-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
200 lines (162 loc) · 9.41 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
import os
import time
import torch
import numpy as np
from models.model import Model
from datasets.data_loader import load_dataset, load_graph_data
from utils.earlystopping import EarlyStopping
from utils.metrices import metric, masked_mae_loss
class Training():
def __init__(self, args, file_logger):
self.args = args
self.adj, self.in_degree, self.out_degree = load_graph_data(os.path.join('datasets', args.dataset, 'graph_data.pkl'))
self.adj = torch.from_numpy(self.adj).to(args.device)
self.in_degree = self.in_degree.to(args.device)
self.out_degree = self.out_degree.to(args.device)
self.load_dataset = load_dataset
self.file_logger = file_logger
# 模型
self.model = Model(self.args).to(args.device)
# 损失函数
self.criterion = masked_mae_loss
# Adam优化器
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=args.learning_rate,
weight_decay=args.weight_decay, eps=args.eps)
# learning rate scheduler
self.lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(self.optimizer, milestones=args.lr_sche_steps,
gamma=args.lr_decay_ratio) if args.if_lr_scheduler else None
def train(self):
# 训练集、验证集数据加载
train_loader, train_scaler = self.load_dataset(os.path.join('./datasets', self.args.dataset, str(self.args.seq_len)),
self.args.batch_size, 'train')
valid_loader, valid_scaler = self.load_dataset(os.path.join('./datasets', self.args.dataset, str(self.args.seq_len)),
self.args.batch_size, 'val')
print("Whole trainining iteration is " + str(train_loader.num_batch))
early_stopping = EarlyStopping(self.args.patience, self.args.save_path, self.file_logger)
train_time, val_time = [], []
batch_num = 0
# ========================== Train ==========================
for epoch in range(self.args.epochs):
start_time = time.time()
train_loss, train_mae, train_rmse = [], [], []
for i, (batch_x, batch_y) in enumerate(train_loader.get_iterator()):
self.model.train()
self.optimizer.zero_grad()
# PEMS-BAY数据集296特征 构造时间戳信息
if self.args.input_dim == 296 and self.args.dataset == 'PEMS-BAY':
B, T, N, _ = batch_x.shape
time_of_day = []
for b in range(B):
tmp = np.eye(288, dtype=np.float32)[batch_x[b, :, 0, 1].astype(np.int32)]
tmp = np.tile(tmp, (N, 1, 1)).transpose((1, 0, 2))
time_of_day.append(tmp)
time_of_day = np.stack(time_of_day, axis=0)
batch_x = np.concatenate([batch_x[:, :, :, 0:1], time_of_day, batch_x[:, :, :, 2:]], axis=-1)
batch_x = torch.tensor(batch_x).float().to(self.args.device)
batch_y = torch.tensor(batch_y).float().to(self.args.device)
output = self.model(batch_x, self.adj, self.in_degree, self.out_degree)
predict = train_scaler.inverse_transform(output)
ground_truth = train_scaler.inverse_transform(batch_y)
loss = self.criterion(predict, ground_truth) # 计算损失
loss.backward()
# 梯度裁剪
if self.args.clip is not None:
torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.args.clip)
self.optimizer.step()
mae, rmse = metric(predict, ground_truth)
train_loss.append(loss.item())
train_mae.append(mae)
train_rmse.append(rmse)
print('{}: {}'.format(i, mae), end='\r', flush=True)
batch_num += 1
end_time = time.time() - start_time
train_time.append(end_time - start_time)
cur_lr = self.optimizer.param_groups[0]['lr']
if self.lr_scheduler:
self.lr_scheduler.step()
mtrain_loss = np.mean(train_loss)
mtrain_mae = np.mean(train_mae)
mtrain_rmse = np.mean(train_rmse)
# ========================== Valid ==========================
valid_start_time = time.time()
mvalid_loss, mvalid_mae, mvalid_rmse = self.valid(valid_loader, valid_scaler)
valid_end_time = time.time()
val_time.append(valid_end_time - valid_start_time)
self.file_logger.info(' | Epoch: {:03d} | Train_Loss: {:.4f} | Train_MAE: {:.4f} | Train_RMSE: {:.4f} | Valid_Loss: {:.4f} | Valid_RMSE: {:.4f} | Valid_MAE: {:.4f} | LR: {:.6f}'.format(
epoch, mtrain_loss, mtrain_mae, mtrain_rmse, mvalid_loss, mvalid_rmse, mvalid_mae, cur_lr))
# Early Stopping
early_stopping(mvalid_loss, self.model)
if early_stopping.early_stop:
self.file_logger.info('Early Stopping !!')
