-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathexp_sea_h.py
executable file
·228 lines (198 loc) · 10.3 KB
/
exp_sea_h.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
import os
from timeit import default_timer
from datetime import datetime, timedelta
import numpy as np
import matplotlib.pyplot as plt
from utils.data_factory import SeaDataset, SeaDatasetMemory
from utils.utilities3 import *
from utils.params import get_args
from utils.adam import Adam
from model_dict import get_model
from tqdm import tqdm
time_str = (datetime.now()).strftime("%Y%m%d_%H%M%S")
torch.manual_seed(0)
np.random.seed(0)
torch.cuda.manual_seed(0)
torch.backends.cudnn.deterministic = True
################################################################
# configs
################################################################
args = get_args(time=time_str)
padding = [int(p) for p in args.padding.split(',')]
ntrain = args.ntrain
ntest = args.ntest
N = args.ntotal
args.in_channels = args.in_dim * args.in_var
args.out_channels = args.out_dim * args.out_var
r1 = args.h_down
r2 = args.w_down
s1 = int(((args.h - 1) / r1) + 1)
s2 = int(((args.w - 1) / r2) + 1)
T_in = args.T_in
T_out = args.T_out
patch_size = tuple(int(x) for x in args.patch_size.split(','))
batch_size = args.batch_size
learning_rate = args.learning_rate
epochs = args.epochs
step_size = args.step_size
gamma = args.gamma
model_save_path = args.model_save_path
model_save_name = args.model_save_name
################################################################
# models
################################################################
model = get_model(args)
################################################################
# load data and data normalization
################################################################
train_dataset = SeaDatasetMemory(args, region=args.region, split='train')
test_dataset = SeaDatasetMemory(args, region=args.region, split='test')
train_loader = train_dataset.loader()
test_loader = test_dataset.loader()
land, sea = get_land_sea_mask(args.data_path, args.fill_value)
if 'DeepLag' in args.model:
model.set_bdydom(land, sea)
if args.resample_strategy == 'uniform' or args.resample_strategy == 'learned':
model.num_samples = min(model.num_samples, s1*s2)
elif args.resample_strategy == 'boundary':
model.num_samples = min(model.num_samples, model.coo_boundary_ms[0].shape[0])
elif args.resample_strategy == 'domain':
model.num_samples = min(model.num_samples, model.coo_domain_ms[0].shape[0])
################################################################
# training and evaluation
################################################################
optimizer = Adam(model.parameters(), lr=learning_rate, weight_decay=1e-4)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=epochs)
myloss = LpLoss(size_average=False, channel_wise=False)
step = 1
min_test_l2_full = 114514
for ep in range(epochs):
model.train()
t1 = default_timer()
train_l2_step = 0
train_l2_full = 0
for xx, yy in tqdm(train_loader):
loss = 0
xx = xx.to(device)
yy = yy.to(device)
if 'DeepLag' in args.model:
h_x_q, h_coo_q, h_coo_offset_q = [], [], []
for i in range(model.num_layers):
if args.resample_strategy == 'uniform':
num_samples = model.num_samples // (4**i)
coo_q = torch.cat([
torch.randint(0,model.img_h_layers[i]-1,(batch_size,num_samples,1)),
torch.randint(0,model.img_w_layers[i]-1,(batch_size,num_samples,1))
], dim=-1).to(torch.float32)
elif args.resample_strategy == 'boundary':
num_samples = min(model.num_samples//(4**i), model.coo_boundary_ms[i].shape[0])
idx_coo_sample = torch.multinomial(1./torch.ones(model.coo_boundary_ms[i].shape[0]), num_samples, replacement=False) # k
coo_q = model.coo_boundary_ms[i][idx_coo_sample][None, ...].repeat(batch_size,1,1).to(torch.float32) # b k 2
elif args.resample_strategy == 'domain':
num_samples = min(model.num_samples//(4**i), model.coo_domain_ms[i].shape[0])
idx_coo_sample = torch.multinomial(1./torch.ones(model.coo_domain_ms[i].shape[0]), num_samples, replacement=False) # k
coo_q = model.coo_domain_ms[i][idx_coo_sample][None, ...].repeat(batch_size,1,1).to(torch.float32) # b k 2
elif args.resample_strategy == 'learned':
num_samples = model.num_samples // (4**i)
coo_q = None
num_chan = args.d_model*(2**i) if i < model.num_layers-1 else args.d_model*(2**(i-1))
h_x_q.append(torch.zeros(batch_size, num_samples, num_chan).to(device))
