Upload model.py
Browse files
model.py
ADDED
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import BertModel,BertConfig
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from huggingface_hub import PyTorchModelHubMixin
|
6 |
+
|
7 |
+
time_gap=10000.0
|
8 |
+
|
9 |
+
class CSIBERT(nn.Module):
|
10 |
+
def __init__(self,bertconfig, input_dim):
|
11 |
+
super().__init__()
|
12 |
+
self.bertconfig=bertconfig
|
13 |
+
self.bert=BertModel(bertconfig)
|
14 |
+
self.hidden_dim=bertconfig.hidden_size
|
15 |
+
self.input_dim=input_dim
|
16 |
+
self.len=bertconfig.max_position_embeddings
|
17 |
+
|
18 |
+
self.Norm1 = nn.LayerNorm(self.input_dim)
|
19 |
+
self.Norm2 = nn.LayerNorm(self.hidden_dim)
|
20 |
+
self.Norm3 = nn.LayerNorm(self.hidden_dim)
|
21 |
+
|
22 |
+
self.csi_emb=nn.Sequential(
|
23 |
+
nn.Linear(input_dim, input_dim),
|
24 |
+
nn.ReLU(),
|
25 |
+
nn.Linear(input_dim, self.hidden_dim),
|
26 |
+
nn.ReLU(),
|
27 |
+
nn.Linear(self.hidden_dim, self.hidden_dim)
|
28 |
+
)
|
29 |
+
|
30 |
+
self.time_emb=nn.Sequential(
|
31 |
+
nn.Linear(input_dim, input_dim),
|
32 |
+
nn.ReLU(),
|
33 |
+
nn.Linear(input_dim, self.hidden_dim),
|
34 |
+
nn.ReLU(),
|
35 |
+
nn.Linear(self.hidden_dim, self.hidden_dim)
|
36 |
+
)
|
37 |
+
|
38 |
+
self.fusion_emb=nn.Sequential(
|
39 |
+
nn.Linear(self.hidden_dim*2, self.hidden_dim*2),
|
40 |
+
nn.ReLU(),
|
41 |
+
nn.Linear(self.hidden_dim*2, self.hidden_dim),
|
42 |
+
nn.ReLU(),
|
43 |
+
nn.Linear(self.hidden_dim, self.hidden_dim)
|
44 |
+
)
|
45 |
+
|
46 |
+
self.arl = nn.Sequential(
|
47 |
+
nn.Linear(self.len, self.len // 2),
|
48 |
+
nn.ReLU(),
|
49 |
+
nn.Linear(self.len // 2, self.len // 4),
|
50 |
+
nn.ReLU(),
|
51 |
+
nn.Linear(self.len // 4, 1)
|
52 |
+
)
|
53 |
+
|
54 |
+
def forward(self,x,timestamp,attention_mask=None):
|
55 |
+
x=x.to(torch.float32)
|
56 |
+
|
57 |
+
x=self.attention(x)
|
58 |
+
x=self.csi_emb(x)
|
59 |
+
x_time=self.time_embedding(timestamp)
|
60 |
+
x = x + x_time
|
61 |
+
y=self.bert(inputs_embeds=x, attention_mask=attention_mask, output_hidden_states=False)
|
62 |
+
y=y.last_hidden_state
|
63 |
+
return y
|
64 |
+
|
65 |
+
def time_embedding(self,timestamp,t=1):
|
66 |
+
device=timestamp.device
|
67 |
+
# timestamp = (timestamp - timestamp[:,0:1]) / time_gap
|
68 |
+
# timestamp = (timestamp - timestamp[:, 0:1]) / (timestamp[:,-1:] - timestamp[:, 0:1])
|
69 |
+
timestamp = (timestamp - timestamp[:, 0:1]) / (timestamp[:,-1:] - timestamp[:, 0:1]) * self.len
|
70 |
+
|
71 |
+
timestamp**=t
|
72 |
+
d_model=self.input_dim
|
73 |
+
dim=torch.tensor(list(range(d_model))).to(device)
|
74 |
+
batch_size,length=timestamp.shape
|
75 |
+
timestamp=timestamp.unsqueeze(2).repeat(1, 1, d_model)
|
76 |
+
dim=dim.reshape([1,1,-1]).repeat(batch_size,length,1)
|
77 |
+
sin_emb = torch.sin(timestamp/10000**(dim//2*2/d_model))
|
78 |
+
cos_emb = torch.cos(timestamp/10000**(dim//2*2/d_model))
|
79 |
+
mask=torch.zeros(d_model).to(device)
|
80 |
+
mask[::2]=1
|
81 |
+
emb=sin_emb*mask+cos_emb*(1-mask)
|
82 |
+
emb=self.time_emb(emb)
|
83 |
+
|
84 |
+
# timestamp = torch.unsqueeze(timestamp, -1)
|
85 |
+
# emb=self.time_emb(timestamp)
|
86 |
+
|
87 |
+
return emb
|
88 |
+
|
89 |
+
# def attention(self,x):
|
90 |
+
# y = torch.transpose(x, -1, -2)
|
91 |
+
# batch_size = y.shape[0]
|
92 |
+
# queries = self.query(y).view(batch_size, -1, self.head_num, self.head_dim).transpose(1, 2)
|
