-
Notifications
You must be signed in to change notification settings - Fork 1
/
DPFs.py
448 lines (348 loc) · 20.6 KB
/
DPFs.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
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
import torch
import torch.nn as nn
import numpy as np
from utils import *
from nf.flows import *
from nf.models import NormalizingFlowModel,NormalizingFlowModel_cond
from torch.distributions import MultivariateNormal
import os
from torch.utils.tensorboard import SummaryWriter
from plot import *
from model.models import *
import cv2
from resamplers.resamplers import resampler
from losses import *
from nf.cglow.CGlowModel import CondGlowModel
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
#6,1994715,10,311,1006,54,23,6,24,98
class DPF(nn.Module):
def __init__(self, args):
super().__init__()
self.param = args
self.NF = args.NF_dyn
self.NFcond = args.NF_cond
self.measurement = args.measurement
self.hidden_size = args.hiddensize # origin: 32
self.state_dim = 2 # 4
self.lr = args.lr
self.alpha = args.alpha
self.seq_len = args.sequence_length
self.num_particle = args.num_particles
self.batch_size = args.batchsize
self.labeledRatio = args.labeledRatio
self.spring_force = 0.1 # 0.1 #0.05 # 0.1 for one object; 0.05 for five objects
self.drag_force = 0.0075 # 0.0075
self.pos_noise = args.pos_noise # 0.1 #0.1
self.vel_noise = args.vel_noise # 2.
self.NF_lr = args.NF_lr
self.n_sequence = 2
self.build_model()
self.eps = args.epsilon
self.scaling = args.scaling
self.threshold = args.threshold
self.max_iter = args.max_iter
self.resampler = resampler(self.param)
def build_model(self):
if self.measurement=='CGLOW':
self.encoder = build_encoder_cglow(self.hidden_size)
self.decoder = build_decoder_cglow(self.hidden_size)
self.build_particle_encoder = build_particle_encoder_cglow
else:
self.encoder = build_encoder(self.hidden_size)
self.decoder = build_decoder(self.hidden_size)
self.build_particle_encoder = build_particle_encoder
self.particle_encoder = self.build_particle_encoder(self.hidden_size, self.state_dim)
self.transition_model = build_transition_model(self.state_dim)
self.motion_update=motion_update
# normalising flow dynamic initialisation
self.nf_dyn = build_conditional_nf(self.n_sequence, 2 * self.state_dim, self.state_dim, init_var=0.01)
self.cond_model = build_conditional_nf(self.n_sequence, 2 * self.state_dim + self.hidden_size, self.state_dim, init_var=0.01)
if self.measurement=='CRNVP':
self.cnf_measurement = build_conditional_nf(self.n_sequence, self.hidden_size, self.hidden_size,
init_var=0.01, prior_std=2.5)
self.measurement_model = measurement_model_cnf(self.particle_encoder, self.cnf_measurement)
elif self.measurement=='cos':
self.measurement_model = measurement_model_cosine_distance(self.particle_encoder)
elif self.measurement=='NN':
self.likelihood_est = build_likelihood(self.hidden_size, self.state_dim)
self.measurement_model = measurement_model_NN(self.particle_encoder, self.likelihood_est)
elif self.measurement=='gaussian':
self.gaussian_distribution = torch.distributions.MultivariateNormal(torch.ones(self.hidden_size).to(device),
100 * torch.eye(self.hidden_size).to(device))
self.measurement_model = measurement_model_Gaussian(self.particle_encoder, self.gaussian_distribution)
elif self.measurement=='CGLOW':
self.cglow_measurement = build_conditional_glow(self.param).to(device)
self.measurement_model = measurement_model_cglow(self.particle_encoder, self.cglow_measurement)
self.prototype_density=compute_normal_density(pos_noise=self.pos_noise, vel_noise= self.vel_noise)
self.optim = torch.optim.Adam(self.parameters(), lr=self.lr)
self.optim_scheduler = torch.optim.lr_scheduler.MultiStepLR(self.optim, milestones=[30*(1+x) for x in range(10)], gamma=1.0)
def forward(self, inputs, train=True):
(start_image, start_state, image, state, q, visible) = inputs
state = state.to(device)
start_state = start_state.to(device)
image = image.permute(0, 1, 4, 2, 3) # for pytorch, the channels are above the width & height
