-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain_bpr_fairmi_lastfm.py
210 lines (170 loc) · 8.11 KB
/
train_bpr_fairmi_lastfm.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
# -*- coding: utf-8 -*-
"""
@author: LMC_ZC
"""
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import TensorDataset, DataLoader
import numpy as np
import pandas as pd
import pickle
from collections import defaultdict
from sklearn.metrics import roc_auc_score
from utils import *
from models import *
from tqdm import tqdm
import scipy.stats as stats
import pdb
import sys
def train_semigcn(gcn, sens, n_users, lr=0.001, num_epochs=1000, device='cpu'):
sens = torch.tensor(sens).to(torch.long).to(device)
optimizer = optim.Adam(gcn.parameters(), lr=lr)
final_loss = 0.0
for _ in tqdm(range(num_epochs)):
_, _, su, _ = gcn()
shuffle_idx = torch.randperm(n_users)
classify_loss = F.cross_entropy(su[shuffle_idx].squeeze(), sens[shuffle_idx].squeeze())
optimizer.zero_grad()
classify_loss.backward()
optimizer.step()
final_loss = classify_loss.item()
print('epoch: %d, classify_loss: %.6f' % (num_epochs, final_loss))
def train_unify_mi(sens_enc, inter_enc, club, dataset, u_sens,
n_users, n_items, train_u2i, test_u2i, args):
optimizer_G = optim.Adam(inter_enc.parameters(), lr=args.lr)
optimizer_D = optim.Adam(club.parameters(), lr=args.lr)
train_loader = DataLoader(dataset, shuffle=True, batch_size=args.batch_size, num_workers=args.num_workers)
e_su, e_si, _, _ = sens_enc.forward()
e_su = e_su.detach().to(args.device)
e_si = e_si.detach().to(args.device)
p_su = conditional_samples(e_su.detach().cpu().numpy())
p_si = conditional_samples(e_si.detach().cpu().numpy())
p_su = torch.tensor(p_su).to(args.device)
p_si = torch.tensor(p_si).to(args.device)
ex_enc = torch.load(args.pretrain_path)
e_xu, e_xi = ex_enc.forward()
e_xu = e_xu.detach().to(args.device)
e_xi = e_xi.detach().to(args.device)
best_perf = 0.0
for epoch in range(args.num_epochs):
train_res = {
'bpr_loss': 0.0,
'emb_loss': 0.0,
'lower_bound': 0.0,
'upper_bound': 0.0,
'mi_loss': 0.0,
}
for uij in train_loader:
u = uij[0].type(torch.long).to(args.device)
i = uij[1].type(torch.long).to(args.device)
j = uij[2].type(torch.long).to(args.device)
main_user_emb, main_item_emb = inter_enc.forward()
bpr_loss, emb_loss = calc_bpr_loss(main_user_emb, main_item_emb, u, i, j)
emb_loss = emb_loss * args.l2_reg
e_zu, e_zi = inter_enc.forward()
lower_bound1 = condition_info_nce_for_embeddings(e_xu[torch.unique(u)], e_zu[torch.unique(u)],
e_su[torch.unique(u)], p_su[torch.unique(u)])
lower_bound2 = condition_info_nce_for_embeddings(e_xi[torch.unique(i)], e_zi[torch.unique(i)],
e_si[torch.unique(i)], p_si[torch.unique(i)],)
