-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathpretrain.py
166 lines (144 loc) · 6.93 KB
/
pretrain.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
from cProfile import label
from macpath import split
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch_geometric.data import DataLoader
from datasets import traffic_dataset
from utils import *
import argparse
import yaml
import time
from pathlib import Path
from tqdm import tqdm
import sys
sys.path.append('./model')
sys.path.append('./model/TSFormer')
sys.path.append('./model/Meta_Models')
from meta_patch import *
from TSmodel_TSFormerTST import *
parser = argparse.ArgumentParser()
parser.add_argument('--config_filename', default='config.yaml', type=str,
help='Configuration filename for restoring the model.')
parser.add_argument('--test_dataset', default='pems-bay', type=str)
parser.add_argument('--data_list', default='chengdu',type=str)
parser.add_argument('--gpu', default=0, type = int)
parser.add_argument('--use_A', default='A', type=str)
args = parser.parse_args()
args.gpu=0
seed=7
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed) # if you are using multi-GPU.
np.random.seed(seed) # Numpy module.
random.seed(seed) # Python random module.
torch.manual_seed(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_default_dtype(torch.float32)
# print(time.strftime('%Y-%m-%d %H:%M:%S'), "meta_dim = ", args.meta_dim,"target_days = ", args.target_days)
def train_batch(start,end,model,source_dataset,loss_fn,opt):
total_loss = []
total_mae = []
total_mse = []
total_rmse = []
total_mape = []
model.train()
for idx in range(start,end):
data_i, A_wave = source_dataset[idx]
# [B, N, L, 2]
x, means, stds = data_i.x, data_i.means, data_i.stds
A_wave = torch.tensor(A_wave,dtype=torch.float32)
# remember that the input of TSFormer is [B, N, 2, L]
x = x.permute(0,1,3,2).to(args.device)
#
out_masked_tokens, label_masked_tokens, plot_args = model(x, A_wave)
# only the masked patch is loss target
loss = loss_fn(out_masked_tokens, label_masked_tokens)
opt.zero_grad()
loss.backward()
opt.step()
# unmask
unnorm_out, unnorm_label = unnorm(out_masked_tokens, means, stds), unnorm(label_masked_tokens,means,stds)
# print(unnorm_out.shape, unnorm_label.shape)
# unnorm_out, unnorm_label = unnorm_out.cpu().detach().numpy(), unnorm_label.cpu().detach().numpy()
MSE,RMSE,MAE,MAPE = calc_metric(unnorm_out, unnorm_label)
total_mse.append(MSE.cpu().detach().numpy())
total_rmse.append(RMSE.cpu().detach().numpy())
total_mae.append(MAE.cpu().detach().numpy())
total_mape.append(MAPE.cpu().detach().numpy())
total_loss.append(loss.item())
return total_mse,total_rmse, total_mae, total_mape, total_loss
# print(out_masked_tokens.shape, label_masked_tokens.shape, list(plot_args.keys()))
def test_batch(start,end,model,source_dataset,loss_fn,opt):
total_loss = []
total_mae = []
total_mse = []
total_rmse = []
total_mape = []
with torch.no_grad():
model.eval()
for idx in range(start,end):
data_i, A_wave = source_dataset[idx]
# [B, N, L, 2]
x, means, stds = data_i.x, data_i.means, data_i.stds
A_wave = torch.tensor(A_wave,dtype=torch.float32)
# remember that the input of TSFormer is [B, N, 2, L]
x = x.permute(0,1,3,2).to(args.device)
#
out_masked_tokens, label_masked_tokens, plot_args = model(x, A_wave)
# unmask
unnorm_out, unnorm_label = unnorm(out_masked_tokens, means, stds), unnorm(label_masked_tokens,means,stds)
MSE,RMSE,MAE,MAPE = calc_metric(unnorm_out, unnorm_label)
total_mse.append(MSE.cpu().detach().numpy())
total_rmse.append(RMSE.cpu().detach().numpy())
total_mae.append(MAE.cpu().detach().numpy())
total_mape.append(MAPE.cpu().detach().numpy())
return total_mse,total_rmse, total_mae, total_mape, total_loss
if __name__ == "__main__":
if torch.cuda.is_available():
args.device = torch.device('cuda:{}'.format(args.gpu))
else:
args.device = torch.device('cpu')
print("INFO: {}".format(args.device))
with open(args.config_filename) as f:
config = yaml.load(f)
data_args, task_args, model_args = config['data'], config['task'], config['model']
train_ratio = 0.7
val_ratio = 0.1
test_ratio = 0.2
model_path = Path('./save/pretrain_model/{}/'.format(args.data_list))
if(not os.path.exists(model_path)):
os.makedirs(model_path)
data_list = get_data_list(args.data_list)
source_dataset = traffic_dataset(data_args, task_args['mae'], data_list, "pretrain", test_data=args.test_dataset)
model = TSFormer(model_args['mae'], args).to(args.device)
model.mode = 'Pretrain'
opt = optim.Adam(model.parameters(),lr = task_args['mae']['lr'])
loss_fn = nn.MSELoss(reduction = 'mean')
batch_size = task_args['mae']['batch_size']
print('pretrain model has {} parameters, batch_size : {}'.format(count_parameters(model), batch_size))
best_loss = 9999999999999.0
best_model = None
for i in range(task_args['mae']['train_epochs']):
length = source_dataset.__len__()
# length=40
train_length = int(train_ratio * length)
val_length = int(val_ratio * length)
print('----------------------')
time_1 = time.time()
total_mse,total_rmse, total_mae, total_mape, total_loss = train_batch(0,train_length, model, source_dataset,loss_fn,opt)
print('Epochs {}/{}'.format(i,task_args['mae']['train_epochs']))
print('in training Unnormed MSE : {:.5f}, RMSE : {:.5f}, MAE : {:.5f}, MAPE: {:.5f}, normed MSE : {:.5f}.'.format(np.mean(total_mse), np.mean(total_rmse), np.mean(total_mae),np.mean(total_mape),np.mean(total_loss)))
total_mse,total_rmse, total_mae, total_mape, total_loss = test_batch(train_length,train_length + val_length, model, source_dataset,loss_fn,opt)
print('in validation Unnormed MSE : {:.5f}, RMSE : {:.5f}, MAE : {:.5f}, MAPE: {:.5f}.'.format(np.mean(total_mse), np.mean(total_rmse), np.mean(total_mae),np.mean(total_mape)))
mae_loss = np.mean(total_mae)
if(mae_loss < best_loss):
best_model = model
best_loss = mae_loss
torch.save(model.state_dict(), model_path / 'best_model.pt')
print('Best model. Saved.')
print('this epoch costs {:.5}s'.format(time.time()-time_1))
total_mse,total_rmse, total_mae, total_mape, total_loss = test_batch(train_length + val_length,length,model, source_dataset,loss_fn,opt)
print('test Unnormed MSE : {:.5f}, RMSE : {:.5f}, MAE : {:.5f}, MAPE: {:.5f}'.format(np.mean(total_mse), np.mean(total_rmse), np.mean(total_mae),np.mean(total_mape)))