/
main.py
241 lines (212 loc) · 12.1 KB
/
main.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
233
234
235
236
237
238
239
240
241
import argparse
import os
import evaluate
import numpy as np
import scipy.sparse as sp
import torch
import torch.backends.cudnn as cudnn
import torch.utils.data as data
from utils import set_seed
from datasets import MultiDataset
from datasets import train_collate_fn
from data_utils import create_adj_mat, load_all, load_pretrain
from model import MF
from LightGCN import LightGCN
from train import TrainModel
def main(param):
seed = param['seed']
set_seed(seed)
cudnn.benchmark = True
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True # cudnn
def worker_init_fn(worker_id):
np.random.seed(param['seed'] + worker_id)
if not os.path.exists(param['folder']):
os.mkdir(param['folder'])
###################### PREPARE DATASET ##########################
data_path = f"{param['datadir']}/{param['dataset']}/"
# load data
if param['train_method'] == "pre":
user_num, item_num, \
train_mat_pv, train_mat_buy, \
user_pos_dict_pv, user_pos_dict_buy, \
train_data_dict_pv, train_data_dict_buy, \
test_data_dict_pv, test_data_dict_buy = load_pretrain(data_path, param['dataset'])
else:
user_num, item_num, \
train_mat_pv, train_mat_buy, \
user_pos_dict_pv, user_pos_dict_buy, \
train_data_dict_pv, train_data_dict_buy, \
test_data_dict_buy = load_all(data_path, param['dataset'])
train_dataset = MultiDataset(user_num=user_num, item_num=item_num,
train_mat_pv=train_mat_pv, train_mat_buy=train_mat_buy,
user_pos_dict_pv=user_pos_dict_pv, user_pos_dict_buy=user_pos_dict_buy,
train_data_dict_pv=train_data_dict_pv, train_data_dict_buy=train_data_dict_buy,
is_pretrain=(param['train_method'] == "pre"))
train_loader = data.DataLoader(train_dataset, batch_size=param.get('batch_size', 2048),
shuffle=True, num_workers=0, pin_memory=True,
worker_init_fn=worker_init_fn, collate_fn=train_collate_fn)
########################### CREATE MODEL #################################
pretrain_model = param['pretrain_model']
model_name = param['model']
try:
norm_adj_buy = sp.load_npz(data_path + param['dataset'] + '_s_pre_adj_mat_buy.npz')
norm_adj_pv = sp.load_npz(data_path + param['dataset'] + '_s_pre_adj_mat_pv.npz')
print("successfully loaded...")
except:
norm_adj_buy = create_adj_mat(train_dataset.train_mat_buy,
train_dataset.user_num, train_dataset.item_num,
data_path, dataset=param['dataset'], mode="buy")
norm_adj_pv = create_adj_mat(train_dataset.train_mat_pv,
train_dataset.user_num, train_dataset.item_num,
data_path, dataset=param['dataset'], mode="pv")
param['norm_adj_buy'] = norm_adj_buy
param['norm_adj_pv'] = norm_adj_pv
if pretrain_model == 'lgn':
model_buy = LightGCN(train_dataset.user_num, train_dataset.item_num,
norm_adj_buy, latent_dim=param['emb_dim'], n_layers=param['num_layers'],
device=param['device'], dropout=param['dropout'])
model_pv = LightGCN(train_dataset.user_num, train_dataset.item_num,
norm_adj_pv, latent_dim=param['emb_dim'], n_layers=param['num_layers'],
device=param['device'], dropout=param['dropout'])
elif pretrain_model == 'MF':
model_buy = MF(train_dataset.user_num, train_dataset.item_num,
param.get("factor_num", 32))
model_pv = MF(train_dataset.user_num, train_dataset.item_num,
param.get("factor_num", 32))
else:
assert False, "模型未实现"
train_model = TrainModel(param=param,
user_num=train_dataset.user_num, item_num=train_dataset.item_num,
train_mat_buy=train_dataset.train_mat_buy, train_mat_pv=train_dataset.train_mat_pv)
data_name = param['dataset']
lambda0 = str(param['lambda0'])
idx = param['idx']
pv_load_model_path = os.path.join(param['folder'],
f"pretrain_pv_{pretrain_model}_{data_name}_{seed}_{lambda0}_{idx}.pt")
buy_load_model_path = os.path.join(param['folder'],
f"pretrain_buy_{pretrain_model}_{data_name}_{seed}_{lambda0}_{idx}.pt")
if param['train_method'] == "pre":
if param['test_only']:
model_buy.load_state_dict(torch.load(buy_load_model_path, map_location=torch.device('cpu')))
recall, NDCG = evaluate.test_all_users(model_buy, 4096, test_data_dict_buy, user_pos_dict_buy,
train_model.top_k, device=train_model.device)
final_perf = "TEST\t recall=[%s], ndcg=[%s]" % \
('\t'.join(['%.4f' % r for r in recall]),
'\t'.join(['%.4f' % r for r in NDCG]))
print(final_perf)
else:
train_model.pretrain(model=model_pv, train_loader=train_loader,
train_mode="pv",
test_data_pos=test_data_dict_pv, user_pos=user_pos_dict_pv,
model_name=pretrain_model, early_stop_rounds=param['pretrain_early_stop_rounds'],
model_save_path=pv_load_model_path)
set_seed(seed)
train_model.pretrain(model=model_buy, train_loader=train_loader,
train_mode="buy",
test_data_pos=test_data_dict_buy, user_pos=user_pos_dict_buy,
model_name=pretrain_model, early_stop_rounds=param['pretrain_early_stop_rounds'],
model_save_path=buy_load_model_path)
elif param['train_method'] == "mba":
is_exist_pretrain = os.path.exists(pv_load_model_path)
