-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_model.py
233 lines (221 loc) · 12.8 KB
/
train_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
import os
import math
import time
import random
import torch
import argparse
import numpy as np
import torch
import torch.nn as nn
# import utils as ut
import torch.optim as optim
from torch import autograd
from torch.utils import data
import torch.utils.data as Data
from torch.autograd import Variable
import dataloader as dataload
from codebase.method.config import get_config
from codebase.method.learner import Learner
import sklearn.metrics as skm
import warnings
from sklearn.metrics import accuracy_score, log_loss
import torch.nn.functional as F
warnings.filterwarnings('ignore')
device = torch.device("cuda:0" if (torch.cuda.is_available()) else "cpu")
print("torch.cuda.is_available():%s" % (torch.cuda.is_available()))
args, _ = get_config()
workstation_path = './'
print('Experiment: epsilon={}, bias={}, iv_lr={}'.format(args.epsilon, args.bias, args.iv_lr))
print('================================================')
# train_dataset_dir = os.path.join(workstation_path, 'dataset/', 'douban', 'dataset.npy')
# test_dataset_dir = os.path.join(workstation_path, 'dataset/', 'douban', 'dataset.npy')
if args.dataset == 'huawei':
train_dataset_dir = os.path.join(workstation_path, 'dataset/', 'huawei', 'train.csv')
test_dataset_dir = os.path.join(workstation_path, 'dataset/', 'huawei', 'train.csv')
train_dataloader = dataload.dataload_huawei(train_dataset_dir, args.batch_size, mode='train')
test_dataloader = dataload.dataload_huawei(test_dataset_dir, args.batch_size, mode='test')
elif args.dataset[:3] == 'non' or args.dataset == 'celeba': # the synthetic data drop here
print(args)
train_dataset_dir = os.path.join(workstation_path, 'dataset/', args.dataset, 'train', 'data_nonuniform.csv')
test_dataset_dir = os.path.join(workstation_path, 'dataset/', args.dataset, 'dev', 'data_uniform.csv')
train_dataloader = dataload.dataload_huawei(train_dataset_dir, args.batch_size, mode='train')
test_dataloader = dataload.dataload_huawei(test_dataset_dir, args.batch_size, mode='test')
elif args.dataset in ['coat', 'yahoo', 'pcic']:
train_dataset_dir = os.path.join(workstation_path, 'dataset', args.dataset, 'train', 'data_nonuniform.csv')
test_dataset_dir = os.path.join(workstation_path, 'dataset', args.dataset, 'dev', 'data_uniform.csv')
if args.dataset in ['coat']:
data_feature_path = os.path.join(workstation_path, 'dataset', args.dataset)
else:
args.feature_data = False
data_feature_path = None
print(args)
pretrain_dataloader = dataload.dataload_ori(train_dataset_dir, args.batch_size, data_feature_path=data_feature_path,
syn=False)
train_dataloader = dataload.dataload_ori(train_dataset_dir, args.batch_size,
data_feature_path=data_feature_path, syn=False)
test_dataloader = dataload.dataload_ori(test_dataset_dir, args.batch_size, data_feature_path=data_feature_path,
syn=False)
# 构建模型
model = Learner(args)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3, betas=(0.9, 0.999))
iv_optimizer = torch.optim.Adam(model.parameters(), lr=args.iv_lr, betas=(0.9, 0.999))
for epoch in range(args.epoch_max):
model.train()
total_auc = 0
total_acc = 0
total_logloss = 0
total_ds_loss = 0
total_kl = 0
total_iv_loss = 0
iv_loss = 0
# for step in range(args.max_step):
if args.mode == 'CausalRep':
iv_count = 1
for x, y in train_dataloader:
optimizer.zero_grad()
loss, ds_loss, kl, z, _ = model.learn(x, y)
if args.downstream == 'MLP':
auc = skm.roc_auc_score(y.cpu().numpy(),
torch.argmax(model.predict(x), dim=1).cpu().detach().numpy())
acc = accuracy_score(y.cpu().numpy(), torch.argmax(model.predict(x), dim=1).cpu().detach().numpy())
logloss = log_loss(y.cpu().numpy(),
F.sigmoid(torch.max(model.predict(x), dim=1)[0]).cpu().detach().numpy())
else:
y_pred = model.predict(x)
auc = skm.roc_auc_score(y.cpu().numpy(), y_pred.cpu().detach().numpy())
acc = skm.accuracy_score(y.cpu().numpy(), torch.where(y_pred > 0.5, torch.ones_like(y_pred),
torch.zeros_like(y_pred)).cpu().detach().numpy())
L = loss.mean() # + iv_loss.mean()
L.backward()
optimizer.step()
total_auc += auc
total_acc += acc
total_logloss += logloss
total_ds_loss += ds_loss
total_kl += kl
m = len(train_dataloader)
# for x, y in train_dataloader:
if args.mode == 'IB':
continue
iv_optimizer.zero_grad()
iv_loss = model.learn_neg(x, y)
# print()
if iv_loss is not None:
# iv_optimizer.zero_grad()
# iv_loss = model.learn_neg(x, y)
L = iv_loss.mean()
L.backward()
iv_optimizer.step()
total_iv_loss += iv_loss
iv_count += 1
else:
continue
# test
if epoch % 1 == 0:
test_advauc = 0