h_coo_q.append(coo_q.to(device) if args.resample_strategy != 'learned' else None)
h_coo_offset_q.append(torch.zeros(batch_size, num_samples, 2).to(device))
for t in range(0, T_out, step):
y = yy[..., t*args.out_var : (t + step)*args.out_var]
if 'DeepLag' in args.model:
im, h_x_q, h_coo_q, h_coo_offset_q, coo_offset_xys = model(xx, h_x_q, h_coo_q, h_coo_offset_q) # B H W T_out*C_out
else:
im = model(xx)
# print(xx.shape, y.shape)
loss += myloss(im, y)
if t == 0:
pred = im
else:
pred = torch.cat((pred, im), -1)
xx = torch.cat((xx[..., step*args.in_var:], im), dim=-1)
train_l2_step += loss.item()
l2_full = myloss(pred, yy)
train_l2_full += l2_full.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
test_l2_step = 0
test_l2_full = 0
with torch.no_grad():
for xx, yy in test_loader:
loss = 0
xx = xx.to(device)
yy = yy.to(device)
if 'DeepLag' in args.model:
h_x_q, h_coo_q, h_coo_offset_q = [], [], []
for i in range(model.num_layers):
if args.resample_strategy == 'uniform':
num_samples = model.num_samples // (4**i)
coo_q = torch.cat([
torch.randint(0,model.img_h_layers[i]-1,(batch_size,num_samples,1)),
torch.randint(0,model.img_w_layers[i]-1,(batch_size,num_samples,1))
], dim=-1).to(torch.float32) # b k 2
elif args.resample_strategy == 'boundary':
num_samples = min(model.num_samples//(4**i), model.coo_boundary_ms[i].shape[0])
idx_coo_sample = torch.multinomial(1./torch.ones(model.coo_boundary_ms[i].shape[0]), num_samples, replacement=False) # k
coo_q = model.coo_boundary_ms[i][idx_coo_sample][None, ...].repeat(batch_size,1,1).to(torch.float32) # b k 2
elif args.resample_strategy == 'domain':
num_samples = min(model.num_samples//(4**i), model.coo_domain_ms[i].shape[0])
idx_coo_sample = torch.multinomial(1./torch.ones(model.coo_domain_ms[i].shape[0]), num_samples, replacement=False) # k
coo_q = model.coo_domain_ms[i][idx_coo_sample][None, ...].repeat(batch_size,1,1).to(torch.float32) # b k 2
elif args.resample_strategy == 'learned':
num_samples = model.num_samples // (4**i)
coo_q = None
num_chan = args.d_model*(2**i) if i < model.num_layers-1 else args.d_model*(2**(i-1))
h_x_q.append(torch.zeros(batch_size, num_samples, num_chan).to(device))
h_coo_q.append(coo_q.to(device) if args.resample_strategy != 'learned' else None)
h_coo_offset_q.append(torch.zeros(batch_size, num_samples, 2).to(device))
for t in range(0, T_out, step):
y = yy[..., t*args.out_var : (t + step)*args.out_var]
if 'DeepLag' in args.model:
im, h_x_q, h_coo_q, h_coo_offset_q, coo_offset_xys = model(xx, h_x_q, h_coo_q, h_coo_offset_q) # B H W T_out*C_out
else:
im = model(xx)
loss += myloss(im, y)
if t == 0:
pred = im
else:
pred = torch.cat((pred, im), -1)
xx = torch.cat((xx[..., step*args.in_var:], im), dim=-1)
test_l2_step += loss.item()
test_l2_full += myloss(pred, yy).item()
t2 = default_timer()
scheduler.step()
if test_l2_full / ntest < min_test_l2_full:
min_test_l2_full = test_l2_full / ntest
print(ep, t2 - t1, train_l2_step / ntrain / (T_out / step), train_l2_full / ntrain,
test_l2_step / ntest / (T_out / step),
test_l2_full / ntest, 'new_best!')
print('save best model')
torch.save(model.state_dict(), os.path.join(args.run_save_path, model_save_name[:-3]+f'_best.pt'))
pd = pred[-1, :, :, -5:].detach().cpu().numpy()
gt = yy[-1, :, :, -5:].detach().cpu().numpy()
vars = ['thetao', 'so', 'uo', 'vo', 'zos']
for i in range(5):
visual(pd[...,i], os.path.join(args.run_save_path, f'best_{vars[i]}_pred.png'))
visual(gt[...,i], os.path.join(args.run_save_path, f'best_{vars[i]}_gt.png'))
visual(np.abs(gt-pd)[...,i], os.path.join(args.run_save_path, f'best_{vars[i]}_err.png'))
else:
print(ep, t2 - t1, train_l2_step / ntrain / (T_out / step), train_l2_full / ntrain,
test_l2_step / ntest / (T_out / step),
test_l2_full / ntest)
if ep % 10 == 0:
# if not os.path.exists(model_save_path):
# os.makedirs(model_save_path)
print('save latest model')
torch.save(model.state_dict(), os.path.join(args.run_save_path, model_save_name[:-3]+f'_latest.pt'))
if ep % 100 == 0:
pd = pred[-1, :, :, -5:].detach().cpu().numpy()
gt = yy[-1, :, :, -5:].detach().cpu().numpy()
vars = ['thetao', 'so', 'uo', 'vo', 'zos']
for i in range(5):
visual(pd[...,i], os.path.join(args.run_save_path, f'ep_{ep}_{vars[i]}_pred.png'))
visual(gt[...,i], os.path.join(args.run_save_path, f'ep_{ep}_{vars[i]}_gt.png'))
visual(np.abs(gt-pd)[...,i], os.path.join(args.run_save_path, f'ep_{ep}_{vars[i]}_err.png'))