-
Notifications
You must be signed in to change notification settings - Fork 2
/
model.py
252 lines (173 loc) · 7.09 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
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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
import torch.nn.functional as F
import torch.nn as nn
import torch
import numpy as np
from utils import sim, to_np
import math
from pdb import set_trace as bp
class f(nn.Module):
def __init__(self, dims):
super(f, self).__init__()
self.net = nn.Sequential(nn.Linear(dims[0], dims[1]), nn.ReLU(), nn.Linear(dims[1], dims[2]))
def forward(self, x):
return self.net(x)
class Base_model(nn.Module):
def __init__(self, user_num, item_num, dim, gpu):
super(Base_model, self).__init__()
self.user_num = user_num
self.item_num = item_num
self.user_list = torch.LongTensor([i for i in range(user_num)]).to(gpu)
self.item_list = torch.LongTensor([i for i in range(item_num)]).to(gpu)
self.user_emb = nn.Embedding(self.user_num, dim)
self.item_emb = nn.Embedding(self.item_num, dim)
nn.init.normal_(self.user_emb.weight, mean=0., std= 0.01)
nn.init.normal_(self.item_emb.weight, mean=0., std= 0.01)
def forward(self, batch_user, batch_pos_item, batch_neg_item):
u = self.user_emb(batch_user)
i = self.item_emb(batch_pos_item)
j = self.item_emb(batch_neg_item)
pos_score = (u * i).sum(dim=1, keepdim=True)
neg_score = (u * j).sum(dim=1, keepdim=True)
output = (pos_score, neg_score)
return output
def get_loss(self, output):
pos_score, neg_score = output[0], output[1]
loss = -(pos_score - neg_score).sigmoid().log().sum()
return loss
def get_embedding(self):
users = self.user_emb(self.user_list)
items = self.item_emb(self.item_list)
return users, items
class FTD(Base_model):
def __init__(self, user_num, item_num, user_emb_teacher, item_emb_teacher, gpu, student_dim):
Base_model.__init__(self, user_num, item_num, student_dim, gpu)
self.student_dim = student_dim
self.gpu = gpu
# Teacher
self.user_emb_teacher = nn.Embedding.from_pretrained(user_emb_teacher)
self.item_emb_teacher = nn.Embedding.from_pretrained(item_emb_teacher)
self.user_emb_teacher.weight.requires_grad = False
self.item_emb_teacher.weight.requires_grad = False
self.teacher_dim = self.user_emb_teacher.weight.size(1)
# topology distillation loss
def get_TD_loss(self, batch_user, batch_item):
s = torch.cat([self.user_emb(batch_user), self.item_emb(batch_item)], 0)
t = torch.cat([self.user_emb_teacher(batch_user), self.item_emb_teacher(batch_item)], 0)
# Full topology
t_dist = sim(t, t).view(-1)
s_dist = sim(s, s).view(-1)
total_loss = ((t_dist - s_dist) ** 2).sum()
return total_loss
class HTD(Base_model):
def __init__(self, user_num, item_num, user_emb_teacher, item_emb_teacher, gpu, student_dim, K, choice):
Base_model.__init__(self, user_num, item_num, student_dim, gpu)
self.student_dim = student_dim
self.gpu = gpu
# Teacher
self.user_emb_teacher = nn.Embedding.from_pretrained(user_emb_teacher)
self.item_emb_teacher = nn.Embedding.from_pretrained(item_emb_teacher)
self.user_emb_teacher.weight.requires_grad = False
self.item_emb_teacher.weight.requires_grad = False
self.teacher_dim = self.user_emb_teacher.weight.size(1)
# Group Assignment related parameters
self.K = K
F_dims = [self.student_dim, (self.teacher_dim + self.student_dim) // 2, self.teacher_dim]
self.user_f = nn.ModuleList([f(F_dims) for i in range(self.K)])
self.item_f = nn.ModuleList([f(F_dims) for i in range(self.K)])