93 |
+
# keys = self.key(y).view(batch_size, -1, self.head_num, self.head_dim).transpose(1, 2)
|
94 |
+
# values = self.value(y).view(batch_size, -1, self.head_num, self.head_dim).transpose(1, 2)
|
95 |
+
# attention_weights = self.softmax(torch.matmul(queries, keys.transpose(-1, -2))/ (self.head_dim ** 0.5))
|
96 |
+
#
|
97 |
+
# # attended_values = torch.matmul(attention_weights,values).transpose(1, 2)
|
98 |
+
# # attended_values = attended_values.reshape(batch_size,self.input_dim,self.len)
|
99 |
+
# # attended_values = self.norm(attended_values)
|
100 |
+
# # y = attended_values.transpose(1, 2)
|
101 |
+
#
|
102 |
+
# attended_values = torch.matmul(attention_weights, values).transpose(-1, -2)
|
103 |
+
# attended_values = attended_values.reshape(batch_size, self.len, self.input_dim)
|
104 |
+
# y = self.norm(attended_values)
|
105 |
+
#
|
106 |
+
# return y+x
|
107 |
+
|
108 |
+
def attention(self, x):
|
109 |
+
y = torch.transpose(x, -1, -2)
|
110 |
+
attn = self.arl(y)
|
111 |
+
y = y * attn
|
112 |
+
y = torch.transpose(y, -1, -2)
|
113 |
+
return y
|
114 |
+
|
115 |
+
class Token_Classifier(nn.Module):
|
116 |
+
def __init__(self,bert,class_num=52):
|
117 |
+
super().__init__()
|
118 |
+
self.bert=bert
|
119 |
+
self.classifier=nn.Sequential(
|
120 |
+
nn.Linear(bert.hidden_dim, bert.hidden_dim//2),
|
121 |
+
nn.ReLU(),
|
122 |
+
nn.Linear(bert.hidden_dim//2, class_num)
|
123 |
+
)
|
124 |
+
|
125 |
+
def forward(self,x,timestamp,attention_mask=None):
|
126 |
+
x=self.bert(x,timestamp,attention_mask=attention_mask)
|
127 |
+
x=self.classifier(x)
|
128 |
+
return x
|
129 |
+
|
130 |
+
class SelfAttention(nn.Module):
|
131 |
+
def __init__(self, input_dim, da, r):
|
132 |
+
super().__init__()
|
133 |
+
self.ws1 = nn.Linear(input_dim, da, bias=False)
|
134 |
+
self.ws2 = nn.Linear(da, r, bias=False)
|
135 |
+
|
136 |
+
def forward(self, h):
|
137 |
+
attn_mat = F.softmax(self.ws2(torch.tanh(self.ws1(h))), dim=1)
|
138 |
+
attn_mat = attn_mat.permute(0, 2, 1)
|
139 |
+
return attn_mat
|
140 |
+
|
141 |
+
|
142 |
+
class Sequence_Classifier(nn.Module):
|
143 |
+
def __init__(self, csibert, class_num, hs=128, da=128, r=4):
|
144 |
+
super().__init__()
|
145 |
+
self.bert = csibert
|
146 |
+
self.attention = SelfAttention(hs, da, r)
|
147 |
+
self.classifier = nn.Sequential(
|
148 |
+
nn.Linear(hs * r, hs * r // 2),
|
149 |
+
nn.ReLU(),
|
150 |
+
nn.Linear(hs * r // 2, class_num)
|
151 |
+
)
|
152 |
+
|
153 |
+
def forward(self, x, timestamp,attention_mask=None):
|
154 |
+
x = self.bert(x, timestamp,attention_mask=attention_mask)
|
155 |
+
attn_mat = self.attention(x)
|
156 |
+
m = torch.bmm(attn_mat, x)
|
157 |
+
flatten = m.view(m.size()[0], -1)
|
158 |
+
res = self.classifier(flatten)
|
159 |
+
return res
|
160 |
+
|
161 |
+
class CSI_BERT2(nn.Module,
|
162 |
+
PyTorchModelHubMixin
|
163 |
+
):
|
164 |
+
def __init__(self, max_len=100, hs=128, layers=6, heads=8, intermediate_size=512, carrier_dim=52):
|
165 |
+
super().__init__()
|
166 |
+
self.config = BertConfig(max_position_embeddings=max_len, hidden_size=hs, num_hidden_layers=layers,num_attention_heads=heads, intermediate_size=intermediate_size)
|
167 |
+
self.model = CSIBERT(self.config,carrier_dim)
|
168 |
+
|
169 |
+
def forward(self, x, timestamp=None, attn_mask=None):
|
170 |
+
return self.model(x,timestamp,attn_mask)
|