lower_bound = lower_bound1 + lower_bound2
lower_bound = args.lower_reg * lower_bound
# our further research found that imposing upper bound constraints on
# the user-side only gives more stable and better results, so codes has been updated here.
upper_bound = club.forward(e_zu[torch.unique(u)], e_su[torch.unique(u)])
upper_bound = args.upper_reg * upper_bound
loss = bpr_loss + emb_loss + lower_bound + upper_bound
optimizer_G.zero_grad()
loss.backward()
optimizer_G.step()
train_res['bpr_loss'] += bpr_loss.item()
train_res['emb_loss'] += emb_loss.item()
train_res['lower_bound'] += lower_bound.item()
train_res['upper_bound'] += upper_bound.item()
train_res['bpr_loss'] = train_res['bpr_loss'] / len(train_loader)
train_res['emb_loss'] = train_res['emb_loss'] / len(train_loader)
train_res['lower_bound'] = train_res['lower_bound'] / len(train_loader)
train_res['upper_bound'] = train_res['upper_bound'] / len(train_loader)
##### training club
e_zu, e_zi = inter_enc.forward()
x_samples = e_zu.detach()
y_samples = e_su.detach()
for _ in range(args.train_step):
mi_loss = club.learning_loss(x_samples, y_samples)
optimizer_D.zero_grad()
mi_loss.backward()
optimizer_D.step()
train_res['mi_loss'] += mi_loss.item()
train_res['mi_loss'] = train_res['mi_loss'] / args.train_step
# print training logs
training_logs = 'epoch: %d, ' % epoch
for name, value in train_res.items():
training_logs += name + ':' + '%.6f' % value + ' '
print(training_logs)
with torch.no_grad():
t_user_emb, t_item_emb = inter_enc.forward()
test_res = ranking_evaluate(
user_emb=t_user_emb.detach().cpu().numpy(),
item_emb=t_item_emb.detach().cpu().numpy(),
n_users=n_users,
n_items=n_items,
train_u2i=train_u2i,
test_u2i=test_u2i,
sens=u_sens,
num_workers=args.num_workers)
p_eval = ''
for keys, values in test_res.items():
p_eval += keys + ':' + '[%.6f]' % values + ' '
print(p_eval)
if best_perf < test_res['ndcg@10']:
best_perf = test_res['ndcg@10']
torch.save(inter_enc, args.param_path)
print('save successful')
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='lastfm_bpr_fairmi',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--bakcbone', type=str, default='bpr')
parser.add_argument('--dataset', type=str, default='./data/lastfm-360k/process/process.pkl')
parser.add_argument('--emb_size', type=int, default=64)
parser.add_argument('--hidden_size', type=int, default=256)
parser.add_argument('--lr', type=float, default=0.001)
parser.add_argument('--l2_reg', type=float, default=0.001)
parser.add_argument('--n_layers', type=int, default=3)
parser.add_argument('--batch_size', type=int, default=4096)
parser.add_argument('--num_workers', type=int, default=6)
parser.add_argument('--log_path', type=str, default='./logs/bpr_fairmi_lastfm.txt')
parser.add_argument('--param_path', type=str, default='./param/bpr_fairmi_lastfm.pth')
parser.add_argument('--pretrain_path', type=str, default='./param/bpr_base_lastfm.pth')
parser.add_argument('--lower_reg', type=float, default=0.1)
parser.add_argument('--upper_reg', type=float, default=0.1)
parser.add_argument('--train_step', type=int, default=50)
parser.add_argument('--num_epochs', type=int, default=500)
parser.add_argument('--device', type=str, default='cuda:0')
args = parser.parse_args()
sys.stdout = Logger(args.log_path)
print(args)
with open(args.dataset, 'rb') as f:
train_u2i = pickle.load(f)
train_i2u = pickle.load(f)
test_u2i = pickle.load(f)
test_i2u = pickle.load(f)
train_set = pickle.load(f)
test_set = pickle.load(f)
user_side_features = pickle.load(f)
n_users, n_items = pickle.load(f)
bprmf = BPRMF(n_users, n_items, args.emb_size, device=args.device)
u_sens = user_side_features['gender'].astype(np.int32)
dataset = BPRTrainLoader(train_set, train_u2i, n_items)
graph = Graph(n_users, n_items, train_u2i)
norm_adj = graph.generate_ori_norm_adj()
sens_enc = SemiGCN(n_users, n_items, norm_adj,
args.emb_size, args.n_layers, args.device,
nb_classes=np.unique(u_sens).shape[0])
inter_enc = BPRMF(n_users, n_items, args.emb_size, args.device)
club = CLUBSample(args.emb_size, args.emb_size, args.hidden_size, args.device)
train_semigcn(sens_enc, u_sens, n_users, device=args.device)
train_unify_mi(sens_enc, inter_enc, club, dataset, u_sens, n_users, n_items, train_u2i, test_u2i, args)
sys.stdout = None