image = image.to(device)
# modify the dimension of hidden state
vel = state[:, :, 2:] + torch.normal(0.0, 4.0, (state[:, :, 2:]).shape).to(device)
particle_list, particle_weight_list, noise_list, likelihood_list, init_weights_log, index_list, jac_list, prior_list, obs_likelihood = self.filtering_pos(image, start_state, vel)
# mask
if train:
mask = self.get_mask() # shape: (batch_size, seq_len)
else:
mask = 1.0
loss_sup, predictions = supervised_loss(particle_list,particle_weight_list, state, mask, train)
loss_ae = autoencoder_loss(image, train, self.encoder, self.decoder)
if self.param.trainType == 'DPF':
lamda1 = 1.0
lamda2 = 0.01
lamda3 = 2.0
loss_pseud_lik = None
total_loss = lamda1 * loss_sup + lamda3 * loss_ae #
elif self.param.trainType == 'SDPF':
lamda1 = 1.0
lamda2 = 0.01
lamda3 = 2.0
# loss_pseud_lik = self.pseudolikelihood_loss(particle_weight_list, noise_list, likelihood_list, index_list)
if self.NF:
loss_pseud_lik = pseudolikelihood_loss_nf(particle_weight_list, noise_list, likelihood_list, index_list,
jac_list, prior_list, self.param.block_length)
else:
loss_pseud_lik = pseudolikelihood_loss(particle_weight_list, noise_list, likelihood_list, index_list, self.param.block_length,
self.param.pos_noise,self.param.vel_noise)
total_loss = lamda1 * loss_sup + lamda2 * loss_pseud_lik + lamda3 * loss_ae
else:
raise ValueError('Please select the training type in DPF (supervised learning) and SDPF (semi-supervised learning)')
return total_loss, loss_sup, loss_pseud_lik, loss_ae, predictions, particle_list, particle_weight_list, state, start_state, image, likelihood_list, noise_list, obs_likelihood
def filtering_pos(self, obs, start_state_vs, vel_input):
start_state = start_state_vs[:, :2]
start_vel = start_state_vs[:, 2:]
batch_size = start_state.shape[0]
initial_particles, init_weights_log=particle_initialization(start_state, self.param.width, self.num_particle, self.state_dim, init_with_true_state=self.param.init_with_true_state)
initial_particle_probs = normalize_log_probs(init_weights_log)
obs_likelihood = 0.0
particles = initial_particles
particle_probs = initial_particle_probs
vel = start_vel
for step in range(self.seq_len):
# index_p shape: (batch, num_p)
index_p = (torch.arange(self.num_particle)+self.num_particle* torch.arange(batch_size)[:, None].repeat((1, self.num_particle))).type(torch.int64).to(device)
ESS = torch.mean(1/torch.sum(particle_probs**2, dim=-1))
if ESS<0.5*self.num_particle:
particles_resampled, particle_probs_resampled, index_p = self.resampler(particles, particle_probs)
particle_probs_resampled = particle_probs_resampled.log()
else:
particles_resampled = particles
particle_probs_resampled = particle_probs.log()
particles_physical, noise = self.motion_update(particles_resampled, vel, pos_noise=self.pos_noise)
vel = vel_input[:, step, :]
particles_dynamical, jac = nf_dynamic_model(self.nf_dyn, particles_physical,particle_probs.shape, NF=self.NF)
encodings = self.encoder(obs[:, step].float()) # encodings shape: (batch, hidden_dim)
propose_particle, lki_log, prior_log, propose_log = proposal_likelihood(self.cond_model,
self.nf_dyn,
self.measurement_model,
particles_dynamical,
particles_physical,
encodings, noise, jac,
self.NF, self.NFcond,
prototype_density= self.prototype_density)
particle_probs_resampled = particle_probs_resampled + lki_log + prior_log - propose_log
particles = propose_particle
particle_probs = particle_probs_resampled
obs_likelihood += particle_probs.mean()
particle_probs = normalize_log_probs(particle_probs)+1e-12
if step ==0:
particle_list= particles[:, None, :, :]
particle_probs_list = particle_probs[:, None, :]
noise_list = noise[:, None, :, :]
likelihood_list = lki_log[:, None, :]
index_list = index_p[:, None, :]
if self.NF:
jac_list = jac[:, None, :]
prior_list = prior_log[:, None, :]
else:
jac_list = None
prior_list = None
else:
particle_list = torch.cat([particle_list, particles[:, None]], dim=1)
