-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathMSE2C.py
390 lines (311 loc) · 14.9 KB
/
MSE2C.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
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
# from layers import *
import torch
import torch.nn as nn
from MSE2C_layers import *
def weights_init(m):
if type(m) in [nn.Conv2d, nn.Linear, nn.ConvTranspose2d]:
torch.nn.init.orthogonal_(m.weight)
class LinearTransitionModel(nn.Module):
def __init__(self, latent_dim, u_dim, num_prob, num_inj):
super(LinearTransitionModel, self).__init__()
self.latent_dim = latent_dim
self.u_dim = u_dim
self.num_prob = num_prob
self.num_inj = num_inj
self.trans_encoder = create_trans_encoder(self.latent_dim + 1)
self.trans_encoder.apply(weights_init)
self.At_layer = nn.Linear(self.latent_dim, self.latent_dim * self.latent_dim)
self.At_layer.apply(weights_init)
self.Bt_layer = nn.Linear(self.latent_dim, self.latent_dim * self.u_dim)
self.Bt_layer.apply(weights_init)
self.Ct_layer = nn.Linear(self.latent_dim, (self.num_prob*2+ self.num_inj)* self.latent_dim)
self.Ct_layer.apply(weights_init)
self.Dt_layer = nn.Linear(self.latent_dim, (self.num_prob*2+ self.num_inj) * self.u_dim)
self.Dt_layer.apply(weights_init)
def forward_nsteps(self, zt, dt, U):
# print(dt.shape)
# print(zt.shape)
zt_expand = torch.cat([zt, dt], dim=-1)
hz = self.trans_encoder(zt_expand)
# print(hz.shape)
At = self.At_layer(hz).view(-1, self.latent_dim, self.latent_dim)
Bt = self.Bt_layer(hz).view(-1, self.latent_dim, self.u_dim)
Ct = self.Ct_layer(hz).view(-1, self.num_prob*2+ self.num_inj, self.latent_dim)
Dt = self.Dt_layer(hz).view(-1, self.num_prob*2+ self.num_inj, self.u_dim)
Zt_k =[]
Yt_k = []
for ut in U:
ut_dt = ut * dt
zt = torch.bmm(At, zt.unsqueeze(-1)).squeeze(-1) + torch.bmm(Bt, ut_dt.unsqueeze(-1)).squeeze(-1)
yt = torch.bmm(Ct, zt.unsqueeze(-1)).squeeze(-1) + torch.bmm(Dt, ut_dt.unsqueeze(-1)).squeeze(-1)
Zt_k.append(zt)
Yt_k.append(yt)
# print('predicted At shape', At.shape)
# return Zt_k, gershgorin_loss(At), gershgorin_loss(Bt)
return Zt_k, Yt_k
def forward(self, zt, dt, ut):
# print(dt.shape)
# print(zt.shape)
zt_expand = torch.cat([zt, dt], dim=-1)
hz = self.trans_encoder(zt_expand)
# print(hz.shape)
At = self.At_layer(hz).view(-1, self.latent_dim, self.latent_dim)
Bt = self.Bt_layer(hz).view(-1, self.latent_dim, self.u_dim)
Ct = self.Ct_layer(hz).view(-1, self.num_prob*2+ self.num_inj, self.latent_dim)
Dt = self.Dt_layer(hz).view(-1, self.num_prob*2+ self.num_inj, self.u_dim)
ut_dt = ut * dt
zt1 = torch.bmm(At, zt.unsqueeze(-1)).squeeze(-1) + torch.bmm(Bt, ut_dt.unsqueeze(-1)).squeeze(-1)
yt1 = torch.bmm(Ct, zt1.unsqueeze(-1)).squeeze(-1) + torch.bmm(Dt, ut_dt.unsqueeze(-1)).squeeze(-1)
# print('predicted latent shape', zt1.shape)
return zt1, yt1
class LinearMultiTransitionModel(nn.Module):
def __init__(self, latent_dim, u_dim, num_prob, num_inj, nsteps):
super(LinearMultiTransitionModel, self).__init__()
self.latent_dim = latent_dim
self.u_dim = u_dim
self.num_prob = num_prob
self.num_inj = num_inj
self.nsteps = nsteps
self.trans_encoder = create_trans_encoder(self.latent_dim + 1)
self.trans_encoder.apply(weights_init)
self.At_layer = nn.Linear(self.latent_dim, self.latent_dim * self.latent_dim)
