-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
178 lines (147 loc) · 6.63 KB
/
train.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
import os
import pickle
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
class TripletUniformPair(Dataset):
def __init__(self, w, num_item, pair, args):
self.neg_w = 1 - w
self.num_item = num_item
self.pair = pair
self.batch_size = args.batch_size
def __getitem__(self, idx):
idx = np.random.choice(self.pair.shape[0], size=args.batch_size)
u = self.pair[idx, 0]
i = self.pair[idx, 1]
j = torch.multinomial(self.neg_w[u,:], num_samples=1).squeeze()
return u, i, j
def __len__(self):
return 10*len(self.pair)
class BPR(nn.Module):
def __init__(self, user_size, item_size, args):
super().__init__()
self.P = nn.Parameter(torch.rand(user_size, args.dim))
self.Q = nn.Parameter(torch.rand(item_size, args.dim))
self.delta_P = nn.Parameter(torch.zeros([user_size, args.dim], dtype=torch.float32))
self.delta_Q = nn.Parameter(torch.zeros([item_size, args.dim], dtype=torch.float32))
self.delta_P.requires_grad = False
self.delta_Q.requires_grad = False
def forward(self, u, i, j,mode):
if mode==1:
return self.apr_loss(u,i,j)
else:
return self.bpr_loss(u,i,j)
def bpr_loss(self, u, i, j): #H[i,:] = items[0], H[j,:] = items[1]
x_ui = torch.mul(self.P[u,:], self.Q[i,:]).sum(dim=1)
x_uj = torch.mul(self.P[u,:], self.Q[j,:]).sum(dim=1)
x_uij = x_ui - x_uj
log_prob = torch.log(torch.sigmoid(x_uij)).mean()
return -log_prob
def apr_loss(self, u, i, j):
x_ui_adv = torch.mul(self.P[u,:]+self.delta_P[u,:], self.Q[i,:]+self.delta_Q[i,:]).sum(dim=1)
x_uj_adv = torch.mul(self.P[u,:]+self.delta_P[u,:], self.Q[j,:]+self.delta_Q[j,:]).sum(dim=1)
x_uij_adv = x_ui_adv - x_uj_adv
log_prob = torch.log(torch.sigmoid(x_uij_adv)).mean()
return -log_prob
def precision_and_recall_k(user_emb, item_emb, train_w, test_w, klist):
"""Compute precision at k using GPU.
Args:
user_emb (torch.Tensor): embedding for user (shape [user_num, dim])
item_emb (torch.Tensor): embedding for item (shape [item_num, dim])
train_w (torch.Tensor): mask array for train record (shape [user_num, item_num])
test_w (torch.Tensor): mask array for test record (shape [user_num, item_num])
k (list(int)): list of k
Returns:
(torch.Tensor, torch.Tensor) Precision and recall at k
"""
# Compute all pair of training and test record
# Reason why do sigmoid is sigmoid and compress value into [0, 1]
# And we are going to make useless value to zero to make smallest value
result = torch.mm(user_emb, item_emb.t())
result = torch.sigmoid(result)
# Mask pred and true
# test_pred represents both test and negative record
test_pred_mask = 1 - (train_w)
test_pred = test_pred_mask * result
# test_true represents only test record indicator
test_true_mask = test_w
test_true = test_true_mask * result
# Sort indice and get test_pred_topk
_, test_indices = torch.topk(test_pred, dim=1, k=max(klist))
precisions, recalls = [], []
for k in klist:
topk_mask = torch.zeros_like(test_pred)
topk_mask.scatter_(dim=1, index=test_indices[:, :k],
src=torch.tensor(1.0))
test_pred_topk = topk_mask * test_pred
# Compare which is not zero and equal with test_true, which means that
# both is not excluded by mask and is true value
acc_result = (test_pred_topk != 0) & (test_pred_topk == test_true)
precisions.append(acc_result.sum().float() / (user_emb.shape[0] * k))
recalls.append((acc_result.float().sum(dim=1) / test_w.sum(dim=1)).mean())
return precisions, recalls
if __name__=='__main__':
# Parse argument
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str,
default=os.path.join('preprocessed', 'bpr-movielens-1m.pickle'),
help="File path for data")
# Model
parser.add_argument('--dim', type=int, default=4,
help="Dimension for embedding")
# Optimizer
parser.add_argument('--lr', type=float, default=1e-3,
help="Learning rate")
parser.add_argument('--weight_decay', type=float, default=0.000025,
help="Weight decay factor")
# Training
parser.add_argument('--n_epochs', type=int, default=10,
help="Number of epoch during training")
parser.add_argument('--batch_size', type=int, default=128,
help="Batch size in one iteration")
parser.add_argument('--print_every', type=int, default=10,
help="Period for printing smoothing loss during training")
parser.add_argument('--eval_every', type=int, default=1000,
help="Period for evaluating precision and recall during training")
parser.add_argument('--save_every', type=int, default=1000,
help="Period for saving model during training")
parser.add_argument('--model', type=str,
default=os.path.join('output', 'bpr.pt'),
help="File path for model")
args = parser.parse_args()
# Load preprocess data
with open(args.data, 'rb') as f:
dataset = pickle.load(f)
user_size, item_size = dataset['user_size'], dataset['item_size']
train_w, test_w = dataset['train_w'], dataset['test_w']
train_pair = dataset['train_pair']
print('Load complete')
# Convert to tensor and move to GPU
train_w = torch.tensor(train_w, dtype=torch.float)
test_w = torch.tensor(test_w, dtype=torch.float)
train_pair = torch.tensor(train_pair, dtype=torch.long)
# Create dataset, model, optimizer
dataset = TripletUniformPair(train_w, item_size, train_pair, args)
model = BPR(user_size, item_size, args)
optimizer = optim.Adam(model.parameters(), lr=args.lr,
weight_decay=args.weight_decay)
# Training
total_loss = 0
for epoch in range(args.n_epochs):
for idx, (u, i, j) in enumerate(iter(dataset)):
optimizer.zero_grad()
loss = model(u, i, j, 0)
loss.backward()
model.delta_P = nn.Parameter(0.1 * F.normalize(model.P, p=2, dim=1))
model.delta_Q = nn.Parameter(0.1 * F.normalize(model.Q, p=2, dim=1))
loss = model(u, i, j, 1)
loss.backward()
optimizer.step()
total_loss += loss
if idx % 100 == 0:
print('loss : %.4f' %loss)
print('Total Loss : %.4f' %(total_loss/len(dataset)))