particle_probs_list = torch.cat([particle_probs_list, particle_probs[:, None]], dim=1)
noise_list = torch.cat([noise_list, noise[:, None]], dim=1)
likelihood_list = torch.cat([likelihood_list, lki_log[:, None]], dim=1)
index_list = torch.cat([index_list, index_p[:,None]], dim=1)
if self.NF:
jac_list = torch.cat([jac_list, jac[:,None]], dim=1)
prior_list = torch.cat([prior_list, prior_log[:, None]], dim=1)
return particle_list, particle_probs_list, noise_list, likelihood_list, init_weights_log, index_list, jac_list, prior_list, obs_likelihood
def get_mask(self):
# number of 0 and 1
N1 = int(self.batch_size*self.seq_len*self.labeledRatio)
N0 = self.batch_size*self.seq_len - N1
arr = np.array([0] * N0 + [1] * N1)
np.random.shuffle(arr)
mask = arr.reshape(self.batch_size, self.seq_len)
mask = torch.tensor(mask).to(device)
return mask
def pretrain_ae(self, train_loader, valid_loader, start_epoch=-1, epoch_num = 100, logger = None):
best_eval_loss = 1e10
best_epoch = -1
for epoch in range(start_epoch + 1, epoch_num):
# train
self.train()
total_loss = []
for batch_idx, inputs in enumerate(train_loader):
(start_image, start_state, image, state, q, visible) = inputs
image = image.permute(0, 1, 4, 2, 3) # for pytorch, the channels are in front of width*height
image = image.reshape(-1, 3, 128, 128)
img = image.to(device)
feature = self.encoder(img)
recontr_img = self.decoder(feature)
loss = F.mse_loss(recontr_img, img)
self.zero_grad()
loss.backward()
self.optim.step()
print(f"Train AE: Iter: {batch_idx}, loss: {loss.detach().cpu().numpy()}")
total_loss.append(loss.detach().cpu().numpy())
print(f"Train AE: Epoch: {epoch}, loss: {np.mean(total_loss)}")
# validation
self.eval()
total_val_loss = []
with torch.no_grad():
for batch_idx, inputs in enumerate(valid_loader):
(start_image, start_state, image, state, q, visible) = inputs
image = image.permute(0, 1, 4, 2, 3) # for pytorch, the channels are in front of width*height
(batchsize, seq_len) = image.shape[:2]
image = image.reshape(-1, 3, 128, 128)
img = image.to(device)
self.zero_grad()
feature = self.encoder(img)
recontr_img = self.decoder(feature)
loss = F.mse_loss(recontr_img, img)
print(f"Evaluation AE: Iter: {batch_idx}, loss: {loss.detach().cpu().numpy()}")
total_val_loss.append(loss.detach().cpu().numpy())
plot_obs(img.reshape(batchsize, seq_len, 3, 128, 128),
recontr_img.reshape(batchsize, seq_len, 3, 128, 128))
eval_loss_sup_mean = np.mean(total_val_loss)
logger.add_scalar('PretrainAE_loss_eval/loss', eval_loss_sup_mean, epoch)
print(f"Evaluation AE: Epoch: {epoch}, loss: {eval_loss_sup_mean}")
# save pretain ae
if eval_loss_sup_mean < best_eval_loss:
best_eval_loss = eval_loss_sup_mean
best_epoch = epoch
print('Save best validation AE-model!')
ckpt_ae = {
"model": self.state_dict(),
'optim': self.optim.state_dict(),
}
torch.save(ckpt_ae, './model/ae_pretrain.pth')
# load the pretrained dynamic model
self.load_state_dict(ckpt_ae['model'])
self.optim.load_state_dict(ckpt_ae['optim'])
def e2e_train(self, train_loader, valid_loader, start_epoch=-1, epoch_num = 100, logger = None, run_id=None):
params = self.param
best_eval_loss = 1e10
best_epoch = -1
if self.param.load_pretrainModel:
print('Load pretrained model')
# load pretrained ae model
ckpt_ae = torch.load('./model/ae_pretrain.pth')
self.model.load_state_dict(ckpt_ae['model'])
eval_loss_epoch = []
for epoch in range(start_epoch + 1, epoch_num):
# train
self.train()
total_sup_loss = []
total_ae_loss = []
for iteration, inputs in enumerate(train_loader):
loss_all, loss_sup, loss_pseud_lik, loss_ae, predictions, particle_list, particle_weight_list, state, start_state, image, likelihood_list, noise_list, obs_likelihood = self.forward(
inputs, train=True)
self.zero_grad()#self.set_zero_grad()
loss_all.backward()
self.optim.step()#self.set_optim_step()
## debug
if params.trainType == 'SDPF':
print(
f"loss_sup: {loss_sup.detach().cpu().numpy()}, loss_pseud_lik: {loss_pseud_lik.detach().cpu().numpy()}, loss_ae: {loss_ae.detach().cpu().numpy()}")