self.At_layer.apply(weights_init)
self.Bt_layer = nn.Linear(self.latent_dim, self.latent_dim * self.u_dim)
self.Bt_layer.apply(weights_init)
self.Ct_layer = nn.Linear(self.latent_dim, (self.num_prob*2+ self.num_inj)* self.latent_dim)
self.Ct_layer.apply(weights_init)
self.Dt_layer = nn.Linear(self.latent_dim, (self.num_prob*2+ self.num_inj) * self.u_dim)
self.Dt_layer.apply(weights_init)
def forward_nsteps(self, zt, dt, U):
# print(dt.shape)
# print(zt.shape)
zt_expand = torch.cat([zt, dt], dim=-1)
hz = self.trans_encoder(zt_expand)
# print(hz.shape)
At = self.At_layer(hz).view(-1, self.latent_dim, self.latent_dim)
Bt = self.Bt_layer(hz).view(-1, self.latent_dim, self.u_dim)
Ct = self.Ct_layer(hz).view(-1, self.num_prob*2+ self.num_inj, self.latent_dim)
Dt = self.Dt_layer(hz).view(-1, self.num_prob*2+ self.num_inj, self.u_dim)
Zt_k =[]
Yt_k = []
for ut in U:
ut_dt = ut * dt
zt = torch.bmm(At, zt.unsqueeze(-1)).squeeze(-1) + torch.bmm(Bt, ut_dt.unsqueeze(-1)).squeeze(-1)
yt = torch.bmm(Ct, zt.unsqueeze(-1)).squeeze(-1) + torch.bmm(Dt, ut_dt.unsqueeze(-1)).squeeze(-1)
Zt_k.append(zt)
Yt_k.append(yt)
# print('predicted At shape', At.shape)
# return Zt_k, gershgorin_loss(At), gershgorin_loss(Bt)
return Zt_k, Yt_k
def forward(self, zt, dt, ut):
# print(dt.shape)
# print(zt.shape)
zt_expand = torch.cat([zt, dt], dim=-1)
hz = self.trans_encoder(zt_expand)
# print(hz.shape)
At = self.At_layer(hz).view(-1, self.latent_dim, self.latent_dim)
Bt = self.Bt_layer(hz).view(-1, self.latent_dim, self.u_dim)
Ct = self.Ct_layer(hz).view(-1, self.num_prob*2+ self.num_inj, self.latent_dim)
Dt = self.Dt_layer(hz).view(-1, self.num_prob*2+ self.num_inj, self.u_dim)
ut_dt = ut * dt
zt1 = torch.bmm(At, zt.unsqueeze(-1)).squeeze(-1) + torch.bmm(Bt, ut_dt.unsqueeze(-1)).squeeze(-1)
yt1 = torch.bmm(Ct, zt1.unsqueeze(-1)).squeeze(-1) + torch.bmm(Dt, ut_dt.unsqueeze(-1)).squeeze(-1)
# print('predicted latent shape', zt1.shape)
return zt1, yt1
class NonLinearTransitionModel(nn.Module):
def __init__(self, latent_dim, u_dim, nsteps):
super(NonLinearTransitionModel, self).__init__()
self.latent_dim = latent_dim
self.latent_dim_total = latent_dim + u_dim
self.u_dim = u_dim
self.node_encoder = create_node_encoder(self.latent_dim_total, self.u_dim)
self.steps = nsteps
# self.feature = ode
self.norm = nn.BatchNorm2d(128)
def forward(self, zt, dt, ut):
hz = self.node_encoder
# ut_dt = ut * dt
zt1 = ode_solve(zt, ut, dt, self.steps, hz)
# zt1 = torch.bmm(At, zt.unsqueeze(-1)).squeeze(-1) + torch.bmm(Bt, ut_dt.unsqueeze(-1)).squeeze(-1)
# print('predicted latent shape', zt1.shape)
return zt1
def ode_solve(z0, ut, dt, nsteps, func):
n_steps = nsteps
z = z0
for i_step in range(n_steps):
z = z + dt/n_steps * func(z, ut)
# t = t + dt
return z
def create_trans_encoder(total_input_dim):
trans_encoder = nn.Sequential(
fc_bn_relu(total_input_dim, 200),
fc_bn_relu(200, 200),
fc_bn_relu(200, total_input_dim - 1)
)
return trans_encoder
def create_nltrans_encoder(total_input_dim, u_dim):
trans_encoder = nn.Sequential(
fc_bn_relu(total_input_dim, 200),
fc_bn_relu(200, 200),
fc_bn_relu(200, total_input_dim - u_dim)
)
return trans_encoder
class create_node_encoder(nn.Module):
