-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmodel.py
130 lines (130 loc) · 6.9 KB
/
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from utils import activation_getter
class RNS(nn.Module):
def load_word2vec(self, vocabulary):
word_vec_path = 'data/glove_twitter_27B/glove25d.txt'
with open(word_vec_path, "r", encoding='UTF-8') as f:
lines = f.readlines()
raw_word2vec = {}
for l in lines:
w, vec = l.split(" ", 1)
raw_word2vec[w] = vec
self.word2vec = None
oov_cnt = 0
for word, index in vocabulary.items():
str_vec = raw_word2vec.get(word, None)
if str_vec is None:
oov_cnt += 1
vec = np.random.randn(25) * 0.1
else:
vec = np.fromstring(str_vec, sep=" ")
vec = np.expand_dims(vec, axis=0)
self.word2vec = np.concatenate((self.word2vec, vec), 0) if self.word2vec is not None else vec
print("word2vec cannot cover %f vocabulary" % (float(oov_cnt) / len(vocabulary)))
def __init__(self, num_users, num_items, model_args, u_text, i_text, vocabulary):
super(RNS, self).__init__()
self.args = model_args
L = self.args.L
self.n_t = self.args.nt
self.n_k = self.args.nk
self.drop_ratio = self.args.drop
self.ac_conv = activation_getter[self.args.ac_conv]
self.ac_fc = activation_getter[self.args.ac_fc]
self.num_users = num_users
self.num_items = num_items
self.word2vec = None
self.w_vec_dim = self.args.dim
self.alpha = self.args.alpha
self.word_embed = nn.Embedding(len(vocabulary), self.w_vec_dim)
self.word_embed.weight.data.normal_(0, 1.0 / self.word_embed.embedding_dim)
self.word_embed.weight.requires_grad = True
self.u_text_fake_embed = nn.Embedding(u_text.shape[0], u_text.shape[1])
self.i_text_fake_embed = nn.Embedding(i_text.shape[0], i_text.shape[1])
self.u_text_fake_embed.weight.data.copy_(torch.from_numpy(u_text))
self.u_text_fake_embed.weight.requires_grad = False
self.i_text_fake_embed.weight.data.copy_(torch.from_numpy(i_text))
self.i_text_fake_embed.weight.requires_grad = False
self.trans_mat = [torch.randn((self.w_vec_dim, self.w_vec_dim), dtype=torch.float, requires_grad=True).to('cuda')
for _ in range(self.n_k)]
lengths = [1, 3, 5, 7, 9]
self.conv_t_item = nn.ModuleList([nn.Conv2d(self.n_k, self.n_t, (i, self.w_vec_dim)) for i in lengths])
self.conv_t_user = nn.ModuleList([nn.Conv2d(self.n_k, self.n_t, (i, self.w_vec_dim)) for i in lengths])
self.k_dim = self.n_t * len(lengths)
self.full_dim = self.n_k * self.k_dim
self.pos_embed = nn.Embedding(L, self.k_dim)
self.conv_d = nn.ModuleList([nn.Conv2d(1, 1, (L, 1)) for _ in range(self.n_k)])
self.dropout = nn.Dropout(self.drop_ratio)
def forward(self, seq_var, user_var, item_var, for_pred=False):
seq_var_word_index = self.i_text_fake_embed(seq_var).long()
seq_var_word_vector = self.word_embed(seq_var_word_index)
item_var_word_index = self.i_text_fake_embed(item_var).long()
item_var_word_vector = self.word_embed(item_var_word_index)
user_var_word_index = self.u_text_fake_embed(user_var).long()
user_var_word_vector = self.word_embed(user_var_word_index)
l1, l2, l3, ll1, ll2, ll3 = [list() for _ in range(6)]
seq_var_aspect_concat = None
item_var_aspect_concat = None
user_var_aspect_concat = None
for i in range(self.n_k):
seq_var_word_trans_vector = torch.einsum("abcd,de->abce", (seq_var_word_vector, self.trans_mat[i]))\
.unsqueeze(1)
seq_var_aspect_concat = torch.cat((seq_var_aspect_concat, seq_var_word_trans_vector), 1) \
if seq_var_aspect_concat is not None else seq_var_word_trans_vector
item_var_word_trans_vector = torch.einsum("abcd,de->abce", (item_var_word_vector, self.trans_mat[i])) \
.unsqueeze(1)
item_var_aspect_concat = torch.cat((item_var_aspect_concat, item_var_word_trans_vector), 1) \
if item_var_aspect_concat is not None else item_var_word_trans_vector
user_var_word_trans_vector = torch.einsum("abcd,de->abce", (user_var_word_vector, self.trans_mat[i])) \
.unsqueeze(1)
user_var_aspect_concat = torch.cat((user_var_aspect_concat, user_var_word_trans_vector), 1) \
if user_var_aspect_concat is not None else user_var_word_trans_vector
for j in range(seq_var_aspect_concat.shape[2]):
s = seq_var_aspect_concat[:, :, j, :, :]
for conv in self.conv_t_item:
conv_out = self.ac_conv(conv(s).squeeze(3))
pool_out = F.max_pool1d(conv_out, conv_out.size(2)).squeeze(2)
l1.append(pool_out)
s1 = torch.cat(l1, 1).unsqueeze(1)
l2.append(s1)
l1 = []
seq_var_repr = torch.cat(l2, 1)
for k in range(item_var.shape[1]):
ss = item_var_aspect_concat[:, :, k, :, :]
for conv in self.conv_t_item:
conv_out = self.ac_conv(conv(ss).squeeze(3))
pool_out = F.max_pool1d(conv_out, conv_out.size(2)).squeeze(2)
ll1.append(pool_out)
ss1 = torch.cat(ll1, 1).unsqueeze(1)
ll2.append(ss1)
ll1 = []
item_repr = torch.cat(ll2, 1)
seq_var_repr += self.pos_embed(torch.arange(0, seq_var.shape[1]).to("cuda"))
l4 = []
for conv in self.conv_t_user:
conv_out = self.ac_conv(conv(user_var_aspect_concat.squeeze(2)).squeeze(3))
pool_out = F.max_pool1d(conv_out, conv_out.size(2)).squeeze(2)
l4.append(pool_out)
user_repr = torch.cat(l4, 1)
al = []
for i in range(item_var.shape[1]):
s = item_repr[:, i, :].unsqueeze(1)
w = F.softmax(torch.sum(seq_var_repr * s, 2), 1).unsqueeze(2)
u = torch.sum(seq_var_repr * w, 1).unsqueeze(1)
m = torch.argmax(w, 1).unsqueeze(1)
index = m.expand(-1, 1, seq_var_repr.size(2))
if for_pred:
seq_var_repr = seq_var_repr.expand(item_var.size(0), -1, -1)
p = seq_var_repr.gather(1, index)
p_u = torch.cat((u, p), 1)
w2 = F.softmax(torch.sum(p_u * s, 2), 1).unsqueeze(2)
ss = torch.sum(p_u * w2, 1).unsqueeze(1)
al.append(ss)
seq_repr = torch.cat(al, 1)
if for_pred:
res = torch.sum((self.alpha * seq_repr + user_repr.unsqueeze(1)).squeeze() * item_repr.squeeze(), 1)
else:
res = torch.sum((self.alpha * seq_repr + user_repr.unsqueeze(1)) * item_repr, 2)
return res