-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathmodels.py
251 lines (204 loc) · 10.1 KB
/
models.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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
import einops
import numpy as np
import pytorch_lightning as pl
import torch
import torch.nn.functional as F
import torch.optim as optim
from omegaconf import OmegaConf
from scipy import stats
import solver as NN_4DVar
from metrics import save_netcdf, nrmse_scores, mse_scores, plot_nrmse, plot_mse, plot_snr, plot_maps, animate_maps, plot_ensemble, maps_score
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
class BiLinUnit(torch.nn.Module):
def __init__(self, dim_in, dim_out, dim, dw, dw2, dropout=0.):
super(BiLinUnit, self).__init__()
self.conv1 = torch.nn.Conv2d(dim_in, 2 * dim, (2 * dw + 1, 2 * dw + 1), padding=dw, bias=False)
self.conv2 = torch.nn.Conv2d(2 * dim, dim, (2 * dw2 + 1, 2 * dw2 + 1), padding=dw2, bias=False)
self.conv3 = torch.nn.Conv2d(2 * dim, dim_out, (2 * dw2 + 1, 2 * dw2 + 1), padding=dw2, bias=False)
self.bilin0 = torch.nn.Conv2d(dim, dim, (2 * dw2 + 1, 2 * dw2 + 1), padding=dw2, bias=False)
self.bilin1 = torch.nn.Conv2d(dim, dim, (2 * dw2 + 1, 2 * dw2 + 1), padding=dw2, bias=False)
self.bilin2 = torch.nn.Conv2d(dim, dim, (2 * dw2 + 1, 2 * dw2 + 1), padding=dw2, bias=False)
self.dropout = torch.nn.Dropout(dropout)
def forward(self, xin):
x = self.conv1(xin)
x = self.dropout(x)
x = self.conv2(F.relu(x))
x = self.dropout(x)
x = torch.cat((self.bilin0(x), self.bilin1(x) * self.bilin2(x)), dim=1)
x = self.dropout(x)
x = self.conv3(x)
return x
class Encoder(torch.nn.Module):
def __init__(self, dim_inp, dim_out, dim_ae, dw, dw2, ss, nb_blocks, rateDropout=0.):
super(Encoder, self).__init__()
self.nb_blocks = nb_blocks
self.dim_ae = dim_ae
# self.conv1HR = torch.nn.Conv2d(dim_inp,self.dim_ae,(2*dw+1,2*dw+1),padding=dw,bias=False)
# self.conv1LR = torch.nn.Conv2d(dim_inp,self.dim_ae,(2*dw+1,2*dw+1),padding=dw,bias=False)
self.pool1 = torch.nn.AvgPool2d(ss)
print(dim_inp, dim_out, dim_ae, dw, dw2, ss, nb_blocks, rateDropout)
self.conv_tr = torch.nn.ConvTranspose2d(dim_out, dim_out, (ss, ss), stride=(ss, ss), bias=False)
# self.nn_tlr = self.__make_ResNet(self.dim_ae,self.nb_blocks,rateDropout)
# self.nn_hr = self.__make_ResNet(self.dim_ae,self.nb_blocks,rateDropout)
self.nn_lr = self.__make_BilinNN(dim_inp, dim_out, self.dim_ae, dw, dw2, self.nb_blocks, rateDropout)
self.nn_hr = self.__make_BilinNN(dim_inp, dim_out, self.dim_ae, dw, dw2, self.nb_blocks, rateDropout)
self.dropout = torch.nn.Dropout(rateDropout)
def __make_BilinNN(self, dim_inp, dim_out, dim_ae, dw, dw2, nb_blocks=2, dropout=0.):
layers = []
layers.append(BiLinUnit(dim_inp, dim_out, dim_ae, dw, dw2, dropout))
for kk in range(0, nb_blocks - 1):
layers.append(BiLinUnit(dim_ae, dim_out, dim_ae, dw, dw2, dropout))
return torch.nn.Sequential(*layers)
def forward(self, xinp):
## LR component
x_lr = self.nn_lr(self.pool1(xinp))
