-
Notifications
You must be signed in to change notification settings - Fork 1
/
Transformer_torch.py
198 lines (149 loc) · 7.98 KB
/
Transformer_torch.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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.utils.data import Dataset, DataLoader
import math
import copy
import time
import gc
import random
from tqdm import tqdm
import numpy as np
import pandas as pd
from sklearn.metrics import roc_auc_score
# MODEL
# FeedForwardNetwork
class FFN(nn.Module):
def __init__(self, state_size = 200, forward_expansion = 1, bn_size=100, dropout=0.2):
super(FFN, self).__init__()
self.state_size = state_size
self.lr1 = nn.Linear(state_size, forward_expansion * state_size)
self.relu = nn.ReLU()
self.bn = nn.BatchNorm1d(bn_size)
self.lr2 = nn.Linear(forward_expansion * state_size, state_size)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
x = self.relu(self.lr1(x))
x = self.bn(x)
x = self.lr2(x)
return self.dropout(x)
FFN()
# Mask
def future_mask(seq_length):
future_mask = (np.triu(np.ones([seq_length, seq_length]), k = 1)).astype('bool')
return torch.from_numpy(future_mask)
future_mask(5)
class TransformerBlock_en(nn.Module):
def __init__(self, embed_dim, heads = 4, MAX_SEQ = 100, dropout = 0.1, forward_expansion = 1):
super(TransformerBlock_en, self).__init__()
self.multi_att = nn.MultiheadAttention(embed_dim=embed_dim, num_heads=heads, dropout=dropout)
self.dropout = nn.Dropout(dropout)
self.layer_normal_q = nn.LayerNorm(embed_dim)
self.layer_normal_k = nn.LayerNorm(embed_dim)
self.layer_normal_v = nn.LayerNorm(embed_dim)
self.ffn = FFN(embed_dim, forward_expansion = forward_expansion, bn_size=MAX_SEQ-1, dropout=dropout)
self.layer_normal_2 = nn.LayerNorm(embed_dim)
def forward(self, query, key, value, att_mask):
query = self.layer_normal_q(query)
key = self.layer_normal_k(key)
value = self.layer_normal_v(value)
att_output, att_weight = self.multi_att(query, key, value, attn_mask=att_mask)
att_output = self.dropout(att_output + query)
att_output = att_output.permute(1, 0, 2) # att_output: [s_len, bs, embed] => [bs, s_len, embed]
att_output = self.layer_normal_2(att_output)
x = self.ffn(att_output)
x = self.dropout(x + att_output)
return x.squeeze(-1), att_weight
class TransformerBlock_de(nn.Module):
def __init__(self, embed_dim = 256, heads_de = 4, MAX_SEQ = 100, dropout = 0.1, forward_expansion = 1):
super(TransformerBlock_de, self).__init__()
self.multi_att_1 = nn.MultiheadAttention(embed_dim=embed_dim, num_heads=heads_de, dropout=dropout)
self.multi_att_2 = nn.MultiheadAttention(embed_dim=embed_dim, num_heads=heads_de, dropout=dropout)
self.dropout_1 = nn.Dropout(dropout)
self.dropout_2 = nn.Dropout(dropout)
self.dropout_3 = nn.Dropout(dropout)
self.layer_normal_de_in = nn.LayerNorm(embed_dim)
self.layer_normal_en_out = nn.LayerNorm(embed_dim)
self.layer_normal_de_out = nn.LayerNorm(embed_dim)
self.layer_normal_1 = nn.LayerNorm(embed_dim)
self.ffn = FFN(embed_dim, forward_expansion = forward_expansion, bn_size=MAX_SEQ-1, dropout=dropout)
def forward(self, de_in, en_out, att_mask):
de_in = self.layer_normal_de_in(de_in)
att_output, att_weight = self.multi_att_1(de_in, de_in, de_in, attn_mask=att_mask)
att_output = self.dropout_1(att_output + de_in)
en_out = self.layer_normal_en_out(en_out)
att_output = self.layer_normal_de_out(att_output)
en_output, en_weight = self.multi_att_2(att_output, en_out, en_out, attn_mask=att_mask)
en_output = self.dropout_2(en_output + att_output)
en_output = en_output.permute(1, 0, 2) # att_output: [s_len, bs, embed] => [bs, s_len, embed]
