-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_tsformer.py
161 lines (154 loc) · 8.17 KB
/
train_tsformer.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
from util import *
import torch
import torch.nn as nn
import numpy as np
import argparse
from torch.utils.data import DataLoader
from model import TSFormer
from engine import tsformer_trainer
import time
import os
parser = argparse.ArgumentParser()
parser.add_argument("--data_path", type=str, default='../data/METR-LA')
parser.add_argument('--device', type=str, default='cuda:0')
parser.add_argument("--history_seq_len", type=int, default=2016)
parser.add_argument('--input_seq_len', default=12, type=int)
parser.add_argument('--output_seq_len', default=12, type=int)
parser.add_argument('--batch_size', type=int, default=16)
parser.add_argument('--mask_ratio', type=float, default=0.75, help='Mask ratio during pre-training')
parser.add_argument('--patch_size', type=int, default=12)
parser.add_argument('--embed_dim', type=int, default=96, help='Embedding/hidden dim for transformer')
parser.add_argument('--encoder_depth', type=int, default=4, help='Number of transformer blocks in encoder')
parser.add_argument('--decoder_depth', type=int, default=1, help='Number of transformer blocks in decoder')
parser.add_argument('--num_heads', type=int, default=4, help='Number of attention heads in TSFormer')
parser.add_argument('--dropout', type=float, default=0.1)
parser.add_argument('--num_token', type=int, default=168)
parser.add_argument('--in_channel', type=int, default=1)
parser.add_argument('--mlp_ratio', type=int, default=4, help='The width of mlp w.r.t the transformer')
parser.add_argument("--learning_rate", type=float, default=0.0005)
parser.add_argument('--weight_decay', type=float, default=0)
parser.add_argument('--pretrain_epoch', type=int, default=50, help='Pre-train epoch for TSFormer')
parser.add_argument("--print_every", type=int, default=200)
parser.add_argument('--expid', type=str, default='test')
parser.add_argument('--model', type=str, default='TSFormer', help='[TSFormer, DistilFormer]')
parser.add_argument("--pretrained_model", type=str,help='Load a pre-trained model')
parser.add_argument("--data_number", type=int, default=0, help='The number of data used to pretrain. in days')
args = parser.parse_args()
def main(args):
args.num_token = int(args.history_seq_len / 12)
train_dataset = TimeSeriesForecastingDataset(args.data_path+'/data_in2016_out12.pkl',
args.data_path+'/index_in2016_out12.pkl', mode = 'train', data_number = args.data_number)
val_dataset = TimeSeriesForecastingDataset(args.data_path+'/data_in2016_out12.pkl',
args.data_path+'/index_in2016_out12.pkl', mode = 'valid')
test_dataset = TimeSeriesForecastingDataset(args.data_path+'/data_in2016_out12.pkl',
args.data_path+'/index_in2016_out12.pkl', mode = 'test')
scaler_pkl = load_pkl(args.data_path+'/scaler_in2016_out12.pkl')
mean, std = scaler_pkl['args']['mean'], scaler_pkl['args']['std']
scaler = Scaler(mean, std)
train_loader = DataLoader(train_dataset, batch_size = args.batch_size, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size = args.batch_size)
test_loader = DataLoader(test_dataset, batch_size = args.batch_size)
engine = tsformer_trainer(args, scaler)
print("Pretraining TSFormer on %s data" % args.data_path)
print("Train data length %d" % len(train_dataset))
print("Val data length %d" % len(val_dataset))
print("Test data length %d" % len(test_dataset))
if 'METR-LA' in args.data_path:
save_tsformer = 'garage_tsformer/METR-LA/' + args.expid + '/'
if not os.path.exists(save_tsformer):
os.makedirs(save_tsformer)
if 'PEMSD7M' in args.data_path:
save_tsformer = 'garage_tsformer/PEMSD7M/' + args.expid + '/'
if not os.path.exists(save_tsformer):
os.makedirs(save_tsformer)