x_lr = self.dropout(x_lr)
x_lr = self.conv_tr(x_lr)
# HR component
x_hr = self.nn_hr(xinp)
return x_lr + x_hr
class Encoder_OI(torch.nn.Module):
def __init__(self, dim_inp, dim_out, dim_ae, dw, dw2, ss, nb_blocks, rateDropout=0.):
super().__init__()
self.nb_blocks = nb_blocks
self.dim_ae = dim_ae
self.nn = self.__make_BilinNN(dim_inp, dim_out, self.dim_ae, dw, dw2, self.nb_blocks, rateDropout)
self.dropout = torch.nn.Dropout(rateDropout)
def __make_BilinNN(self, dim_inp, dim_out, dim_ae, dw, dw2, nb_blocks=2, dropout=0.):
layers = []
layers.append(BiLinUnit(dim_inp, dim_out, dim_ae, dw, dw2, dropout))
for kk in range(0, nb_blocks - 1):
layers.append(BiLinUnit(dim_ae, dim_out, dim_ae, dw, dw2, dropout))
return torch.nn.Sequential(*layers)
def forward(self, xinp):
# HR component
x = self.nn(xinp)
return x
class Decoder(torch.nn.Module):
def __init__(self):
super(Decoder, self).__init__()
def forward(self, x):
return torch.mul(1., x)
class CorrelateNoise(torch.nn.Module):
def __init__(self, shape_data, dim_cn):
super(CorrelateNoise, self).__init__()
self.conv1 = torch.nn.Conv2d(shape_data, dim_cn, (3, 3), padding=1, bias=False)
self.conv2 = torch.nn.Conv2d(dim_cn, 2 * dim_cn, (3, 3), padding=1, bias=False)
self.conv3 = torch.nn.Conv2d(2 * dim_cn, shape_data, (3, 3), padding=1, bias=False)
def forward(self, w):
w = self.conv1(F.relu(w)).to(device)
w = self.conv2(F.relu(w)).to(device)
w = self.conv3(w).to(device)
return w
class RegularizeVariance(torch.nn.Module):
def __init__(self, shape_data, dim_rv):
super(RegularizeVariance, self).__init__()
self.conv1 = torch.nn.Conv2d(shape_data, dim_rv, (3, 3), padding=1, bias=False)
self.conv2 = torch.nn.Conv2d(dim_rv, 2 * dim_rv, (3, 3), padding=1, bias=False)
self.conv3 = torch.nn.Conv2d(2 * dim_rv, shape_data, (3, 3), padding=1, bias=False)
def forward(self, v):
v = self.conv1(F.relu(v)).to(device)
v = self.conv2(F.relu(v)).to(device)
v = self.conv3(v).to(device)
return v
class Phi_r(torch.nn.Module):
def __init__(self, shape_data, DimAE, dw, dw2, ss, nb_blocks, rateDr, stochastic=False):
super().__init__()
print(shape_data, DimAE, dw, dw2, ss, nb_blocks, rateDr, stochastic)
self.stochastic = stochastic
self.encoder = Encoder(shape_data, shape_data, DimAE, dw, dw2, ss, nb_blocks, rateDr)
#self.encoder = Encoder(shape_data, 2*shape_data, DimAE, dw, dw2, ss, nb_blocks, rateDr)
self.decoder = Decoder()
self.correlate_noise = CorrelateNoise(shape_data, 10)
self.regularize_variance = RegularizeVariance(shape_data, 10)
def forward(self, x):
white = True
if self.stochastic == True:
# pure white noise
z = torch.randn([x.shape[0],x.shape[1],x.shape[2],x.shape[3]]).to(device)
# correlated noise with regularization of the variance
# z = torch.mul(self.regularize_variance(x),self.correlate_noise(z))
z = z/torch.std(x)
x = self.encoder(x+z)
else:
x = self.encoder(x)
x = self.decoder(x)
return x
class Phi_r_OI(torch.nn.Module):
def __init__(self, shape_data, DimAE, dw, dw2, ss, nb_blocks, rateDr, stochastic=False):
