-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathnas4lfm.py
executable file
·158 lines (136 loc) · 6.31 KB
/
nas4lfm.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
import torch
import torch.nn
import numpy as np
from copy import deepcopy
import random
from torch.autograd import Variable
from torch.nn import functional as F
def init_param(param_list):
for param in param_list:
torch.nn.init.normal(param,mean=0,std=0.01)
class AdaptiveNet(torch.nn.Module):
def __init__(self, config,p_dim,relu=0):
super(AdaptiveNet,self).__init__()
self.ifeature_dim = config['ifeature_dim']
self.ufeature_dim = config['ufeature_dim']
self.embedding_dim = config['embedding_dim']
self.rating_range=config['rating_range_lfm']
self.context_dim=8*self.embedding_dim
self.p_dim=p_dim
self.relu=relu
self.ctx_fc=torch.nn.Linear(self.embedding_dim*2+1,self.context_dim)
self.out_layer=torch.nn.Linear(self.context_dim,self.p_dim)
def forward(self,emb,ys):
x=torch.cat((emb,ys.view(-1,1)),1)
x=self.ctx_fc(x)
x=F.leaky_relu(x)
x=torch.mean(x,0)
x=self.out_layer(x)
if self.relu:x=F.relu(x)
return x.view(4,-1)
class SuperNet(torch.nn.Module):
def __init__(self, config):
super(SuperNet,self).__init__()
self.ifeature_dim = config['ifeature_dim']
self.ufeature_dim = config['ufeature_dim']
self.embedding_dim = config['embedding_dim']
self.rating_range=config['rating_range_lfm']
self.use_cuda=1
self.hidden_units=torch.tensor(config['hidden_units'])
ap_dim=int(self.hidden_units.sum())
self.hidden_units=config['hidden_units']
self.item_emb = torch.nn.Linear(in_features=self.ifeature_dim,out_features=self.embedding_dim)
self.user_emb = torch.nn.Linear(in_features=self.ufeature_dim,out_features=self.embedding_dim)
self.adp=AdaptiveNet(config,(self.embedding_dim*2+ap_dim)*4)
self.fc1=torch.nn.Linear(self.embedding_dim*2,self.hidden_units[0])
self.fc2=torch.nn.Linear(self.hidden_units[0],self.hidden_units[1])
self.fc3=torch.nn.Linear(self.hidden_units[1],self.hidden_units[2])
self.linear_out = torch.nn.Linear(self.hidden_units[2], 1)
self.alpha=torch.nn.Parameter(torch.zeros(4,4).cuda())
self.alpha.requires_grad=True
init_param(list(self.parameters()))
self.optim=torch.optim.Adam(self.parameters(),config['lr_i'])
def modulate(self,x,maxp,minp,mutp,addp,alpha):
x1=F.normalize(x,1.0,-1)
x=alpha[0]*F.normalize(torch.maximum(x,maxp),1.0,-1)+(1-alpha[0])*x1
x=alpha[1]*F.normalize(torch.minimum(x,minp),1.0,-1)+(1-alpha[1])*x1
x=alpha[2]*F.normalize(x*mutp,1.0,-1)+(1-alpha[2])*x1
x=alpha[3]*F.normalize((x+addp),1.0,-1)+(1-alpha[3])*x1
#x+=x0
return x
def forward(self, xs,ys,xq, training = True):
item_x = Variable(xs[:, 0:3846], requires_grad=False).float()
user_x = Variable(xs[:, 3846:], requires_grad=False).float()
item_emb = self.item_emb(item_x)
user_emb = self.user_emb(user_x)
emb = torch.cat((item_emb, user_emb), 1)
p=self.adp(emb,ys)
maxp=p[0]
minp=p[1]
mutp=F.relu(p[2])
addp=p[3]
alpha=torch.sigmoid(self.alpha)
item_x = Variable(xq[:, 0:3846], requires_grad=False).float()
user_x = Variable(xq[:, 3846:], requires_grad=False).float()
item_emb = self.item_emb(item_x)
user_emb = self.user_emb(user_x)
emb = torch.cat((item_emb, user_emb), 1)
d=0
x=emb
x=self.modulate(x,maxp[:self.embedding_dim*2],minp[:self.embedding_dim*2],mutp[:self.embedding_dim*2],addp[:self.embedding_dim*2],alpha[0])
x=x.normalize
x=F.leaky_relu(x)
x=self.fc1(x)
d+=self.embedding_dim*2
x=self.modulate(x,maxp[d:d+self.hidden_units[0]],minp[d:d+self.hidden_units[0]],mutp[d:d+self.hidden_units[0]],addp[d:d+self.hidden_units[0]],alpha[1])
x=F.leaky_relu(x)
x=self.fc2(x)
d+=self.hidden_units[0]
x=self.modulate(x,maxp[d:d+self.hidden_units[1]],minp[d:d+self.hidden_units[1]],mutp[d:d+self.hidden_units[1]],addp[d:d+self.hidden_units[1]],alpha[2])
x=F.leaky_relu(x)
x=self.fc3(x)
d+=self.hidden_units[1]
x=self.modulate(x,maxp[d:d+self.hidden_units[2]],minp[d:d+self.hidden_units[2]],mutp[d:d+self.hidden_units[2]],addp[d:d+self.hidden_units[2]],alpha[2])
x=F.leaky_relu(x)
x=self.linear_out(x)
return self.rating_range*torch.sigmoid(x)
def global_update(self, xs,ys,xq,yq):
batch_sz = len(xs)
loss=0
self.optim.zero_grad()
if self.use_cuda:
for i in range(batch_sz):
xs[i] = xs[i].cuda()
ys[i] = ys[i].cuda()
xq[i] = xq[i].cuda()
yq[i] = yq[i].cuda()
for i in range(batch_sz):
y_pred=self.forward(xs[i],ys[i],xq[i],0).reshape(-1,1)
#y_pred = torch.clip(y_pred,1e-6,1-1e-6)
loss+=F.mse_loss(y_pred,yq[i].view(-1,1))
loss=loss/batch_sz
self.optim.zero_grad()
loss.backward()
self.optim.step()
def query_rec(self, support_set_xs, support_set_ys, query_set_xs, query_set_ys):
batch_sz = 1
# used for calculating the rmse.
losses_q = []
losses_mae=[]
if self.use_cuda:
for i in range(batch_sz):
support_set_xs[i] = support_set_xs[i].cuda()
support_set_ys[i] = support_set_ys[i].cuda()
query_set_xs[i] = query_set_xs[i].cuda()
query_set_ys[i] = query_set_ys[i].cuda()
for i in range(batch_sz):
#query_set_y_pred = self.forward(support_set_xs[i], support_set_ys[i], query_set_xs[i], num_local_update)
query_set_y_pred = self.forward(support_set_xs[i], support_set_ys[i], query_set_xs[i], 0)
loss_q = F.mse_loss(query_set_y_pred, query_set_ys[i].view(-1, 1))
loss_mae=F.l1_loss(query_set_y_pred, query_set_ys[i].view(-1, 1))
losses_q.append(loss_q)
losses_mae.append(loss_mae)
losses_q = torch.stack(losses_q).mean(0)
losses_mae = torch.stack(losses_mae).mean(0)
output_list, recommendation_list = query_set_y_pred.view(-1).sort(descending=True)
return losses_q.item(),losses_mae.item(), recommendation_list