/
Nets.py
159 lines (94 loc) · 5.07 KB
/
Nets.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as I
class EmbedAttention(nn.Module):
def __init__(self, att_size):
super(EmbedAttention, self).__init__()
self.att_w = nn.Linear(att_size,1,bias=False)
def forward(self,input,len_s):
att = self.att_w(input).squeeze(-1)
out = self._masked_softmax(att,len_s).unsqueeze(-1)
return out
def _masked_softmax(self,mat,len_s):
#print(len_s.type())
len_s = len_s.type_as(mat.data)#.long()
idxes = torch.arange(0,int(len_s[0]),out=mat.data.new(int(len_s[0])).long()).unsqueeze(1)
mask = (idxes.float()<len_s.unsqueeze(0)).float()
exp = torch.exp(mat) * mask
sum_exp = exp.sum(0,True)+0.0001
return exp/sum_exp.expand_as(exp)
class AttentionalBiRNN(nn.Module):
def __init__(self, inp_size, hid_size, dropout=0, RNN_cell=nn.GRU):
super(AttentionalBiRNN, self).__init__()
self.natt = hid_size*2
self.rnn = RNN_cell(input_size=inp_size,hidden_size=hid_size,num_layers=1,bias=True,batch_first=True,dropout=dropout,bidirectional=True)
self.lin = nn.Linear(hid_size*2,self.natt)
self.att_w = nn.Linear(self.natt,1,bias=False)
self.emb_att = EmbedAttention(self.natt)
def forward(self, packed_batch):
rnn_sents,_ = self.rnn(packed_batch)
enc_sents,len_s = torch.nn.utils.rnn.pad_packed_sequence(rnn_sents)
emb_h = F.tanh(self.lin(enc_sents))
attended = self.emb_att(emb_h,len_s) * enc_sents
return attended.sum(0,True).squeeze(0)
class UIAttentionalBiRNN(AttentionalBiRNN):
def __init__(self, inp_size, hid_size, dropout=0, RNN_cell=nn.LSTM):
super(UIAttentionalBiRNN, self).__init__(inp_size, hid_size, dropout, RNN_cell)
self.register_buffer("mask",torch.FloatTensor())
self.att_h = nn.Linear(inp_size*2+self.natt,self.natt,bias=True)
def forward(self, packed_batch,user_embs,item_embs):
rnn_sents,_ = self.rnn(packed_batch)
enc_sents,len_s = torch.nn.utils.rnn.pad_packed_sequence(rnn_sents)
uit = torch.cat([user_embs.expand_as(enc_sents),item_embs.expand_as(enc_sents),enc_sents],dim=-1)
summed = F.tanh(self.att_h(uit))
return torch.sum(enc_sents * self.emb_att(summed,len_s),0)
class HAN(nn.Module):
def __init__(self, ntoken, num_class, emb_size=200, hid_size=50):
super(HAN, self).__init__()
self.emb_size = emb_size
self.embed = nn.Embedding(ntoken, emb_size,padding_idx=0)
self.word = AttentionalBiRNN(emb_size, hid_size)
self.sent = AttentionalBiRNN(hid_size*2, hid_size)
self.lin_out = nn.Linear(hid_size*2,num_class)
def set_emb_tensor(self,emb_tensor):
self.emb_size = emb_tensor.size(-1)
self.embed.weight.data = emb_tensor
def _reorder_sent(self,sents,sent_order):
sents = F.pad(sents,(0,0,1,0)) #adds a 0 to the top
revs = sents[sent_order.view(-1)]
revs = revs.view(sent_order.size(0),sent_order.size(1),sents.size(1))
return revs
def forward(self, batch_reviews,sent_order,ls,lr):
emb_w = F.dropout(self.embed(batch_reviews),training=self.training)
packed_sents = torch.nn.utils.rnn.pack_padded_sequence(emb_w, ls,batch_first=True)
sent_embs = self.word(packed_sents)
rev_embs = self._reorder_sent(sent_embs,sent_order)
packed_rev = torch.nn.utils.rnn.pack_padded_sequence(rev_embs, lr,batch_first=True)
doc_embs = self.sent(packed_rev)
out = self.lin_out(doc_embs)
return out
class NSCUPA(HAN):
def __init__(self, ntoken, nusers, nitems, num_class, emb_size=200, hid_size=100):
super(NSCUPA, self).__init__(ntoken, num_class, emb_size, hid_size)
self.users = nn.Embedding(nusers, emb_size)
I.normal(self.users.weight.data,0.01,0.01)
self.items = nn.Embedding(nitems, emb_size)
I.normal(self.items.weight.data,0.01,0.01)
self.word = UIAttentionalBiRNN(emb_size, emb_size//2)
self.sent = UIAttentionalBiRNN(emb_size, emb_size//2)
def forward(self, batch_reviews,users,items,sent_order,ui_indexs,ls,lr):
u = users[ui_indexs]
i = items[ui_indexs]
emb_w = F.dropout(self.embed(batch_reviews),training=self.training)
emb_u = F.dropout(self.users(u),training=self.training)
emb_i = F.dropout(self.items(i),training=self.training)
packed_sents = torch.nn.utils.rnn.pack_padded_sequence(emb_w, ls,batch_first=True)
sent_embs = self.word(packed_sents,emb_u,emb_i)
rev_embs = self._reorder_sent(sent_embs,sent_order)
packed_rev = torch.nn.utils.rnn.pack_padded_sequence(rev_embs, lr,batch_first=True)
emb_u = F.dropout(self.users(users),training=self.training)
emb_i = F.dropout(self.items(items),training=self.training)
doc_embs = self.sent(packed_rev,emb_u,emb_i)
out = self.lin_out(doc_embs)
return out