def __init__(self, total_input_dim, u_dim):
super(create_node_encoder, self).__init__()
# self.input_dim = input_dim
self.NNODE = create_nltrans_encoder(total_input_dim, u_dim)
def forward(self, x, u):
zt_expand = torch.cat([x, u], dim=-1)
out = self.NNODE(zt_expand)
return out
def fc_bn_relu(input_dim, output_dim):
return nn.Sequential(
nn.Linear(input_dim, output_dim),
nn.BatchNorm1d(output_dim),
nn.ReLU()
)
class Encoder(nn.Module):
def __init__(self, latent_dim, input_shape, sigma=0.0):
super(Encoder, self).__init__()
self.sigma = sigma
self.fc_layers = nn.Sequential(
conv_bn_relu(input_shape[0], 16, 3, 3, stride=2),
conv_bn_relu(16, 32, 3, 3, stride=1),
conv_bn_relu(32, 64, 3, 3, stride=2),
conv_bn_relu(64, 128, 3, 3, stride=1)
)
self.fc_layers.apply(weights_init)
self.res_layers = nn.Sequential(
ResidualConv(128, 128, 3, 3, stride=1),
ResidualConv(128, 128, 3, 3, stride=1),
ResidualConv(128, 128, 3, 3, stride=1)
)
self.res_layers.apply(weights_init)
self.flatten = nn.Flatten()
self.fc_mean = nn.Linear(128 * int(input_shape[1] / 4) * int(input_shape[2] / 4), latent_dim)
def forward(self, x):
x = self.fc_layers(x)
x = self.res_layers(x)
x = self.flatten(x)
xi_mean = self.fc_mean(x)
return xi_mean
class Decoder(nn.Module):
def __init__(self, latent_dim, input_shape):
super(Decoder, self).__init__()
self.fc_layers = nn.Sequential(
nn.Linear(latent_dim, int(input_shape[1] * input_shape[2] / 16 * 128)),
nn.ReLU()
)
self.fc_layers.apply(weights_init)
self.upsample_layers = nn.Sequential(
ResidualConv(128, 128, 3, 3),
ResidualConv(128, 128, 3, 3),
ResidualConv(128, 128, 3, 3)
)
self.upsample_layers.apply(weights_init)
self.deconv_layers = nn.Sequential(
# dconv_bn_nolinear(128, 64, 3, 3, stride=(1, 1)),
# dconv_bn_nolinear(64, 32, 3, 3, stride=(2, 2)),
# dconv_bn_nolinear(32, 16, 3, 3, stride=(1, 1)),
# dconv_bn_nolinear(16, input_shape[0], 3, 3, stride=(2, 2)),
dconv_bn_nolinear(128, 64, 3, 3, stride=(1, 1), padding =1),
dconv_bn_nolinear(64, 32, 2, 2, stride=(2, 2)),
dconv_bn_nolinear(32, 16, 3, 3, stride=(1, 1), padding =1),
dconv_bn_nolinear(16, 16, 2, 2, stride=(2, 2)),
nn.Conv2d(16, input_shape[0], kernel_size=(3, 3), padding='same')
)
self.deconv_layers.apply(weights_init)
self.input_shape=input_shape
def forward(self, z):
x = self.fc_layers(z)
# print(self.input_shape)
x = x.view(-1, 128, int(self.input_shape[1] / 4), int(self.input_shape[2] / 4))
x = self.upsample_layers(x)
y = self.deconv_layers(x)
return y
class MSE2C(nn.Module):
def __init__(self, latent_dim, u_dim, num_prob, num_inj, input_shape, perm_shape, prod_loc_shape, n_steps, sigma=0.0):
super(MSE2C, self).__init__()
self._build_model(latent_dim, u_dim, num_prob, num_inj, input_shape, sigma)
self.perm_shape = perm_shape
self.prod_loc_shape = prod_loc_shape
self.n_steps = n_steps
def _build_model(self, latent_dim, u_dim, num_prob, num_inj, input_shape, sigma):
self.encoder = Encoder(latent_dim, input_shape, sigma=sigma)
self.decoder = Decoder(latent_dim, input_shape)
self.transition = LinearTransitionModel(latent_dim, u_dim, num_prob, num_inj)
def forward(self, inputs):
# X, U, Y, dt, perm = inputs
X, U, Y, dt = inputs
X_next = X[1:]
x0 = X[0]
z0 = self.encoder(x0)