en_output = self.layer_normal_1(en_output)
x = self.ffn(en_output)
x = self.dropout_3(x+en_output)
return x.squeeze(-1), att_weight
class Encoder(nn.Module):
def __init__(self, total_ex, total_cat, embed_dim, heads_en, max_seq, dropout, forward_expansion, num_layers):
super(Encoder, self).__init__()
self.embed_dim = embed_dim
self.embedding_id = nn.Embedding(total_ex , embed_dim)
self.pos_embedding = nn.Embedding(max_seq - 1, embed_dim)
self.embedding_part = nn.Embedding(total_cat + 1, embed_dim)
self.layers = nn.ModuleList([TransformerBlock_en(embed_dim, heads = heads_en, MAX_SEQ = max_seq, dropout = dropout, forward_expansion = forward_expansion) for _ in range(num_layers)])
self.dropout = nn.Dropout(dropout)
def forward(self, e_id, part_id):
device = e_id.device
e_id = self.embedding_id(e_id)
pos_id = torch.arange(e_id.size(1)).unsqueeze(0).to(device)
pos_x = self.pos_embedding(pos_id)
part_id = self.embedding_part(part_id)
x = self.dropout(e_id + part_id + pos_x)
x = x.permute(1, 0, 2) # x: [bs, s_len, embed] => [s_len, bs, embed]
for layer in self.layers:
att_mask = future_mask(x.size(0)).to(device)
x, att_weight = layer(x, x, x, att_mask=att_mask)
x = x.permute(1, 0, 2)
x = x.permute(1, 0, 2)
return x, att_weight
class Decoder(nn.Module):
def __init__(self, total_in , total_task , total_lag , total_p , heads_de, max_seq, embed_dim, dropout, forward_expansion, num_layers):
super(Decoder, self).__init__()
self.embed_dim = embed_dim
self.embedding_in = nn.Embedding(total_in, embed_dim)
self.embedding_task = nn.Embedding(total_task, embed_dim)
self.embedding_lag = nn.Embedding(total_lag, embed_dim)
#self.embedding_p = nn.Embedding(total_p, embed_dim)
self.pos_embedding = nn.Embedding(max_seq - 1, embed_dim)
self.p_time_con = nn.Linear(1, embed_dim, bias=False)
#self.embedding_part = nn.Embedding(total_cat+1, embed_dim)
self.layers = nn.ModuleList([TransformerBlock_de(embed_dim, heads_de = heads_de, MAX_SEQ = max_seq, dropout = dropout, forward_expansion = forward_expansion) for _ in range(num_layers)])
self.dropout = nn.Dropout(dropout)
def forward(self, de_id, de_task, de_time, p_time, en_out):
device = de_id.device
de_in = self.embedding_in(de_id)
#de_task = self.embedding_task(de_task)
p_time = p_time.unsqueeze(0).permute(1, 2, 0).float()
de_time = self.embedding_lag(de_time)
p_time = self.p_time_con(p_time)
pos_id = torch.arange(de_in.size(1)).unsqueeze(0).to(device)
pos_x = self.pos_embedding(pos_id)
x = self.dropout(de_in + de_time + p_time + pos_x)
x = x.permute(1, 0, 2) # x: [bs, s_len, embed] => [s_len, bs, embed]
en_out = en_out.permute(1, 0, 2)
for layer in self.layers:
att_mask = future_mask(x.size(0)).to(device)
x, att_weight = layer(x, en_out, att_mask=att_mask)
x = x.permute(1, 0, 2)
x = x.permute(1, 0, 2)
return x, att_weight
class TransformerModel(nn.Module):
def __init__(self, total_ex, total_cat, total_in, total_task, total_lag, total_p, embed_dim, heads_en, heads_de, max_seq, dropout, forward_expansion = 1, enc_layers=3, dec_layers=3):
super(TransformerModel, self).__init__()
self.encoder = Encoder(total_ex, total_cat, embed_dim, heads_en, max_seq, dropout, forward_expansion, num_layers=enc_layers)
self.decoder = Decoder(total_in, total_task, total_lag, total_p, heads_de, max_seq, embed_dim, dropout, forward_expansion, num_layers=dec_layers)
self.pred = nn.Linear(embed_dim, 1)
def forward(self, e_id, part_id, de_task, de_time, p_time, de_in,):
en_y, att_weight = self.encoder(e_id, part_id)
x, att_weight_de = self.decoder(de_in, de_task, de_time, p_time, en_y)
x = self.pred(x)
return x.squeeze(-1), att_weight