import *
from sparse_attention import *
import numpy as np
# 前馈神经网络
class PointWiseFeedForward(torch.nn.Module):
def __init__(self, hidden_units, dropout_rate):
super(PointWiseFeedForward, self).__init__()
# 一维卷积，输入向量维度，输出向量维度kernel_size卷积步长
self.conv1 = torch.nn.Conv1d(hidden_units, hidden_units, kernel_size=1)
self.dropout1 = torch.nn.Dropout(p=dropout_rate)
self.relu = torch.nn.ReLU()
self.conv2 = torch.nn.Conv1d(hidden_units, hidden_units, kernel_size=1)
self.dropout2 = torch.nn.Dropout(p=dropout_rate)
def forward(self, inputs):
# 最后两维转置，然后进行一维卷积，然后dropout，然后relu激活函数，然后再进行一次一维卷积
outputs = self.dropout2(self.conv2(self.relu(self.dropout1(self.conv1(inputs.transpose(-1, -2))))))
# as Conv1D requires (N, C, Length) 转置回（128，200，50）
outputs = outputs.transpose(-1, -2)
outputs += inputs
return outputs
class SGST(torch.nn.Module):
def __init__(self, args):
super(SGST, self).__init__()
self.dev = args.device
self.attention_layernorms = torch.nn.ModuleList()
self.attention_layers = torch.nn.ModuleList()
self.forward_layernorms = torch.nn.ModuleList()
self.forward_layers = torch.nn.ModuleList()
self.feature1_attention_layernorms = torch.nn.ModuleList()
self.feature1_attention_layers = torch.nn.ModuleList()
self.feature1_forward_layernorms = torch.nn.ModuleList()
self.feature1_forward_layers = torch.nn.ModuleList()
self.feature2_attention_layernorms = torch.nn.ModuleList()
self.feature2_attention_layers = torch.nn.ModuleList()
self.feature2_forward_layernorms = torch.nn.ModuleList()
self.feature2_forward_layers = torch.nn.ModuleList()
self.feature3_attention_layernorms = torch.nn.ModuleList()
self.feature3_attention_layers = torch.nn.ModuleList()
self.feature3_forward_layernorms = torch.nn.ModuleList()
self.feature3_forward_layers = torch.nn.ModuleList()
self.pos_attention_layernorms = torch.nn.ModuleList()
self.pos_attention_layers = torch.nn.ModuleList()
self.pos_forward_layernorms = torch.nn.ModuleList()
self.pos_forward_layers = torch.nn.ModuleList()
self.neg_attention_layernorms = torch.nn.ModuleList()
self.neg_attention_layers = torch.nn.ModuleList()
self.neg_forward_layernorms = torch.nn.ModuleList()
self.neg_forward_layers = torch.nn.ModuleList()
self.sub_attention_layernorms = torch.nn.ModuleList()
self.sub_attention_layers = torch.nn.ModuleList()
self.sub_forward_layernorms = torch.nn.ModuleList()
self.sub_forward_layers = torch.nn.ModuleList()
self.sparse_attention_layers = torch.nn.ModuleList()
self.seqs_embedd = torch.nn.Linear(1, args.hidden_units).cuda()
self.feature1_embedd = torch.nn.Linear(1, args.hidden_units).cuda()
self.feature2_embedd = torch.nn.Linear(1, args.hidden_units).cuda()
self.feature3_embedd = torch.nn.Linear(1, args.hidden_units).cuda()
self.pos_embedd = torch.nn.Linear(1, args.hidden_units).cuda()
self.neg_embedd = torch.nn.Linear(1, args.hidden_units).cuda()
self.sub_embedd = torch.nn.Linear(1, args.hidden_units).cuda()
self.seq_last_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature1_last_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature2_last_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature3_last_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.price_concat_layernorm = torch.nn.LayerNorm(args.hidden_units * 6, eps=1e-8).cuda()
self.price_lstm_layer = torch.nn.LSTM(input_size=args.hidden_units * 6, hidden_size=args.hidden_units,
batch_first=True).cuda()
self.price_final_emd = torch.nn.Linear(args.hidden_units*2, 1).cuda()
self.final_norm = torch.nn.LayerNorm(args.hidden_units*2, eps=1e-8)
self.sparse_lstm_layer = torch.nn.LSTM(input_size=args.hidden_units, hidden_size=args.hidden_units,
batch_first=True).cuda()
self.seqs_final = torch.nn.Linear(args.hidden_units, 1).cuda()
