-
Notifications
You must be signed in to change notification settings - Fork 2
/
DPJFMBS.py
216 lines (185 loc) · 10.3 KB
/
DPJFMBS.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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
import torch
import torch.nn.functional as F
from torch.autograd import Variable
from torch.nn.utils.rnn import pack_padded_sequence
import numpy as np
import os
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
WORD_EMBED_DIM = 32
USER_EMBED_DIM = 32
MAX_PROFILELEN = 20
MAX_TERMLEN = 50
MEMORY_SLOT = 5
class Model(torch.nn.Module):
def __init__(self, word_embeddings, behavior_typenum):
super(Model, self).__init__()
self.behavior_typenum = behavior_typenum
print('MEMORY_SLOT=', MEMORY_SLOT)
# profile: word embeddings for look_up
# embedding_matrix = [[0...0], [...], ...[]]
self.word_embeddings = torch.nn.Embedding.from_pretrained(word_embeddings, padding_idx=0)
self.word_embeddings.weight.requires_grad = False
# bilstm: int(USER_EMBED_DIM/2) * 2 = USER_EMBED_DIM
self.words_gru = torch.nn.GRU(input_size=WORD_EMBED_DIM, hidden_size=int(USER_EMBED_DIM/2),
num_layers=1, batch_first=True, bidirectional=True, dropout=0.)
# attention layer
self.words_attention_layer = torch.nn.Sequential(
torch.nn.Linear(USER_EMBED_DIM, USER_EMBED_DIM),
torch.nn.Tanh(), # LeakyReLU
torch.nn.Linear(USER_EMBED_DIM, 1, bias=False),
)
# attention layer
self.terms_attention_layer = torch.nn.Sequential(
torch.nn.Linear(USER_EMBED_DIM, USER_EMBED_DIM),
torch.nn.Tanh(),
torch.nn.Linear(USER_EMBED_DIM, 1, bias=False),
)
self.KEY_MATs = []
for i in range(self.behavior_typenum):
# [MEMORY_SLOT, USER_EMBED_DIM]
self.KEY_MATs.append(Variable(torch.randn((MEMORY_SLOT, USER_EMBED_DIM)), requires_grad = True).float().to(DEVICE))
self.update_mlp = torch.nn.Sequential(
torch.nn.Linear(USER_EMBED_DIM, USER_EMBED_DIM),
torch.nn.Sigmoid(),
)
self.e2j_MTL_MLPs = []
for i in range(self.behavior_typenum-1):
self.e2j_MTL_MLPs.append(torch.nn.Sequential(
torch.nn.Linear( 4 *USER_EMBED_DIM, USER_EMBED_DIM),
torch.nn.Tanh(),
torch.nn.Linear( USER_EMBED_DIM, 1),
torch.nn.Sigmoid()).to(DEVICE))
self.j2e_MTL_MLPs = []
for i in range(self.behavior_typenum-1):
self.j2e_MTL_MLPs.append(torch.nn.Sequential(
torch.nn.Linear( 4 *USER_EMBED_DIM, USER_EMBED_DIM),
torch.nn.Tanh(),
torch.nn.Linear( USER_EMBED_DIM, 1),
torch.nn.Sigmoid()).to(DEVICE))
self.match_MLP = torch.nn.Sequential(
torch.nn.Linear( 4 *USER_EMBED_DIM, USER_EMBED_DIM),
torch.nn.Tanh(),
torch.nn.Linear( USER_EMBED_DIM, 1),
torch.nn.Sigmoid()).to(DEVICE)
self.e2j_MTL_MLPs.append(self.match_MLP)
self.j2e_MTL_MLPs.append(self.match_MLP)
# profiles: [batch_size, MAX_PROFILELEN, MAX_TERMLEN] = (40, 15, 50), word idx
# return: [batch_size, USER_EMBED_DIM]
def get_profile_embeddings(self, profiles, isexpect):
# word level:
# [batch_size, MAX_PROFILELEN, MAX_TERMLEN] (40, 15, 50) ->
# [batch_size * MAX_PROFILELEN, MAX_TERMLEN](40 * 15, 50)
shape = profiles.shape
