forked from PattonYu/CUFAR
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_joint.py
178 lines (155 loc) · 6.91 KB
/
train_joint.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
import os
device = '0'
os.environ["CUDA_VISIBLE_DEVICES"] = device
import time
import torch
import torch.nn.functional as F
from src.metrics import get_MSE, get_MAE, get_MAPE
from src.utils import get_dataloader_joint, print_model_parm_nums
from src.args import get_args
from model.UrbanFM import UrbanFM
from model.UrbanODE import UrbanODE
from model.DeepLGR import DeepLGR
from model.CUFAR import CUFAR
from model.FODE import FODE
from model.modules.ODE import *
import numpy as np
args = get_args()
save_path = 'experiments/joint/{}-{}-{}'.format(
args.model,
args.dataset,
args.n_channels)
torch.manual_seed(args.seed)
print("mk dir {}".format(save_path))
os.makedirs(save_path, exist_ok=True)
print('device:', device)
cuda = True if torch.cuda.is_available() else False
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
def get_learning_rate(optimizer):
lr=[]
for param_group in optimizer.param_groups:
lr +=[ param_group['lr'] ]
return lr[-1]
def choose_model():
if args.model == 'CUFAR':
model = CUFAR(height=args.height, width=args.width, use_exf=args.use_exf,
scale_factor=args.scale_factor, channels=args.n_channels,
sub_region= args.sub_region,
scaler_X=args.scaler_X, scaler_Y=args.scaler_Y, args= args)
elif args.model == 'UrbanFM':
model = UrbanFM(in_channels=1, out_channels=1, n_residual_blocks=16,
base_channels= args.n_channels, img_width= args.width,
img_height= args.height, ext_flag= args.use_exf,
scaler_X=args.scaler_X, scaler_Y=args.scaler_Y)
elif args.model == 'FODE':
model = FODE(in_channels=1, out_channels=1, n_residual_blocks=16,
base_channels= args.n_channels, img_width= args.width,
img_height= args.height, ext_flag= args.use_exf,
scaler_X=args.scaler_X, scaler_Y=args.scaler_Y)
elif args.model == 'UrbanODE':
model = UrbanODE(in_channels=1, out_channels=1, n_residual_blocks=16,
base_channels= args.n_channels, img_width= args.width,
img_height= args.height, ext_flag= args.use_exf,
scaler_X=args.scaler_X, scaler_Y=args.scaler_Y)
elif args.model == 'DeepLGR':
model = DeepLGR(in_channels=1, out_channels=1, n_residual_blocks=12,
base_channels= args.n_channels, img_width= args.width,
img_height= args.height, ext_flag= args.use_exf, predictor='td',
scaler_X=args.scaler_X, scaler_Y=args.scaler_Y)
return model
def load_model(task):
load_path = '{}/best_epoch_{}.pt'.format(save_path, task)
print("load from {}".format(load_path))
model_state_dict = torch.load(load_path)["model_state_dict"]
model = choose_model()
model.load_state_dict(model_state_dict)
if cuda:
model = model.cuda()
return model
else:
return model
criterion = F.mse_loss
total_datapath = 'datasets'
train_sequence = ["P1", "P2", "P3", "P4"]
total_mses = {"P1":[np.inf], "P2":[np.inf], "P3":[np.inf], "P4":[np.inf]}
best_epoch = {"P1":0, "P2":0, "P3":0, "P4":0}
task_id = 0
start_time = time.time()
for task in train_sequence[task_id:]:
print('='*15,'Start to train {}'.format(task),'='*15)
task_id += 1
model = choose_model()
if cuda:
model = model.cuda()
print_model_parm_nums(model, args.model)
optimizer = torch.optim.Adam(
model.parameters(), lr= args.lr, betas=(args.b1, args.b1))
train_ds = get_dataloader_joint(args,
datapath= total_datapath, dataset= args.dataset,
batch_size= args.batch_size, mode= 'train', task_id= task_id)
valid_task = get_dataloader_joint(args,
datapath= total_datapath, dataset= args.dataset,
batch_size= 32, mode= 'valid', task_id= task_id)
test_ds = get_dataloader_joint(args,
datapath= total_datapath, dataset= args.dataset,
batch_size= 32, mode= 'test', task_id= task_id)
for epoch in range(0, args.n_epochs):
epoch_start_time = time.time()
train_loss = 0
# training phase
for i, (c_map, f_map, exf) in enumerate(train_ds):
model.train()
optimizer.zero_grad()
pred_f_map = model(c_map, exf) * args.scaler_Y
loss = criterion(pred_f_map, f_map * args.scaler_Y)
loss.backward()
optimizer.step()
train_loss += loss.item() * len(c_map)
train_loss /= len(train_ds.dataset)
# validating phase
model.eval()
val_mse, mse = 0, 0
for j, (c_map, f_map, exf) in enumerate(valid_task):
pred_f_map = model(c_map, exf)
pred = pred_f_map.cpu().detach().numpy() * args.scaler_Y
real = f_map.cpu().detach().numpy() * args.scaler_Y
mse += get_MSE(pred=pred, real=real) * len(c_map)
val_mse = mse / len(valid_task.dataset)
if val_mse < np.min(total_mses[task]):
state = {'model_state_dict': model.state_dict(), 'optimizer': optimizer.state_dict(), 'epoch': epoch, 'task': task}
best_epoch[task] = epoch
torch.save(state, '{}/best_epoch_{}.pt'.format(save_path, task))
total_mses[task].append(val_mse)
# log print
log = ('Task:{}|Epoch:{}|Loss:{:.3f}|Val_MSE\t{:.3f}|Time_Cost:{:.2f}|Best_Epoch:{}|lr:{}'.format(
task, epoch, train_loss,
total_mses[train_sequence[task_id -1]][-1],
time.time() - epoch_start_time, best_epoch[task], get_learning_rate(optimizer)))
print(log)
f = open('{}/train_process.txt'.format(save_path), 'a')
f.write(log+'\n')
model = load_model(task)
model.eval()
total_mse, total_mae, total_mape = 0, 0, 0
for i, (c_map, f_map, eif) in enumerate(test_ds):
pred_f_map = model(c_map, eif)
pred = pred_f_map.cpu().detach().numpy() * args.scaler_Y
real = f_map.cpu().detach().numpy() * args.scaler_Y
total_mse += get_MSE(pred=pred, real=real) * len(c_map)
total_mae += get_MAE(pred=pred, real=real) * len(c_map)
total_mape += get_MAPE(pred=pred, real=real) * len(c_map)
mse = total_mse / len(test_ds.dataset)
mae = total_mae / len(test_ds.dataset)
mape = total_mape / len(test_ds.dataset)
log = ('{} test: MSE={:.6f}, MAE={:.6f}, MAPE={:.6f}'.format(task, mse, mae, mape))
f = open('{}/test_results.txt'.format(save_path), 'a')
f.write(log+'\n')
print(log)
print('*' * 64)
log = (
f'Total running time: {(time.time()-start_time)//60:.0f}mins {(time.time()-start_time)%60:.0f}s')
print(log)
f = open('{}/test_results.txt'.format(save_path), 'a')
f.write(log+'\n')
f.close()
print('*' * 64)