-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
303 lines (253 loc) · 15.9 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
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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
import os
import time
import argparse
import pickle
import random
import numpy as np
import pandas as pd
import torch.nn.functional as F
from torch.optim.lr_scheduler import StepLR
import torch
from utils import newloss_generate_batch_data, get_loc2cat, get_unseen_user, get_adj
from Mymodel.hgat_newloss import Model
def settings(param=[]):
parser = argparse.ArgumentParser()
# data params
parser.add_argument('--path_in', type=str, default='./data/', help="input data path")
parser.add_argument('--path_out', type=str, default='./results/', help="output data path")
parser.add_argument('--data_name', type=str, default='NYC', help="data name")
parser.add_argument('--cat_contained', action='store_false', default=True, help="whether contain category")
parser.add_argument('--out_filename', type=str, default='', help="output data filename")
# train params
parser.add_argument('--gpu', type=str, default='1', help="GPU index to choose")
parser.add_argument('--setting', type=int, default=0, help="0: full, 1: transductive, 2: inductive")
parser.add_argument('--seed', type=int, default=0, help="random seed")
parser.add_argument('--run_num', type=int, default=1, help="run number")
parser.add_argument('--epoch_num', type=int, default=80, help="epoch number")
parser.add_argument('--batch_size', type=int, default=64, help="batch size")
parser.add_argument('--learning_rate', type=float, default=2e-4, help="learning rate")
parser.add_argument('--weight_decay', type=float, default=1e-4, help="weight decay")
parser.add_argument('--evaluate_step', type=int, default=1, help="evaluate step")
parser.add_argument('--lam_c', type=float, default=1, help="category loss term factor")
parser.add_argument('--lam_l', type=float, default=1, help="location loss term factor")
parser.add_argument('--gt_w', type=float, default=0.8, help="gt weight")
parser.add_argument('--sum_loss', action='store_true', default=False, help="sum loss, average loss else-wise")
parser.add_argument('--schedule_w', action='store_true', default=False, help="schedule weight")
# model params
parser.add_argument('--res_factor', type=float, default=0.1, help="residual factor")
parser.add_argument('--self_loop', action='store_true', default=False, help="add self loop")
# embedding
parser.add_argument('--u_dim', type=int, default=20, help="user embedding dimension")
parser.add_argument('--l_dim', type=int, default=200, help="loc embedding dimension")
parser.add_argument('--th_dim', type=int, default=20, help="time hour embedding dimension")
parser.add_argument('--tw_dim', type=int, default=10, help="time week embedding dimension")
parser.add_argument('--c_dim', type=int, default=200, help="category embedding dimension")
# adj
parser.add_argument('--dist', type=float, default=0.1, help="distance for l_adj, km")
parser.add_argument('--weighted', action='store_true', default=False, help="whether use weighted adj")
# l_gat
parser.add_argument('--alpha', type=float, default=0.2, help='Alpha for the leaky_relu.')
parser.add_argument('--num_heads', type=int, default=2, help='Number of head attentions.')
parser.add_argument('--nhid', type=int, default=16, help='Number of hidden units.')