test_advacc = 0
total_test_auc = 0
total_test_acc = 0
total_test_advauc = 0
total_test_advacc = 0
total_test_logloss = 0
total_test_dc = 0
total_test_mic = 0
total_test_tic = 0
for test_x, test_y in test_dataloader:
test_a = x[:, args.user_dim:]
# print(test_x)
if args.downstream in ['MLP', 'bprBPR', 'mlpBPR', 'gmfBPR', 'NeuBPR']:
# print(test_y)
# test_y_pred = model.predict(test_x, test_a)
test_auc = skm.roc_auc_score(test_y.cpu().numpy(),
torch.argmax(model.predict(test_x), dim=1).cpu().detach().numpy())
test_acc = accuracy_score(test_y.cpu().numpy(),
torch.argmax(model.predict(test_x), dim=1).cpu().detach().numpy())
test_adv_y_pred = model.eval_adv(test_x, test_y)
test_advauc = skm.roc_auc_score(test_y.cpu().numpy(),
torch.argmax(test_adv_y_pred, dim=1).cpu().detach().numpy())
test_advacc = accuracy_score(test_y.cpu().numpy(),
torch.argmax(test_adv_y_pred, dim=1).cpu().detach().numpy())
else:
test_y_pred = model.predict(test_x, test_a)
test_adv_y_pred = model.eval_adv(test_x, test_y, test_a)
test_auc = skm.roc_auc_score(test_y.cpu().numpy(), test_y_pred.cpu().detach().numpy())
test_acc = skm.accuracy_score(test_y.cpu().numpy(),
torch.where(test_y_pred > 0.5, torch.ones_like(test_y_pred),
torch.zeros_like(
test_y_pred)).cpu().detach().numpy())
test_advauc = skm.roc_auc_score(test_y.cpu().numpy(), test_adv_y_pred.cpu().detach().numpy())
test_advacc = skm.accuracy_score(test_y.cpu().numpy(),
torch.where(test_adv_y_pred > 0.5, torch.ones_like(test_adv_y_pred),
torch.zeros_like(
test_adv_y_pred)).cpu().detach().numpy())
total_test_auc += test_auc
total_test_acc += test_acc
total_test_advauc += test_advauc
total_test_advacc += test_advacc
test_dataloader_len = len(test_dataloader)
train_dataloader_len = len(train_dataloader)
print("Epoch:{}\n test_auc:{}, test_acc:{}, test_adv_auc:{}, test_adv_acc:{}".format(epoch, float(
total_test_auc / test_dataloader_len), float(total_test_acc / test_dataloader_len), float(
total_test_advauc / test_dataloader_len), float(total_test_advacc / test_dataloader_len)))
if args.dataset in ['yahoo', 'coat']:
test_dataset_dir_non = os.path.join(workstation_path, 'dataset', args.dataset, 'dev', 'data_nonuniform.csv')
test_dataloader_non = dataload.dataload_ori(test_dataset_dir_non, args.batch_size,
data_feature_path=data_feature_path,
syn=False)
if epoch % 1 == 0:
test_advauc = 0
test_advacc = 0
total_test_auc = 0
total_test_acc = 0
total_test_advauc = 0
total_test_advacc = 0
total_test_logloss = 0
for test_x, test_y in test_dataloader_non:
test_a = x[:, args.user_dim:]
if args.downstream == 'MLP':
test_auc = skm.roc_auc_score(test_y.cpu().numpy(),
torch.argmax(model.predict(test_x), dim=1).cpu().detach().numpy())
test_acc = accuracy_score(test_y.cpu().numpy(),
torch.argmax(model.predict(test_x), dim=1).cpu().detach().numpy())
test_adv_y_pred = model.eval_adv(test_x, test_y)
test_advauc = skm.roc_auc_score(test_y.cpu().numpy(),
torch.argmax(test_adv_y_pred, dim=1).cpu().detach().numpy())
test_advacc = accuracy_score(test_y.cpu().numpy(),
torch.argmax(test_adv_y_pred, dim=1).cpu().detach().numpy())
else:
test_y_pred = model.predict(test_x, test_a)
test_adv_y_pred = model.eval_adv(test_x, test_y, test_a)
test_auc = skm.roc_auc_score(test_y.cpu().numpy(), test_y_pred.cpu().detach().numpy())
test_acc = skm.accuracy_score(test_y.cpu().numpy(),
torch.where(test_y_pred > 0.5, torch.ones_like(test_y_pred),
torch.zeros_like(
test_y_pred)).cpu().detach().numpy())
test_advauc = skm.roc_auc_score(test_y.cpu().numpy(), test_adv_y_pred.cpu().detach().numpy())
test_advacc = skm.accuracy_score(test_y.cpu().numpy(),
torch.where(test_adv_y_pred > 0.5,
torch.ones_like(test_adv_y_pred),
torch.zeros_like(
test_adv_y_pred)).cpu().detach().numpy())
total_test_auc += test_auc
total_test_acc += test_acc
total_test_advauc += test_advauc
total_test_advacc += test_advacc
test_dataloader_len = len(test_dataloader_non)
train_dataloader_len = len(train_dataloader)
print("Epoch:{}\n test_auc_non:{}, test_acc_non:{}, test_adv_auc_non:{}, test_adv_acc_non:{}".format(epoch,
float(
total_test_auc / test_dataloader_len),
float(
total_test_acc / test_dataloader_len),
float(
total_test_advauc / test_dataloader_len),
float(
total_test_advacc / test_dataloader_len)))
'''
print("Epoch:{}\n train_auc:{}, train_acc:{}, train_logloss:{}, train_ds_loss:{}, train_kl:{}, train_iv_loss:{}".\
format(epoch, float(total_auc / train_dataloader_len), float(total_acc / train_dataloader_len),\
float(total_logloss / train_dataloader_len), float(total_ds_loss / train_dataloader_len), \
float(total_kl / train_dataloader_len), float(total_iv_loss / iv_count)))
'''