# 2层
for _ in range(args.num_blocks):
# 归一化层
new_attn_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.attention_layernorms.append(new_attn_layernorm)
new_attn_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature1_attention_layernorms.append(new_attn_layernorm)
new_attn_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature2_attention_layernorms.append(new_attn_layernorm)
new_attn_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature3_attention_layernorms.append(new_attn_layernorm)
new_attn_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.pos_attention_layernorms.append(new_attn_layernorm)
new_attn_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.neg_attention_layernorms.append(new_attn_layernorm)
new_attn_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.sub_attention_layernorms.append(new_attn_layernorm)
# attention_layers添加多头注意力层
new_attn_layer = Multihead_attention(num_units=args.hidden_units, num_heads=args.num_heads,
dropout_rate=args.dropout_rate, causality=True)
self.attention_layers.append(new_attn_layer)
new_attn_layer = Multihead_attention(num_units=args.hidden_units, num_heads=args.num_heads,
dropout_rate=args.dropout_rate, causality=True)
self.feature1_attention_layers.append(new_attn_layer)
new_attn_layer = Multihead_attention(num_units=args.hidden_units, num_heads=args.num_heads,
dropout_rate=args.dropout_rate, causality=True)
self.feature2_attention_layers.append(new_attn_layer)
new_attn_layer = Multihead_attention(num_units=args.hidden_units, num_heads=args.num_heads,
dropout_rate=args.dropout_rate, causality=True)
self.feature3_attention_layers.append(new_attn_layer)
new_attn_layer = Multihead_attention(num_units=args.hidden_units, num_heads=args.num_heads,
dropout_rate=args.dropout_rate, causality=True)
self.pos_attention_layers.append(new_attn_layer)
new_attn_layer = Multihead_attention(num_units=args.hidden_units, num_heads=args.num_heads,
dropout_rate=args.dropout_rate, causality=True)
self.neg_attention_layers.append(new_attn_layer)
new_attn_layer = Multihead_attention(num_units=args.hidden_units, num_heads=args.num_heads,
dropout_rate=args.dropout_rate, causality=True)
self.sub_attention_layers.append(new_attn_layer)
# FNN
new_fwd_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.forward_layernorms.append(new_fwd_layernorm)
new_fwd_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature1_forward_layernorms.append(new_fwd_layernorm)
new_fwd_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature2_forward_layernorms.append(new_fwd_layernorm)
new_fwd_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.feature3_forward_layernorms.append(new_fwd_layernorm)
new_fwd_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.pos_forward_layernorms.append(new_fwd_layernorm)
new_fwd_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.neg_forward_layernorms.append(new_fwd_layernorm)
new_fwd_layernorm = torch.nn.LayerNorm(args.hidden_units, eps=1e-8)
self.sub_forward_layernorms.append(new_fwd_layernorm)
new_fwd_layer = PointWiseFeedForward(args.hidden_units, args.dropout_rate)
self.forward_layers.append(new_fwd_layer)
new_fwd_layer = PointWiseFeedForward(args.hidden_units, args.dropout_rate)
self.feature1_forward_layers.append(new_fwd_layer)
new_fwd_layer = PointWiseFeedForward(args.hidden_units, args.dropout_rate)
self.feature2_forward_layers.append(new_fwd_layer)
new_fwd_layer = PointWiseFeedForward(args.hidden_units, args.dropout_rate)
self.feature3_forward_layers.append(new_fwd_layer)
new_fwd_layer = PointWiseFeedForward(args.hidden_units, args.dropout_rate)
self.pos_forward_layers.append(new_fwd_layer)
new_fwd_layer = PointWiseFeedForward(args.hidden_units, args.dropout_rate)
self.neg_forward_layers.append(new_fwd_layer)