if args.pretrained_model is not None:
print("Load a pretrained model from %s" % args.pretrained_model)
sd = torch.load(args.pretrained_model, map_location = args.device)
try:
engine.tsformer_model.load_state_dict(sd['model_state_dict'])
except:
engine.tsformer_model.load_state_dict(sd)
val_maes = []
val_rmses = []
val_mapes = []
t1 = time.time()
for y, x in val_loader:
metrics = engine.eval_pretrain(x)
val_maes.append(metrics[0])
val_rmses.append(metrics[1])
val_mapes.append(metrics[2])
t2 = time.time()
print("Evaluate the loaded model on validation data")
print("Time %.4fs, mae %.4f, rmse %.4f, mape %.4f"%(t2-t1, np.mean(val_maes), np.mean(val_rmses), np.mean(val_mapes)))
test_maes = []
test_rmses = []
test_mapes = []
t1 = time.time()
for y, x in test_loader:
metrics = engine.eval_pretrain(x)
test_maes.append(metrics[0])
test_rmses.append(metrics[1])
test_mapes.append(metrics[2])
t2 = time.time()
print("Evaluate the loaded model on test data")
print("Time %.4fs, mae %.4f, rmse %.4f, mape %.4f"%(t2-t1, np.mean(test_maes), np.mean(test_rmses), np.mean(test_mapes)))
val_loss = []
train_loss = []
train_time = []
val_time = []
for ep in range(1, args.pretrain_epoch+1):
epoch_trainloss = []
epoch_trainrmse = []
epoch_trainmape = []
epoch_valloss = []
epoch_valrmse = []
epoch_valmape = []
s1 = time.time()
for i, (y, x) in enumerate(train_loader):
# print(i, time.time() - s1)
# print('x', x.shape)
# print('y', y.shape)
# reconstructed, label = model(x[:,:,:,:2])
# print(x[0,:288,0,1] * 288)
# print(x[0,:288,0,2])
# break
if args.history_seq_len != 2016:
x = x[:,-args.history_seq_len:,:,:]
metrics = engine.pre_train(x)
# print('reconstructed', reconstructed.shape)
# print('label', label.shape)
epoch_trainloss.append(metrics[0])
epoch_trainrmse.append(metrics[1])
epoch_trainmape.append(metrics[2])
if i % args.print_every == 0:
print("Pretrain Epoch %d, Iter %d, MAE %.4f, RMSE %.4f, MAPE %.4f, Time spent %.4fs" % (ep, i, metrics[0], metrics[1], metrics[2], time.time() - s1))
engine.scheduler.step()
s2 = time.time()
train_time.append(s2-s1)
train_loss.append(np.mean(epoch_trainloss))
print("Pretrain Epoch %d, train MAE %.4f, train RMSE %.4f, train MAPE %.4f, Time %.4fs, current lr %.5f" % (ep, train_loss[-1], np.mean(epoch_trainrmse), np.mean(epoch_trainmape), s2-s1, engine.scheduler.get_last_lr()[0]))
t1 = time.time()
for y, x in val_loader:
if args.history_seq_len != 2016:
x = x[:,-args.history_seq_len:,:,:]
metrics = engine.eval_pretrain(x)
epoch_valloss.append(metrics[0])
epoch_valrmse.append(metrics[1])
epoch_valmape.append(metrics[2])
t2 = time.time()
val_time.append(t2-t1)
val_loss.append(np.mean(epoch_valloss))
print("Pretrain Epoch %d, val MAE %.4f, val RMSE %.4f, val MAPE %.4f, Time %.4fs" % (ep, val_loss[-1], np.mean(epoch_valrmse), np.mean(epoch_valmape), t2-t1))
torch.save(engine.tsformer_model.state_dict(), save_tsformer + 'epoch_%d_%.4f.pth' % (ep, val_loss[-1]))
best_val_idx = np.argmin(val_loss)
print("Pre-training finish. ")
print("Average epoch train time %.4fs, val time %.4fs" % (np.mean(train_time), np.mean(val_time)))
print("The epoch with best pre-training validation loss is %d, best loss %.4f" % (best_val_idx, val_loss[best_val_idx]))
engine.tsformer_model.load_state_dict(torch.load(save_tsformer+'epoch_%d_%.4f.pth' % (best_val_idx+1, val_loss[best_val_idx])))
save_path = save_tsformer+'best_pretrained_%s_%.4f.pth' % (args.expid, val_loss[best_val_idx])
torch.save(engine.tsformer_model.state_dict(), save_tsformer + 'best_pretrained_%s_%.4f.pth' % (args.expid, val_loss[best_val_idx]))
print("Saving the best pre-trained TSFormer to %s" % save_path)
np.save(save_tsformer+'train_loss.npy', arr = np.array(train_loss))
np.save(save_tsformer+'val_loss.npy', arr = np.array(val_loss))
if __name__ == '__main__':
main(args)