parser.add_argument('--gat_outdim', type=int, default=50, help="gat_outdim dimension")
# rnn
parser.add_argument('--rnn_type', type=str, default='gru', help="rnn type")
parser.add_argument('--rnn_dim', type=int, default=600, help="rnn hidden dimension")
parser.add_argument('--rnn_layer', type=int, default=1, help="rnn layer number")
parser.add_argument('--dropout', type=float, default=0.1, help="drop out for rnn")
# optimizer
parser.add_argument('--step_size', type=int, default=1, help="drop out for rnn")
parser.add_argument('--gamma', type=float, default=0.8, help="drop out for rnn")
if __name__ == '__main__' and param == []:
params = parser.parse_args()
else:
params = parser.parse_args(param)
if not os.path.exists(params.path_out):
os.mkdir(params.path_out)
return params
def train(params, dataset):
torch.manual_seed(params.seed)
torch.cuda.manual_seed(params.seed)
np.random.seed(params.seed)
random.seed(params.seed)
# dataset info
params.uid_size = len(dataset['uid_list'])
params.pid_size = len(dataset['pid_dict'])
params.cid_size = len(dataset['cid_dict']) if params.cat_contained else 0
print(f'============Our Model============')
print(f'params: \n{params}\n')
pid_cid_dict = dataset['pid_cid_dict']
# generate input data
data_train, train_id = dataset['train_data'], dataset['train_id']
data_test, test_id = dataset['test_data'], dataset['test_id']
pid_lat_lon_radians = torch.tensor([[0, 0]] + list(dataset['pid_lat_lon_radians'].values()))
pid_lat_lon = torch.tensor([[0, 0]] + list(dataset['pid_lat_lon'].values()))
params.loc2cat, params.cat2loc, unseen_loc, test_mask_uid_sid, train_mask_uid_sid = \
get_loc2cat(data_train, train_id, data_test, test_id, params.batch_size, params.pid_size, params.setting)
mask_uid_sid = get_unseen_user(data_test, test_id, params.device, params.batch_size, params.cat_contained, loc_unseen=unseen_loc)
# model and optimizer
model = Model(params).to(params.device)
optimizer = torch.optim.Adam(model.parameters(), lr=params.learning_rate, weight_decay=params.weight_decay)
# optimizer = torch.optim.SGD(model.parameters(), lr=params.learning_rate, momentum=0.8)
scheduler = StepLR(optimizer, step_size=params.step_size, gamma=params.gamma)
# iterate epoch
best_info_test = {'epoch': 0,
'R_Cat@1': 0, 'R_Cat@5': 0, 'R_Cat@10': 0, 'R_Cat@20': 0,
'R_Loc@1': 0, 'R_Loc@5': 0, 'R_Loc@10': 0, 'R_Loc@20': 0,
'N_Cat@1': 0, 'N_Cat@5': 0, 'N_Cat@10': 0, 'N_Cat@20': 0,
'N_Loc@1': 0, 'N_Loc@5': 0, 'N_Loc@10': 0, 'N_Loc@20': 0
} # best metrics
print('=' * 10, ' Training')
l_adj, c_adj, lc_adj = get_adj(data_train, train_id, pid_lat_lon, pid_cid_dict, params)
l_adj = torch.tensor(l_adj)
l_adj = l_adj.to(params.device)
c_adj = torch.tensor(c_adj)
c_adj = c_adj.to(params.device)
lc_adj = torch.tensor(lc_adj)
lc_adj = lc_adj.to(params.device)
l_adj_1c = torch.cat((l_adj, lc_adj), dim=1)
l_adj_1c_down = F.pad(lc_adj.T, (0, c_adj.shape[0]))
l_adj_1c = torch.cat((l_adj_1c, l_adj_1c_down), dim=0)
l_pad = torch.zeros(l_adj.shape).to(l_adj.device)
c_pad = torch.zeros(c_adj.shape).to(c_adj.device)
c_h_adj_top = torch.cat((l_pad, lc_adj), dim=1)
c_h_adj_down = torch.cat((lc_adj.T, c_pad), dim=1)
c_h_adj = torch.cat((c_h_adj_top, c_h_adj_down), dim=0)