profiles_ = profiles.view([-1, shape[-1]])
# sort expects_sample_: large to small
# sorted [batch_size * MAX_PROFILELEN, MAX_TERMLEN](40 * 15, 50)
lens = (profiles_ > 0).sum(dim=-1)
lens_sort, ind_sort = lens.sort(dim=0, descending=True)
profiles_sort = profiles_[ind_sort]
# embeddings: [batch_size * MAX_PROFILELEN, MAX_TERMLEN, EMBED_DIM]
profile_embed = self.word_embeddings(profiles_sort).float()
# compress: [batch_size * MAX_PROFILELEN, MAX_TERMLEN, EMBED_DIM]
profile_pack = pack_padded_sequence(profile_embed, lens_sort, batch_first=True)
words_output, _ = self.words_gru(profile_pack)
# output: [batch_size * MAX_PROFILELEN, MAX_TERMLEN, words_lstm_hidden_dims * 2]
words_output_, _ = torch.nn.utils.rnn.pad_packed_sequence(words_output, batch_first=True)
# attention: [batch_size * MAX_PROFILELEN, MAX_TERMLEN, words_lstm_hidden_dims * 2]
words_attention = F.softmax(self.words_attention_layer(words_output_), dim=-2)
# [batch_size * MAX_PROFILELEN, MAX_TERMLEN, words_lstm_hidden_dims * 2] ->
# [batch_size * MAX_PROFILELEN, words_lstm_hidden_dims * 2]
terms_ = (words_attention * words_output_).sum(-2)\
.view([shape[0], MAX_PROFILELEN, USER_EMBED_DIM])
# (batch, 1, USER_EMBED_DIM) * (batch_size, MAX_PROFILELEN, hidden_size) ->
# (batch, MAX_PROFILELEN, USER_EMBED_DIM) -> softmax(batch, MAX_PROFILELEN, 1)
attention = torch.softmax(self.terms_attention_layer(terms_), dim=-2)
# (batch, MAX_PROFILELEN, 1) * (batch_size, MAX_PROFILELEN, hidden_size) ->
# (batch, hidden_size)
profile_embeddings = torch.sum(attention * terms_, dim=1)
return profile_embeddings
# memory: [batch, MEMORY_SLOT, USER_EMBED_DIM]
# b_embedding： [batch, USER_EMBED_DIM]
# col_label: []
# return: [batch, USER_EMBED_DIM]
def read(self, memory, b_embedding, col_label):
preference = None
for i in range(self.behavior_typenum):
# [batch, USER_EMBED_DIM] mm [MEMORY_SLOT, USER_EMBED_DIM].T -> [batch, MEMORY_SLOT]
attention = torch.softmax(torch.matmul(b_embedding, torch.transpose(self.KEY_MATs[i], dim0=1, dim1=0)), dim=-1)
# [batch, 1, MEMORY_SLOT] bmm [batch, MEMORY_SLOT, USER_EMBED_DIM] -> [batch, 1, USER_EMBED_DIM]
# -> [batch, USER_EMBED_DIM]
embedding = torch.squeeze(torch.bmm(torch.unsqueeze(attention, 1), memory), 1)
mask_label = torch.from_numpy(np.array([label == i for label in col_label])).float().to(DEVICE)
if i==0:
preference = embedding * torch.unsqueeze(mask_label, -1)
else:
preference += embedding * torch.unsqueeze(mask_label, -1)
return preference
# memory: [batch, MEMORY_SLOT, USER_EMBED_DIM]
# b_embedding： [batch, USER_EMBED_DIM]
# col_label: []
# col_mask: tensor([])
# return: [batch, MEMORY_SLOT, USER_EMBED_DIM]
def update(self, memory, b_embedding, col_label, col_mask, isaexpect):
for i in range(self.behavior_typenum):
# [batch, USER_EMBED_DIM] mm [MEMORY_SLOT, USER_EMBED_DIM].T -> [batch, MEMORY_SLOT]
attention = torch.softmax(torch.matmul(b_embedding, torch.transpose(self.KEY_MATs[i], dim0=1, dim1=0)), dim=-1)
