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
| import torch
from torch import nn
import math
from tensorboardX import SummaryWriter
class NeuMF(nn.Module):
def __init__(self, num_users, num_items, latent_dim_mf, latent_dim_mlp, config_layers, learning_rate):
super(NeuMF, self).__init__()
self.num_users = num_users
self.num_items = num_items
self.latent_dim_mf = latent_dim_mf
self.latent_dim_mlp = latent_dim_mlp
self.config_layers = config_layers
self.learning_rate = learning_rate
self._writer = SummaryWriter(log_dir='logs')
self.embedding_user_mlp = nn.Embedding(num_embeddings=self.num_users, embedding_dim=self.latent_dim_mlp)
self.embedding_item_mlp = nn.Embedding(num_embeddings=self.num_items, embedding_dim=self.latent_dim_mlp)
self.embedding_user_mf = nn.Embedding(num_embeddings=self.num_users, embedding_dim=self.latent_dim_mf)
self.embedding_item_mf = nn.Embedding(num_embeddings=self.num_items, embedding_dim=self.latent_dim_mf)
self.fc_layers = nn.ModuleList()
input_size = self.latent_dim_mlp * 2
for output_size in self.config_layers:
self.fc_layers.append(nn.Linear(input_size, output_size))
input_size = output_size
self.output_layer = nn.Linear(self.config_layers[-1] + self.latent_dim_mf, 1)
self.loss = nn.BCELoss()
self.optimizer = torch.optim.Adam(self.parameters(), lr=self.learning_rate)
self._metrics = Metrics(top_k=10)
self._init_weight_()
def _init_weight_(self):
nn.init.normal_(self.embedding_user_mlp.weight, std=0.01)
nn.init.normal_(self.embedding_item_mlp.weight, std=0.01)
nn.init.normal_(self.embedding_user_mf.weight, std=0.01)
nn.init.normal_(self.embedding_item_mf.weight, std=0.01)
for fc in self.fc_layers:
nn.init.xavier_uniform_(fc.weight)
nn.init.zeros_(fc.bias)
nn.init.kaiming_uniform_(self.output_layer.weight, nonlinearity='sigmoid')
nn.init.zeros_(self.output_layer.bias)
def forward(self, user, item):
user_mlp = self.embedding_user_mlp(user)
item_mlp = self.embedding_item_mlp(item)
mlp_vector = torch.cat([user_mlp, item_mlp], dim=-1)
for fc in self.fc_layers:
mlp_vector = torch.relu(fc(mlp_vector))
user_mf = self.embedding_user_mf(user)
item_mf = self.embedding_item_mf(item)
mf_vector = torch.mul(user_mf, item_mf)
output_vector = torch.cat([mlp_vector, mf_vector], dim=-1)
rating_logit = torch.sigmoid(self.output_layer(output_vector))
return rating_logit
def train_single_batch(self, users, items, ratings):
if torch.cuda.is_available():
users, items, ratings = users.cuda(), items.cuda(), ratings.cuda()
self.train()
self.optimizer.zero_grad()
ratings_pred = self.forward(users, items)
loss = self.loss(ratings_pred.view(-1), ratings)
loss.backward()
self.optimizer.step()
return loss.item()
def train_single_epoch(self, train_loader, epoch_id):
self.train()
total_loss = 0
for batch_id, (users, items, ratings) in enumerate(train_loader):
loss = self.train_single_batch(users, items, ratings)
print('Epoch {} Batch {} Loss {}'.format(epoch_id, batch_id, loss))
total_loss += loss
self._writer.add_scalar('model/loss', total_loss, epoch_id)
return total_loss
def evaluate(self, evaluate_data):
self.eval()
test_users, test_items, negative_users, negative_items = evaluate_data
if torch.cuda.is_available():
test_users, test_items, negative_users, negative_items = test_users.cuda(), test_items.cuda(), negative_users.cuda(), negative_items.cuda()
test_scores = self.forward(test_users, test_items)
negative_scores = self.forward(negative_users, negative_items)
self._metrics.set_subjects(
test_users=test_users.cpu(),
test_items=test_items.cpu(),
test_scores=test_scores.cpu(),
negative_users=negative_users.cpu(),
negative_items=negative_items.cpu(),
negative_scores=negative_scores.cpu()
)
hit_ratio = self._metrics.cal_hit_ratio()
ndcg = self._metrics.cal_ndcg()
self._writer.add_scalar('model/hit_ratio', hit_ratio, 0)
self._writer.add_scalar('model/ndcg', ndcg, 0)
print('Hit Ratio is {:.6f}, NDCG is {:.6f}'.format(hit_ratio, ndcg))
return hit_ratio, ndcg
class Metrics(object):
def __init__(self, top_k):
self._top_k = top_k
def set_subjects(self, test_users, test_items, test_scores, negative_users, negative_items, negative_scores):
self._subjects = []
for i in range(len(test_users)):
user = test_users[i].item()
test_item = test_items[i].item()
test_score = test_scores[i].view(-1).item()
neg_items = negative_items[i*99:(i+1)*99].detach().numpy()
neg_scores = negative_scores[i*99:(i+1)*99].view(-1).detach().numpy()
items = np.concatenate(([test_item], neg_items))
scores = np.concatenate(([test_score], neg_scores))
ranked_idx = np.argsort(-scores)
ranked_items = items[ranked_idx]
self._subjects.append({
'user': user,
'ranked_items': ranked_items,
'test_item': test_item
})
def cal_hit_ratio(self):
hits = 0
for subject in self._subjects:
if subject['test_item'] in subject['ranked_items'][:self._top_k]:
hits += 1
return hits / len(self._subjects)
def cal_ndcg(self):
total_ndcg = 0
for subject in self._subjects:
try:
rank = np.where(subject['ranked_items'] == subject['test_item'])[0][0] + 1
total_ndcg += math.log(2) / math.log(1 + rank)
except IndexError:
continue
return total_ndcg / len(self._subjects)
|