cc_adj = torch.eye(c_adj.shape[0]).to(c_adj.device)
cc_pad = torch.zeros(c_adj.shape).to(c_adj.device)
cc_adj_top = torch.cat((c_adj, cc_adj), dim=1)
cc_adj_down = torch.cat((cc_adj.T, cc_pad), dim=1)
cc_adj = torch.cat((cc_adj_top, cc_adj_down), dim=0).to(c_adj.device)
model = model.float()
if params.schedule_w:
th_epoch = (1 - params.gt_w) / params.epoch_num
for i in range(params.epoch_num):
if params.schedule_w:
params.gt_w += th_epoch
# print(f'gt_w: {params.gt_w}')
epoch_start_time = time.time()
loss_epoch, l_train_epoch_recall, l_test_epoch_recall, c_train_epoch_recall, c_test_epoch_recall, num_test_batch = \
[], np.zeros(4), np.zeros(4), np.zeros(4), np.zeros(4), 0
l_train_epoch_ndcg, l_test_epoch_ndcg, c_train_epoch_ndcg, c_test_epoch_ndcg = np.zeros(4), np.zeros(4), np.zeros(4), np.zeros(4)
num_train_batch = 0
model.train()
for mask_batch, target_batch, data_batch, info_batch in newloss_generate_batch_data(
data_train, train_id, params.device, params.batch_size,
params.cat_contained, 'train', train_mask_uid_sid, test_mask_uid_sid, mask_uid_sid, params.setting):
l_pred, c_pred = model(data_batch, mask_batch, l_adj, c_h_adj, cc_adj)
l_loss = model.calculate_loss(l_pred, torch.squeeze(target_batch[0]), info_batch[0], info_batch[1], info_batch[2], params.gt_w)
c_loss = model.calculate_loss(c_pred, torch.squeeze(target_batch[2]), info_batch[3], info_batch[4], info_batch[5], params.gt_w)
loss = params.lam_l * l_loss + params.lam_c * c_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_epoch.append(loss.item())
l_train_batch_ndcg = model.calculate_ndcg(l_pred, torch.squeeze(target_batch[0]) - 1)
l_train_batch_recall = model.calculate_recall(l_pred, torch.squeeze(target_batch[0]) - 1)
c_train_batch_ndcg = model.calculate_ndcg(c_pred, torch.squeeze(target_batch[2]) - 1)
c_train_batch_recall = model.calculate_recall(c_pred, torch.squeeze(target_batch[2]) - 1)
l_train_epoch_recall += l_train_batch_recall
l_train_epoch_ndcg += l_train_batch_ndcg
c_train_epoch_recall += c_train_batch_recall
c_train_epoch_ndcg += c_train_batch_ndcg
num_train_batch += 1
with torch.no_grad():
model.eval()
for mask_batch, target_batch, data_batch, unique_loc_batch in newloss_generate_batch_data(
data_test, test_id, params.device, params.batch_size,
params.cat_contained, 'test', train_mask_uid_sid, test_mask_uid_sid, mask_uid_sid, params.setting):
# l_pred, c_pred, c_l_pred = model(data_batch, mask_batch, target_batch[1])
l_pred, c_pred = model(data_batch, mask_batch, l_adj, c_h_adj, cc_adj)
l_test_batch_ndcg = model.calculate_ndcg(l_pred, torch.squeeze(target_batch[0])-1)
l_test_batch_recall = model.calculate_recall(l_pred, torch.squeeze(target_batch[0])-1)
c_test_batch_ndcg = model.calculate_ndcg(c_pred, torch.squeeze(target_batch[2])-1)
c_test_batch_recall = model.calculate_recall(c_pred, torch.squeeze(target_batch[2])-1)
l_test_epoch_recall += l_test_batch_recall
l_test_epoch_ndcg += l_test_batch_ndcg
c_test_epoch_recall += c_test_batch_recall
c_test_epoch_ndcg += c_test_batch_ndcg
num_test_batch += 1
l_train_epoch_recall /= num_train_batch
l_train_epoch_ndcg /= num_train_batch
c_train_epoch_recall /= num_train_batch
c_train_epoch_ndcg /= num_train_batch
l_test_epoch_recall /= num_test_batch