self.user_v = nn.Sequential(nn.Linear(self.teacher_dim, K), nn.Softmax(dim=1))
self.item_v = nn.Sequential(nn.Linear(self.teacher_dim, K), nn.Softmax(dim=1))
self.sm = nn.Softmax(dim = 1)
self.T = 0.1
# Group-Level topology design choices
self.choice = choice
def get_group_result(self, batch_entity, is_user=True):
with torch.no_grad():
if is_user:
t = self.user_emb_teacher(batch_entity)
v = self.user_v
else:
t = self.item_emb_teacher(batch_entity)
v = self.item_v
z = v(t).max(-1)[1]
if not is_user:
z = z + self.K
return z
# For Adaptive Group Assignment
def get_GA_loss(self, batch_entity, is_user=True):
if is_user:
s = self.user_emb(batch_entity)
t = self.user_emb_teacher(batch_entity)
f = self.user_f
v = self.user_v
else:
s = self.item_emb(batch_entity)
t = self.item_emb_teacher(batch_entity)
f = self.item_f
v = self.item_v
alpha = v(t)
g = torch.distributions.Gumbel(0, 1).sample(alpha.size()).to(self.gpu)
alpha = alpha + 1e-10
z = self.sm((alpha.log() + g) / self.T)
z = torch.unsqueeze(z, 1)
z = z.repeat(1, self.teacher_dim, 1)
f_hat = [f[i](s).unsqueeze(-1) for i in range(self.K)]
f_hat = torch.cat(f_hat, -1)
f_hat = f_hat * z
f_hat = f_hat.sum(2)
GA_loss = ((t-f_hat) ** 2).sum(-1).sum()
return GA_loss
def get_TD_loss(self, batch_user, batch_item):
if self.choice == 'first':
return self.get_TD_loss1(batch_user, batch_item)
else:
return self.get_TD_loss2(batch_user, batch_item)
# Topology Distillation Loss (with Group(P,P))
def get_TD_loss1(self, batch_user, batch_item):
s = torch.cat([self.user_emb(batch_user), self.item_emb(batch_item)], 0)
t = torch.cat([self.user_emb_teacher(batch_user), self.item_emb_teacher(batch_item)], 0)
z = torch.cat([self.get_group_result(batch_user, is_user=True), self.get_group_result(batch_item, is_user=False)], 0)
G_set = z.unique()
Z = F.one_hot(z).float()
# Compute Prototype
with torch.no_grad():
tmp = Z.T
tmp = tmp / (tmp.sum(1, keepdims=True) + 1e-10)
P_s = tmp.mm(s)[G_set]
P_t = tmp.mm(t)[G_set]
# entity_level topology
entity_mask = Z.mm(Z.T)
t_sim_tmp = sim(t, t) * entity_mask
t_sim_dist = t_sim_tmp[t_sim_tmp > 0.]
s_sim_dist = sim(s, s) * entity_mask
s_sim_dist = s_sim_dist[t_sim_tmp > 0.]
# # Group_level topology
t_proto_dist = sim(P_t, P_t).view(-1)
s_proto_dist = sim(P_s, P_s).view(-1)
total_loss = ((s_sim_dist - t_sim_dist) ** 2).sum() + ((s_proto_dist - t_proto_dist) ** 2).sum()
return total_loss
# Topology Distillation Loss (with Group(P,e))
def get_TD_loss2(self, batch_user, batch_item):
s = torch.cat([self.user_emb(batch_user), self.item_emb(batch_item)], 0)
t = torch.cat([self.user_emb_teacher(batch_user), self.item_emb_teacher(batch_item)], 0)
z = torch.cat([self.get_group_result(batch_user, is_user=True), self.get_group_result(batch_item, is_user=False)], 0)
G_set = z.unique()
Z = F.one_hot(z).float()
# Compute Prototype
with torch.no_grad():
tmp = Z.T
tmp = tmp / (tmp.sum(1, keepdims=True) + 1e-10)
P_s = tmp.mm(s)[G_set]
P_t = tmp.mm(t)[G_set]
# entity_level topology
entity_mask = Z.mm(Z.T)
t_sim_tmp = sim(t, t) * entity_mask
t_sim_dist = t_sim_tmp[t_sim_tmp > 0.]
s_sim_dist = sim(s, s) * entity_mask
s_sim_dist = s_sim_dist[t_sim_tmp > 0.]
# # Group_level topology
# t_proto_dist = (sim(P_t, t) * (1 - Z.T)[G_set]).view(-1)
# s_proto_dist = (sim(P_s, s) * (1 - Z.T)[G_set]).view(-1)
t_proto_dist = sim(P_t, t).view(-1)
s_proto_dist = sim(P_s, s).view(-1)
total_loss = ((s_sim_dist - t_sim_dist) ** 2).sum() + ((s_proto_dist - t_proto_dist) ** 2).sum()
return total_loss