assert is_exist_pretrain, "点击预训练模型路径不存在"
is_exist_pretrain = os.path.exists(buy_load_model_path)
assert is_exist_pretrain, "购买预训练模型路径不存在"
if param['test_only']:
model_buy.load_state_dict(torch.load(buy_load_model_path, map_location=torch.device('cpu')))
if pretrain_model == 'MF':
target_buy = MF(train_model.user_num, train_model.item_num, param.get("factor_num", 32))
else:
target_buy = LightGCN(train_dataset.user_num, train_dataset.item_num,
norm_adj_buy, latent_dim=param['emb_dim'], n_layers=param['num_layers'],
device=param['device'], dropout=param['dropout'])
target_model_path = os.path.join(param['folder'],
f"{param['dataset']}_target_{model_name}_{seed}_{str(param['lambda1'])}.pt")
target_buy.load_state_dict(torch.load(target_model_path, map_location=torch.device('cpu')))
recall, NDCG = evaluate.test_all_users_with_two_model(target_buy, model_buy, 4096,
test_data_dict_buy, user_pos_dict_buy, train_model.top_k,
device=train_model.device,
beta1=train_model.beta1, beta2=train_model.beta2)
final_perf = "TEST\t recall=[%s], ndcg=[%s]" % \
('\t'.join(['%.4f' % r for r in recall]),
'\t'.join(['%.4f' % r for r in NDCG]))
print(final_perf)
else:
model_pv.load_state_dict(torch.load(pv_load_model_path, map_location=torch.device('cpu')))
model_buy.load_state_dict(torch.load(buy_load_model_path, map_location=torch.device('cpu')))
train_model.train(pv_pre_model=model_pv, buy_pre_model=model_buy, train_loader=train_loader,
test_data_pos=test_data_dict_buy, user_pos=user_pos_dict_buy,
data_source='buy', model_name=model_name)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# flexible parameters
parser.add_argument('--datadir', type=str, default=r'./data')
parser.add_argument('--folder', type=str, default='./output')
parser.add_argument('--model', type=str, default='MF')
parser.add_argument('--h_model', type=str, default='MF')
parser.add_argument('--dataset', type=str, default='beibei',
help='dataset used for training, options: beibei, taobao')
parser.add_argument("--epochs", type=int, default=400, help="training epoches")
parser.add_argument("--top_k", type=list, default=[10, 20], help="compute metrics@top_k")
parser.add_argument("--C_1", default=1000, type=int, help='the large number used in DP')
parser.add_argument("--C_2", default=1000, type=int, help='the large number used in DN')
parser.add_argument("--alpha", type=float, default=1.0, help='weight between two KL divergence')
parser.add_argument("--lambda0", type=float, default=1e-4, help='regularization parameter')
parser.add_argument("--lambda1", type=float, default=1e-6, help='regularization parameter')
parser.add_argument("--save_model", type=int, default=1)
parser.add_argument("--seed", type=int, default=2020)
# following parameters are fixed during my implementation
parser.add_argument("--dropout", type=float, default=0.0, help="dropout rate")
parser.add_argument("--lr", type=float, default=0.001, help="learning rate")
parser.add_argument("--batch_size", type=int, default=2048, help="batch size for training")
parser.add_argument("--factor_num", type=int, default=32, help="predictive factors numbers in the model")
parser.add_argument("--num_layers", type=int, default=3, help="number of layers in MLP model")
parser.add_argument("--NSR", type=int, default=1, help="sample negative items for training")
parser.add_argument("--device", default='cuda', help='cuda or cpu')
parser.add_argument("--emb_dim", type=int, default=32, help='embedding dimension of the Gamma model')
parser.add_argument("--early_stop_rounds", type=int, default=30,
help='early stop after how many iteration rounds non decreasing')
# 新加入的参数
parser.add_argument("--train_method", type=str, default="mba", help="mba: MBA, pre: pretrain")
parser.add_argument("--pretrain_early_stop_rounds", type=int, default=20)
parser.add_argument("--idx", type=int, default=0)
parser.add_argument("--pretrain_model", type=str, default="MF")
parser.add_argument("--denoise_type", type=str, default="DP")
parser.add_argument("--beta", type=float, default=0.8)
parser.add_argument('--test_only', action='store_true')
args = parser.parse_args()
main({
'datadir': args.datadir,
'folder': args.folder,
'model': args.model,
'h_model': args.h_model,
'dataset': args.dataset,
'epochs': args.epochs,
'top_k': args.top_k,
'C_2': args.C_2,
'C_1': args.C_1,
'alpha': args.alpha,
'lambda0': args.lambda0,
'lambda1': args.lambda0,
'save_model': args.save_model,
"seed": args.seed,
'dropout': args.dropout,
'lr': args.lr,
'batch_size': args.batch_size,
'factor_num': args.factor_num,
'num_layers': args.num_layers,
'NSR': args.NSR,
'device': args.device,
'emb_dim': args.emb_dim,
'early_stop_rounds': args.early_stop_rounds,
# 新加入的参数
'train_method': args.train_method,
'pretrain_early_stop_rounds': args.pretrain_early_stop_rounds,
'idx': args.idx,
'pretrain_model': args.pretrain_model,
'denoise_type': args.denoise_type,
'beta1': args.beta,
'beta2': 1.0 - args.beta,
'test_only': args.test_only
})