break
# ========================== Test ==========================
del train_loader, train_scaler, valid_loader, valid_scaler
test_loader, test_scaler = self.load_dataset(os.path.join('./datasets', self.args.dataset, str(self.args.seq_len)),
self.args.batch_size, 'test')
self.model.load_state_dict(torch.load(self.args.save_path))
self.test(self.model, self.adj, self.in_degree, self.out_degree, test_loader, test_scaler, self.args, self.file_logger)
def valid(self, valid_loader, valid_scaler):
valid_loss, valid_mae, valid_rmse = [], [], []
self.model.eval()
for i, (batch_x, batch_y) in enumerate(valid_loader.get_iterator()):
# PEMS-BAY数据集296特征 构造时间戳信息
if self.args.input_dim == 296 and self.args.dataset == 'PEMS-BAY':
B, T, N, _ = batch_x.shape
time_of_day = []
for b in range(B):
tmp = np.eye(288, dtype=np.float32)[batch_x[b, :, 0, 1].astype(np.int32)]
tmp = np.tile(tmp, (N, 1, 1)).transpose((1, 0, 2))
time_of_day.append(tmp)
time_of_day = np.stack(time_of_day, axis=0)
batch_x = np.concatenate([batch_x[:, :, :, 0:1], time_of_day, batch_x[:, :, :, 2:]], axis=-1)
batch_x = torch.tensor(batch_x).float().to(self.args.device)
batch_y = torch.tensor(batch_y).float().to(self.args.device)
output = self.model(batch_x, self.adj, self.in_degree, self.out_degree)
predict = valid_scaler.inverse_transform(output)
ground_truth = valid_scaler.inverse_transform(batch_y)
loss = self.criterion(predict, ground_truth)
mae, rmse = metric(predict, ground_truth)
print("validation: {}".format(mae), end='\r', flush=True)
valid_loss.append(loss.item())
valid_mae.append(mae)
valid_rmse.append(rmse)
mvalid_loss = np.mean(valid_loss)
mvalid_mae = np.mean(valid_mae)
mvalid_rmse = np.mean(valid_rmse)
return mvalid_loss, mvalid_mae, mvalid_rmse
@staticmethod
def test(model, adj, in_degree, out_degree, test_loader, test_scaler, params, file_logger):
file_logger.info('Begin testing ...')
model.eval()
predict = []
ground_truth = []
for i, (batch_x, batch_y) in enumerate(test_loader.get_iterator()):
# PEMS-BAY数据集296特征 构造时间戳信息
if params.input_dim == 296 and params.dataset == 'PEMS-BAY':
B, T, N, _ = batch_x.shape
time_of_day = []
for b in range(B):
tmp = np.eye(288, dtype=np.float32)[batch_x[b, :, 0, 1].astype(np.int32)]
tmp = np.tile(tmp, (N, 1, 1)).transpose((1, 0, 2))
time_of_day.append(tmp)
time_of_day = np.stack(time_of_day, axis=0)
batch_x = np.concatenate([batch_x[:, :, :, 0:1], time_of_day, batch_x[:, :, :, 2:]], axis=-1)
batch_x = torch.tensor(batch_x).float().to(params.device)
batch_y = torch.tensor(batch_y).float().to(params.device)
output = model(batch_x, adj, in_degree, out_degree)
output = test_scaler.inverse_transform(output)
batch_y = test_scaler.inverse_transform(batch_y)
predict.append(output.detach())
ground_truth.append(batch_y.detach())
predict = torch.cat(predict, dim=0)
ground_truth = torch.cat(ground_truth, dim=0)
# print('saving...')
# np.save('predictions' + str(params.seq_len) + '.npy', [predict.detach().cpu().numpy(), ground_truth.detach().cpu().numpy()])
mae, rmse = metric(predict, ground_truth)
file_logger.info('(On average over {} horizons) Test MAE: {:.2f} | Test RMSE: {:.2f}'.format(params.seq_len, mae, rmse))
if params.show_step_err:
for i in range(params.seq_len):
pred = predict[:, i, :, :]
real = ground_truth[:, i, :, :]
step_mae, step_rmse = metric(pred, real)
file_logger.info('Horizon {}: Test MAE: {:.2f} | Test RMSE: {:.2f}'.format(i, step_mae, step_rmse))