l_test_epoch_ndcg /= num_test_batch
c_test_epoch_recall /= num_test_batch
c_test_epoch_ndcg /= num_test_batch
scheduler.step()
print(f'Epoch {i+1} | loss: {np.mean(loss_epoch):.3f} | TRAIN |==| Recall | Cat: ['
f'{c_train_epoch_recall[0]:.3f}; {c_train_epoch_recall[1]:.3f}; {c_train_epoch_recall[2]:.3f}; {c_train_epoch_recall[3]:.3f}] | Loc: ['
f'{l_train_epoch_recall[0]:.3f}; {l_train_epoch_recall[1]:.3f}; {l_train_epoch_recall[2]:.3f}; {l_train_epoch_recall[3]:.3f}] |==| NDCG | Cat: ['
f'{c_train_epoch_ndcg[0]:.3f}; {c_train_epoch_ndcg[1]:.3f}; {c_train_epoch_ndcg[2]:.3f}; {c_train_epoch_ndcg[3]:.3f}] | Loc: ['
f'{l_train_epoch_ndcg[0]:.3f}; {l_train_epoch_ndcg[1]:.3f}; {l_train_epoch_ndcg[2]:.3f}; {l_train_epoch_ndcg[3]:.3f}]'
f'\nEpoch {i+1} | loss: {np.mean(loss_epoch):.3f} | TEST |==| Recall | Cat: ['
f'{c_test_epoch_recall[0]:.3f}; {c_test_epoch_recall[1]:.3f}; {c_test_epoch_recall[2]:.3f}; {c_test_epoch_recall[3]:.3f}] | Loc: ['
f'{l_test_epoch_recall[0]:.3f}; {l_test_epoch_recall[1]:.3f}; {l_test_epoch_recall[2]:.3f}; {l_test_epoch_recall[3]:.3f}] |==| NDCG | Cat: ['
f'{c_test_epoch_ndcg[0]:.3f}; {c_test_epoch_ndcg[1]:.3f}; {c_test_epoch_ndcg[2]:.3f}; {c_test_epoch_ndcg[3]:.3f}] | Loc: ['
f'{l_test_epoch_ndcg[0]:.3f}; {l_test_epoch_ndcg[1]:.3f}; {l_test_epoch_ndcg[2]:.3f}; {l_test_epoch_ndcg[3]:.3f}]'
f' | Time: '
f'{time.time()-epoch_start_time:.3f}\n')
if params.setting == 2:
if best_info_test["R_Loc@5"] < l_test_epoch_recall[1]:
best_info_test["epoch"] = i + 1
best_info_test["R_Cat@1"] = c_test_epoch_recall[0]
best_info_test["R_Cat@5"] = c_test_epoch_recall[1]
best_info_test["R_Cat@10"] = c_test_epoch_recall[2]
best_info_test["R_Cat@20"] = c_test_epoch_recall[3]
best_info_test["R_Loc@1"] = l_test_epoch_recall[0]
best_info_test["R_Loc@5"] = l_test_epoch_recall[1]
best_info_test["R_Loc@10"] = l_test_epoch_recall[2]
best_info_test["R_Loc@20"] = l_test_epoch_recall[3]
best_info_test["N_Cat@1"] = c_test_epoch_ndcg[0]
best_info_test["N_Cat@5"] = c_test_epoch_ndcg[1]
best_info_test["N_Cat@10"] = c_test_epoch_ndcg[2]
best_info_test["N_Cat@20"] = c_test_epoch_ndcg[3]
best_info_test["N_Loc@1"] = l_test_epoch_ndcg[0]
best_info_test["N_Loc@5"] = l_test_epoch_ndcg[1]
best_info_test["N_Loc@10"] = l_test_epoch_ndcg[2]
best_info_test["N_Loc@20"] = l_test_epoch_ndcg[3]
else:
if best_info_test["R_Loc@1"] < l_test_epoch_recall[0]:
best_info_test["epoch"] = i+1
best_info_test["R_Cat@1"] = c_test_epoch_recall[0]
best_info_test["R_Cat@5"] = c_test_epoch_recall[1]
best_info_test["R_Cat@10"] = c_test_epoch_recall[2]
best_info_test["R_Cat@20"] = c_test_epoch_recall[3]
best_info_test["R_Loc@1"] = l_test_epoch_recall[0]
best_info_test["R_Loc@5"] = l_test_epoch_recall[1]
best_info_test["R_Loc@10"] = l_test_epoch_recall[2]
best_info_test["R_Loc@20"] = l_test_epoch_recall[3]
best_info_test["N_Cat@1"] = c_test_epoch_ndcg[0]
best_info_test["N_Cat@5"] = c_test_epoch_ndcg[1]
best_info_test["N_Cat@10"] = c_test_epoch_ndcg[2]
best_info_test["N_Cat@20"] = c_test_epoch_ndcg[3]
best_info_test["N_Loc@1"] = l_test_epoch_ndcg[0]
best_info_test["N_Loc@5"] = l_test_epoch_ndcg[1]
best_info_test["N_Loc@10"] = l_test_epoch_ndcg[2]
best_info_test["N_Loc@20"] = l_test_epoch_ndcg[3]