# [batch, MEMORY_SLOT, 1] * [batch, 1, USER_EMBED_DIM] -> [batch, MEMORY_SLOT, USER_EMBED_DIM
update = torch.unsqueeze(attention, -1) * torch.unsqueeze(self.update_mlp(b_embedding), 1)
# [batch, MEMORY_SLOT, USER_EMBED_DIM] * [batch, MEMORY_SLOT, 1]
new_memory = memory * (1 - torch.unsqueeze(attention.to(DEVICE), -1)) + update
# [batch]
mask_label = torch.from_numpy(np.array([label==i for label in col_label])).float().to(DEVICE)
mask = mask_label * col_mask
memory = memory * (1-mask.view([len(col_label),1,1])) + new_memory * mask.view([len(col_label),1,1])
return memory
# b_profiless: [batch, max_seq_len, sent, word] [1, 3, 20, 50], gpu tensor
# b_seq_lens: [], list
# b_seq_labels: [[], ...], 0,1,2,3
# return: [batch, MEMORY_SLOT, USER_EMBED_DIM]
def process_seq(self, b_seq_profiless, b_seq_lens, b_seq_tlabels, isaexpect=True):
# memory
batch_memory = torch.from_numpy(np.zeros((len(b_seq_lens), MEMORY_SLOT, USER_EMBED_DIM))).float().to(DEVICE)
for i in range(max(b_seq_lens)):
# [1,0,... ]
col_mask = torch.from_numpy((np.array(b_seq_lens)-i>0)+0.).float().to(DEVICE)
col_label = [bstls[i] if len(bstls)>i+1 else 0 for bstls in b_seq_tlabels]
# [batch, USER_EMBED_DIM]
batch_b_embedding = self.get_profile_embeddings(b_seq_profiless[:, i, :, :].contiguous(), not isaexpect)
batch_memory = self.update(batch_memory, batch_b_embedding, col_label, col_mask, isaexpect)
return batch_memory
# expect_memory: [batch, MEMORY_SLOT, USER_EMBED_DIM]
# job_profiles: [batch, sent, word]
# job_memory: [batch, MEMORY_SLOT, USER_EMBED_DIM]
# expect_profiles: [batch, sent, word]
# label: [], list
def predict(self, expect_memory, job_profiles, job_memory, expect_profiles, tlabel, ise2j):
# read expect
job_embedding = self.get_profile_embeddings(job_profiles, False)
expect_preference = self.read(expect_memory, job_embedding, tlabel)
# read job
expect_embedding = self.get_profile_embeddings(expect_profiles, True)
job_preference = self.read(job_memory, expect_embedding, tlabel)
# interact
e2j_interact = expect_preference * job_embedding
j2e_interact = job_preference * expect_embedding
feature = torch.cat([e2j_interact, job_embedding, j2e_interact, expect_embedding], dim=-1)
if ise2j:
e2j_scores = None
for i in range(self.behavior_typenum):
mask_label = torch.from_numpy(np.array([label == i for label in tlabel])).float().to(DEVICE)
score = torch.squeeze(self.e2j_MTL_MLPs[i](feature), -1)
if i == 0:
e2j_scores = score * mask_label
else:
e2j_scores += score * mask_label
return e2j_scores
else:
j2e_scores = None
for i in range(self.behavior_typenum):
mask_label = torch.from_numpy(np.array([label == i for label in tlabel])).float().to(DEVICE)
score = torch.squeeze(self.j2e_MTL_MLPs[i](feature), -1)
if i==0:
j2e_scores = score * mask_label
else:
j2e_scores += score * mask_label
return j2e_scores