x0_rec = self.decoder(z0)
X_next_pred = []
Z_next = []
# Z_next_pred = []
# q_z = q_z0
z = z0
# Z_next_pred, g_A, g_B = self.transition.forward_nsteps(z, dt, U)
Z_next_pred, Y_next_pred = self.transition.forward_nsteps(z, dt, U)
# print(self.n_steps)
for i_step in range(len(Z_next_pred)):
z_next = Z_next_pred[i_step]
x_next_pred = self.decoder(z_next)
z_next = self.encoder(X[i_step + 1])
X_next_pred.append(x_next_pred)
Z_next.append(z_next)
# self.zt1_pred = self.transition(self.zt, self.dt, self.ut)
# self.xt1_pred = self.decoder(self.zt1_pred)
# return X_next_pred, X_next, Z_next_pred, Z_next, Y_next_pred, Y, z0, x0, x0_rec, perm
return X_next_pred, X_next, Z_next_pred, Z_next, Y_next_pred, Y, z0, x0, x0_rec
def predict(self, inputs):
# xt, ut, yt, dt, perm = inputs
xt, ut, yt, dt = inputs
# assert self.n_steps == len(U) == len(X) - 1
zt = self.encoder(xt)
zt_next, yt_next = self.transition(zt, dt, ut)
xt_next_pred = self.decoder(zt_next)
# self.zt1_pred = self.transition(self.zt, self.dt, self.ut)
# self.xt1_pred = self.decoder(self.zt1_pred)
# return X_next_pred, X_next, Z_next_pred, Z_next, z0, x0, x0_rec, g_A, g_B, perm, prod_loc
return xt_next_pred, yt_next
def predict_latent(self, zt, dt, ut):
zt_next, yt_next = self.transition(zt, dt, ut)
return zt_next, yt_next
def load_weights_from_file(self, encoder_file, decoder_file, transition_file):
self.encoder.load_state_dict(torch.load(encoder_file))
self.decoder.load_state_dict(torch.load(decoder_file))
self.transition.load_state_dict(torch.load(transition_file))
def save_weights_to_file(self, encoder_file, decoder_file, transition_file):
torch.save(self.encoder.state_dict(), encoder_file)
torch.save(self.decoder.state_dict(), decoder_file)
torch.save(self.transition.state_dict(), transition_file)
class MSE2CNODE(nn.Module):
def __init__(self, latent_dim, u_dim, input_shape, perm_shape, prod_loc_shape, ode_steps, sigma=0.0):
super(MSE2CNODE, self).__init__()
self._build_model(latent_dim, u_dim, input_shape, sigma, ode_steps)
self.perm_shape = perm_shape
self.prod_loc_shape = prod_loc_shape
self.steps = ode_steps
def _build_model(self, latent_dim, u_dim, input_shape, sigma, nsteps):
self.encoder = Encoder(latent_dim, input_shape, sigma=sigma)
self.decoder = Decoder(latent_dim, input_shape)
self.transition = NonLinearTransitionModel(latent_dim, u_dim, nsteps)
def forward(self, inputs):
self.xt, self.ut, self.dt, self.perm, self.prod_loc = inputs
# nsteps = self.steps
self.zt = self.encoder(self.xt)
self.xt_rec = self.decoder(self.zt)
self.zt1_pred = self.transition(self.zt, self.dt, self.ut)
self.xt1_pred = self.decoder(self.zt1_pred)
return self.xt1_pred, self.zt1_pred, self.zt, self.xt_rec, self.perm, self.prod_loc
def load_weights_from_file(self, encoder_file, decoder_file, transition_file):
self.encoder.load_state_dict(torch.load(encoder_file))
self.decoder.load_state_dict(torch.load(decoder_file))
self.transition.load_state_dict(torch.load(transition_file))
def save_weights_to_file(self, encoder_file, decoder_file, transition_file):
torch.save(self.encoder.state_dict(), encoder_file)
torch.save(self.decoder.state_dict(), decoder_file)
torch.save(self.transition.state_dict(), transition_file)