total_sup_loss.append(loss_sup.detach().cpu().numpy())
total_ae_loss.append(loss_ae.detach().cpu().numpy())
self.optim_scheduler.step()
train_loss_sup_mean = np.mean(total_sup_loss)
total_ae_loss_mean = np.mean(total_ae_loss)
logger.add_scalar('Sup_loss/loss', train_loss_sup_mean, epoch)
print(f"End-to-end loss: epoch: {epoch}, loss: {train_loss_sup_mean}, loss_ae: {total_ae_loss_mean}, obs_likelihood: {obs_likelihood}")
# evaluate
self.eval()
total_sup_eval_loss = []
with torch.no_grad():
for iteration, inputs in enumerate(valid_loader):
self.zero_grad()
loss_all, loss_sup, loss_pseud_lik, loss_ae, predictions, particle_list, particle_weight_list, state, start_state, image, likelihood_list, noise_list,obs_likelihood = self.forward(
inputs, train=False)
total_sup_eval_loss.append(loss_sup.detach().cpu().numpy())
eval_loss_sup_mean = np.mean(total_sup_eval_loss)
logger.add_scalar('Sup_loss_eval/loss', eval_loss_sup_mean, epoch)
print(f"End-to-end loss evaluation: epoch: {epoch}, loss: {eval_loss_sup_mean}, obs_likelihood: {obs_likelihood}", self.NF)
eval_loss_epoch.append(eval_loss_sup_mean)##############
np.save(os.path.join('logs', run_id, "data", 'eval_loss_epoch.npy'), eval_loss_epoch)
if eval_loss_sup_mean < best_eval_loss:
best_eval_loss = eval_loss_sup_mean
best_epoch = epoch
print('Save best validation model')
np.savez(os.path.join('logs', run_id, "data", 'eval_result_best.npz'),
particle_list=particle_list.detach().cpu().numpy(),
particle_weight_list=particle_weight_list.detach().cpu().numpy(),
likelihood_list = likelihood_list.detach().cpu().numpy(),
pred=predictions.detach().cpu().numpy(),
state=state.detach().cpu().numpy(),
loss= total_sup_eval_loss)
checkpoint_e2e = checkpoint_state(self, epoch)
torch.save(checkpoint_e2e, os.path.join('logs', run_id, "models", 'e2e_model_bestval_e2e.pth'))
def load_model(self, file_name):
ckpt_e2e = torch.load(file_name)
load_model(self, ckpt_e2e)
epoch = ckpt_e2e['epoch']
print(f'Load epcoh: {epoch}')
def train_val(self, train_loader, valid_loader, run_id):
params = self.param
epoch_num = params.num_epochs
dirs = ['result', 'model', 'checkpoint', 'logger']
flags = [os.path.isdir(dir) for dir in dirs]
for i, flag in enumerate(flags):
if not flag:
os.mkdir(dirs[i])
logger = SummaryWriter('./logger')
start_epoch = -1
if params.resume:
print('Resume training!')
self.load_model('./model/e2e_model_bestval_e2e.pth')
if params.pretrain_ae:#False
print("Pretrain autoencoder model!")
self.pretrain_ae(train_loader, valid_loader, start_epoch=start_epoch, epoch_num=300, logger=logger)
if params.e2e_train:#True
# end-to-end training
print('End-to-end training!')
self.e2e_train(train_loader, valid_loader, start_epoch=start_epoch, epoch_num=epoch_num, logger=logger, run_id = run_id)
def testing(self, test_loader, run_id, model_path='./model/e2e_model_bestval_e2e.pth'):
params = self.param
if self.param.testing:
print('Testing!')
print('Load trained model')
self.load_model(os.path.join(model_path, 'e2e_model_bestval_e2e.pth'))
for epoch in range(1):
# test
self.eval()
total_sup_eval_loss = []
with torch.no_grad():
for iteration, inputs in enumerate(test_loader):
self.zero_grad()
loss_all, loss_sup, loss_pseud_lik, loss_ae, predictions, particle_list, particle_weight_list, state, start_state, image, likelihood_list, noise_list,obs_likelihood = self.forward(
inputs, train=False)
total_sup_eval_loss.append(loss_sup.detach().cpu().numpy())
np.save(os.path.join('logs', run_id, "data", 'test_loss_epoch.npy'), total_sup_eval_loss)
print(f"End-to-end loss testing: loss: {np.mean(total_sup_eval_loss)}")
np.savez(os.path.join('logs', run_id, "data",'test_result.npz'),
particle_list= particle_list.detach().cpu().numpy(),
particle_weight_list=particle_weight_list.detach().cpu().numpy(),
likelihood_list=likelihood_list.detach().cpu().numpy(),
state=state.detach().cpu().numpy(),
pred=predictions.detach().cpu().numpy(),
images=image.detach().cpu().numpy(),
noise=noise_list.detach().cpu().numpy())