File size: 6,809 Bytes
a3d0be2 |
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 |
from transformers import BertConfig,BertModel
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Function
from huggingface_hub import PyTorchModelHubMixin
class CSIBERT(nn.Module):
def __init__(self,bertconfig,input_dim,carrier_attention=False, time_emb=True):
super().__init__()
self.bertconfig=bertconfig
self.auto_pos=time_emb
self.bert=BertModel(bertconfig)
self.hidden_dim=bertconfig.hidden_size
self.input_dim=input_dim
self.carrier_attention=carrier_attention
if carrier_attention:
self.attention = SelfAttention(bertconfig.max_position_embeddings, 128, input_dim)
self.emb=nn.Sequential(
nn.Linear(input_dim, 64),
nn.ReLU(),
nn.Linear(64, self.hidden_dim)
)
else:
self.emb=nn.Sequential(
nn.Linear(input_dim, 64),
nn.ReLU(),
nn.Linear(64, self.hidden_dim)
)
def forward(self,x,attn_mask=None,timestamp=None):
if self.carrier_attention:
x=x.permute(0,2,1)
attn_mat = self.attention(x)
x = torch.bmm(attn_mat, x)
x = x.permute(0, 2, 1)
x=self.emb(x)
if timestamp is not None:
pos_emb=self.positional_embedding(timestamp)
x=x+pos_emb
y=self.bert(inputs_embeds=x,attention_mask=attn_mask, output_hidden_states=True)
y=y.hidden_states[-1]
return y
def mask(self,batch_size=1,min=None,max=None,std=None,avg=None):
if std is not None and avg is not None:
device=std.device
result=torch.randn((batch_size, self.bertconfig.max_position_embeddings ,self.input_dim)).to(device)
result=result*std+avg
else:
result=torch.rand((batch_size, self.bertconfig.max_position_embeddings ,self.input_dim))
if min is not None and max is not None:
device = max.device
result=result.to(device)
result=result*(max-min)+min
return result
def positional_embedding(self,timestamp,t=1):
timestamp**=t
device=timestamp.device
min=torch.min(timestamp,dim=-1,keepdim=True)[0]
max=torch.max(timestamp,dim=-1,keepdim=True)[0]
ran=timestamp.shape[-1]
timestamp=(timestamp-min)/(max-min)*ran
d_model=self.hidden_dim
dim=torch.tensor(list(range(d_model))).to(device)
batch_size,length=timestamp.shape
timestamp=timestamp.unsqueeze(2).repeat(1, 1, d_model)
dim=dim.reshape([1,1,-1]).repeat(batch_size,length,1)
sin_emb = torch.sin(timestamp/10000**(dim//2*2/d_model))
cos_emb = torch.cos(timestamp/10000**(dim//2*2/d_model))
mask=torch.zeros(d_model).to(device)
mask[::2]=1
emb=sin_emb*mask+cos_emb*(1-mask)
return emb
class Token_Classifier(nn.Module):
def __init__(self,bert,class_num=52):
super().__init__()
self.bert=bert
self.classifier=nn.Sequential(
nn.Linear(bert.hidden_dim, 64),
nn.ReLU(),
nn.Linear(64, class_num)
)
def forward(self,x,attn_mask=None,timestamp=None):
x=self.bert(x,attn_mask,timestamp)
x=self.classifier(x)
return x
# GRL
class GRL(Function):
@staticmethod
def forward(ctx, x, alpha=1):
ctx.alpha = alpha
return x.view_as(x)
@staticmethod
def backward(ctx, grad_output):
output = grad_output.neg() * ctx.alpha
return output, None
class Sequence_Classifier(nn.Module):
def __init__(self,bert,class_num=6):
super().__init__()
self.bert=bert
self.query = nn.Linear(bert.hidden_dim, 64)
self.key = nn.Linear(bert.hidden_dim, 64)
self.value = nn.Linear(bert.hidden_dim, 64)
self.self_attention = nn.MultiheadAttention(embed_dim=64, num_heads=4, dropout=0, batch_first=True)
self.norm1=nn.BatchNorm1d(64)
self.Linear=nn.Linear(64, 64)
self.norm2 = nn.BatchNorm1d(bert.bertconfig.max_position_embeddings * 64)
self.classifier=nn.Sequential(
nn.Linear(bert.bertconfig.max_position_embeddings * 64, 64),
nn.ReLU(),
nn.Linear(64, class_num)
)
self.GRL = GRL()
def forward(self,x,attn_mask=None,timestamp=None,adversarial=False,alpha=1):
x=self.bert(x,attn_mask,timestamp)
if adversarial:
x = self.GRL.apply(x,alpha)
batch_size,length,hidden_dim=x.shape
x_attn, _ = self.self_attention(self.query(x), self.key(x), self.value(x))
x = x + x_attn
x1 = x.reshape(-1, 64)
x1 = self.norm1(x1)
x1 = self.Linear(x1)
x2 = x1.reshape(batch_size, -1)
x2 = self.norm2(x2)
x2=self.classifier(x2)
return x2
class SelfAttention(nn.Module):
def __init__(self, input_dim, da, r):
super().__init__()
self.ws1 = nn.Linear(input_dim, da, bias=False)
self.ws2 = nn.Linear(da, r, bias=False)
def forward(self, h):
attn_mat = F.softmax(self.ws2(torch.tanh(self.ws1(h))), dim=1)
attn_mat = attn_mat.permute(0, 2, 1)
return attn_mat
class Classification(nn.Module):
def __init__(self, csibert, class_num, hs=64, da=128, r=4):
super().__init__()
self.bert = csibert
self.attention = SelfAttention(hs, da, r)
self.classifier = nn.Sequential(
nn.Linear(hs * r, 256),
nn.ReLU(),
nn.Linear(256, class_num)
)
self.GRL = GRL()
def forward(self, x, attn=None, timestamp=None,adversarial=False):
x = self.bert(x, attn, timestamp)
if adversarial:
x = self.GRL.apply(x)
attn_mat = self.attention(x)
m = torch.bmm(attn_mat, x)
flatten = m.view(m.size()[0], -1)
res = self.classifier(flatten)
return res
class CSI_BERT( nn.Module,
PyTorchModelHubMixin
):
def __init__(self, max_len=100, hs=64, layers=4, heads=4, intermediate_size=128, carrier_dim=52, carrier_attn=False, time_embedding=True):
super().__init__()
self.config = BertConfig(max_position_embeddings=max_len, hidden_size=hs, num_hidden_layers=layers,num_attention_heads=heads, intermediate_size=intermediate_size)
self.model = CSIBERT(self.config,carrier_dim,carrier_attn,time_embedding)
def forward(self, x, attn_mask=None, timestamp=None):
return self.model(x,attn_mask,timestamp) |