new_fwd_layer = PointWiseFeedForward(args.hidden_units, args.dropout_rate)
self.sub_forward_layers.append(new_fwd_layer)
# 稀疏注意力
new_spare_att = sparseAttention(nb_head=args.num_heads, size_per_head=args.hidden_units // args.num_heads,
rate=args.sparse_rate, key_size=None)
# 添加稀疏注意力
self.sparse_attention_layers.append(new_spare_att)
def log2feats(self, seqs, feature1, feature2, feature3, pos_seqs, neg_seqs):
global sparse_attention
seqs = self.seqs_embedd(seqs)
feature1 = self.feature1_embedd(feature1)
feature2 = self.feature2_embedd(feature2)
feature3 = self.feature3_embedd(feature3)
pos_seqs = self.pos_embedd(pos_seqs)
neg_seqs = self.neg_embedd(neg_seqs)
senti_emb_concat = torch.cat((pos_seqs, neg_seqs), -1)
for i in range(len(self.attention_layers)):
seqs_Q = self.attention_layernorms[i](seqs)
feature1_Q = self.feature1_attention_layernorms[i](feature1)
feature2_Q = self.feature2_attention_layernorms[i](feature2)
feature3_Q = self.feature3_attention_layernorms[i](feature3)
pos_seqs_Q = self.pos_attention_layernorms[i](pos_seqs)
neg_seqs_Q = self.neg_attention_layernorms[i](neg_seqs)
seqs_ = self.attention_layers[i](seqs_Q, seqs, seqs)
feature1 = self.feature1_attention_layers[i](feature1_Q, seqs, seqs)
feature2 = self.feature1_attention_layers[i](feature2_Q, seqs, seqs)
feature3 = self.feature1_attention_layers[i](feature3_Q, seqs, seqs)
pos_seqs = self.pos_attention_layers[i](pos_seqs_Q, seqs, seqs)
neg_seqs = self.neg_attention_layers[i](neg_seqs_Q, seqs, seqs)
seqs = self.forward_layernorms[i](seqs_)
seqs = self.forward_layers[i](seqs)
feature1 = self.feature1_forward_layernorms[i](feature1)
feature1 = self.feature1_forward_layers[i](feature1)
feature2 = self.feature2_forward_layernorms[i](feature2)
feature2 = self.feature2_forward_layers[i](feature2)
feature3 = self.feature3_forward_layernorms[i](feature3)
feature3 = self.feature3_forward_layers[i](feature3)
pos_seqs = self.pos_forward_layernorms[i](pos_seqs)
pos_seqs = self.pos_forward_layers[i](pos_seqs)
neg_seqs = self.neg_forward_layernorms[i](neg_seqs)
neg_seqs = self.neg_forward_layers[i](neg_seqs)
sparse_attention = self.sparse_attention_layers[i](senti_emb_concat)
log_feats = self.seq_last_layernorm(seqs)
log_feature1 = self.feature1_last_layernorm(feature1)
log_feature2 = self.feature1_last_layernorm(feature2)
log_feature3 = self.feature1_last_layernorm(feature3)
log_price = torch.cat((log_feats,log_feature1,log_feature2,log_feature3,pos_seqs,neg_seqs),-1)
log_price = self.price_concat_layernorm(log_price)
p_ge_price, (p_ge_ht, p_ge_ct) = self.price_lstm_layer(log_price)
p_ge_states, (p_ge_ht, p_ge_ct) = self.sparse_lstm_layer(sparse_attention)
log_feats = torch.cat((p_ge_price, p_ge_states), -1)
log_feats = self.final_norm(log_feats)
log_feats = self.price_final_emd(log_feats)
seqs = self.seqs_final(seqs)
return log_feats, seqs
def forward(self, seqs, feature1, feature2, feature3, pos_seqs, neg_seqs):
seqs = torch.Tensor(seqs).cuda()
feature1, feature2, feature3 = torch.Tensor(feature1).cuda(),torch.Tensor(feature2).cuda(),torch.Tensor(feature3).cuda()
pos_seqs, neg_seqs = torch.Tensor(pos_seqs).cuda(), torch.Tensor(neg_seqs).cuda()
log_feats, seqs = self.log2feats(seqs, feature1, feature2, feature3, pos_seqs, neg_seqs)
return log_feats, seqs
def predict(self, seqs, feature1, feature2, feature3, pos_seqs, neg_seqs):
seqs = torch.Tensor(seqs).cuda()
feature1, feature2, feature3 = torch.Tensor(feature1).cuda(), torch.Tensor(feature2).cuda(), torch.Tensor(feature3).cuda()
pos_seqs, neg_seqs = torch.Tensor(pos_seqs).cuda(), torch.Tensor(neg_seqs).cuda()
log_feats, seqs = self.log2feats(seqs, feature1, feature2, feature3, pos_seqs, neg_seqs)
return log_feats, seqs