super().__init__()
self.stochastic = stochastic
self.encoder = Encoder_OI(shape_data, shape_data, DimAE, dw, dw2, ss, nb_blocks, rateDr)
self.decoder = Decoder()
self.correlate_noise = CorrelateNoise(shape_data, 10)
self.regularize_variance = RegularizeVariance(shape_data, 10)
def forward(self, x):
white = True
if self.stochastic == True:
# pure white noise
z = torch.randn([x.shape[0],x.shape[1],x.shape[2],x.shape[3]]).to(device)
z = z/torch.std(x)
x = self.encoder(x+z)
else:
x = self.encoder(x)
x = self.decoder(x)
return x
class Model_H(torch.nn.Module):
def __init__(self, shape_data):
super(Model_H, self).__init__()
self.dim_obs = 1
self.dim_obs_channel = np.array([shape_data])
def forward(self, x, y, mask):
dyout = (x - y) * mask
return dyout
class Model_HwithSST(torch.nn.Module):
def __init__(self, shape_data, dT=5, dim=5):
super(Model_HwithSST, self).__init__()
self.dim_obs = 2
self.dim_obs_channel = np.array([shape_data, dim])
self.conv11 = torch.nn.Conv2d(shape_data, self.dim_obs_channel[1], (3, 3), padding=1, bias=False)
self.conv21 = torch.nn.Conv2d(dT, self.dim_obs_channel[1], (3, 3), padding=1, bias=False)
self.conv_m = torch.nn.Conv2d(dT, self.dim_obs_channel[1], (3, 3), padding=1, bias=False)
self.sigmoid = torch.nn.Sigmoid() # torch.nn.Softmax(dim=1)
def forward(self, x, y, mask):
dyout = (x - y[0]) * mask[0]
y1 = y[1] * mask[1]
dyout1 = self.conv11(x) - self.conv21(y1)
dyout1 = dyout1 * self.sigmoid(self.conv_m(mask[1]))
return [dyout, dyout1]
class Gradient_img(torch.nn.Module):
def __init__(self):
super(Gradient_img, self).__init__()
a = np.array([[1., 0., -1.], [2., 0., -2.], [1., 0., -1.]])
self.conv_gx = torch.nn.Conv2d(1, 1, kernel_size=3, stride=1, padding=1, bias=False)
self.conv_gx.weight = torch.nn.Parameter(torch.from_numpy(a).float().unsqueeze(0).unsqueeze(0),
requires_grad=False)
b = np.array([[1., 2., 1.], [0., 0., 0.], [-1., -2., -1.]])
self.conv_gy = torch.nn.Conv2d(1, 1, kernel_size=3, stride=1, padding=1, bias=False)
self.conv_gy.weight = torch.nn.Parameter(torch.from_numpy(b).float().unsqueeze(0).unsqueeze(0),
requires_grad=False)
self.eps=10**-3
# self.eps=0.
def forward(self, im):
if im.size(1) == 1:
g_x = self.conv_gx(im)
g_y = self.conv_gy(im)
g = torch.sqrt(torch.pow(0.5 * g_x, 2) + torch.pow(0.5 * g_y, 2) + self.eps)
else:
for kk in range(0, im.size(1)):
g_x = self.conv_gx(im[:, kk, :, :].view(-1, 1, im.size(2), im.size(3)))
g_y = self.conv_gy(im[:, kk, :, :].view(-1, 1, im.size(2), im.size(3)))
g_x = g_x.view(-1, 1, im.size(2) - 2, im.size(3) - 2)
g_y = g_y.view(-1, 1, im.size(2) - 2, im.size(3) - 2)
ng = torch.sqrt(torch.pow(0.5 * g_x, 2) + torch.pow(0.5 * g_y, 2)+ self.eps)
if kk == 0:
g = ng.view(-1, 1, im.size(2) - 2, im.size(3) - 2)
else:
g = torch.cat((g, ng.view(-1, 1, im.size(2) - 2, im.size(3) - 2)), dim=1)
return g
class ModelLR(torch.nn.Module):
def __init__(self):
super().__init__()
self.pool = torch.nn.AvgPool2d((16, 16))
def forward(self, im):
return self.pool(im)