-
Notifications
You must be signed in to change notification settings - Fork 17
/
fce.py
400 lines (318 loc) · 14.6 KB
/
fce.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
### Pytorch implementation of training EBMs via FCE
#
#
import contextlib
import numpy as np
import torch
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
import torch.optim as optim
from torch import nn
from torch.autograd import Variable
from torch.utils.data import DataLoader
from data.imca import ContrastiveConditionalDataset, SimpleDataset
from data.utils import to_one_hot
torch.set_default_tensor_type('torch.cuda.FloatTensor')
class ConditionalFCE(object):
"""
train an energy based model using noise contrastive estimation
where we assume we observe data from multiple segments/classes
this is useful for nonlinear ICA and semi supervised learning !
"""
def __init__(self, data, segments, energy_MLP, flow_model, verbose=False):
self.data = data
self.segments = segments
self.contrastSegments = (np.ones(self.segments.shape) / self.segments.shape[1]).astype(np.float32)
self.energy_MLP = energy_MLP
self.ebm_norm = -5.
self.hidden_dim = self.energy_MLP.linearLast.weight.shape[0]
self.n_segments = self.segments.shape[1]
self.ebm_finalLayer = torch.tensor(np.ones((self.hidden_dim, self.n_segments)).astype(np.float32))
# self.ebm_finalLayer = torch.tensor( np.random.random(( self.hidden_dim, self.n_segments )).astype(np.float32) )
self.flow_model = flow_model # flow model, must have sample and log density capabilities
self.noise_samples = None
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.verbose = verbose
def sample_noise(self, n):
if self.device == 'cuda':
return self.flow_model.module.sample(n)[-1].detach().cpu().numpy()
else:
return self.flow_model.sample(n)[-1].detach().numpy()
def noise_logpdf(self, dat):
"""
compute log density under flow model
"""
zs, prior_logprob, log_det = self.flow_model(dat)
flow_logdensity = (prior_logprob + log_det)
return flow_logdensity
def compute_ebm_logpdf(self, dat, seg, logNorm, augment=False):
act_allLayer = torch.mm(self.energy_MLP(dat), self.ebm_finalLayer)
if augment:
# we augment the feature extractor
act_allLayer += torch.mm(self.energy_MLP(dat) * self.energy_MLP(dat),
self.ebm_finalLayer * self.ebm_finalLayer)
# now select relevant layers by multiplying by mask matrix and reducing (and adding log norm)
act_segment = (act_allLayer * seg).sum(1) + logNorm
return act_segment
def train_ebm_fce(self, epochs=500, lr=.0001, cutoff=None, augment=False, finalLayerOnly=False, useVAT=False):
"""
FCE training of EBM model
"""
if self.verbose:
print('Training energy based model using FCE' + useVAT * ' with VAT penalty')
if cutoff is None:
cutoff = 1.00 # will basically only stop with perfect classification
# sample noise data
n = self.data.shape[0]
self.noise_samples = self.sample_noise(n) # self.noise_dist.sample( n )
# define classification labels
y = np.array([0] * n + [1] * n)
# define
dat_fce = ContrastiveConditionalDataset(np.vstack((self.data, self.noise_samples)).astype(np.float32),
to_one_hot(y)[0].astype(np.float32),
np.vstack((self.segments, self.contrastSegments)), device=self.device)
fce_loader = DataLoader(dat_fce, shuffle=True, batch_size=128)
# define log normalization constant
ebm_norm = self.ebm_norm # -5.
logNorm = torch.from_numpy(np.array(ebm_norm).astype(np.float32)).float().to(
self.device) # , device=dat_fce.device, requires_grad=True )
logNorm.requires_grad_()
#
self.ebm_finalLayer.requires_grad_()
# define optimizer
if finalLayerOnly:
# only train the final layer, this is the equivalent of g(y) in
# IMCA manuscript.
optimizer = optim.Adam([self.ebm_finalLayer] + [logNorm], lr=lr)
else:
optimizer = optim.Adam(list(self.energy_MLP.parameters()) + [self.ebm_finalLayer] + [logNorm], lr=lr)
self.energy_MLP.to(self.device)
self.energy_MLP.train()
# begin optimization
loss_criterion = nn.BCELoss()
use_cuda = torch.cuda.is_available()
if use_cuda:
self.energy_MLP.cuda()
self.energy_MLP = torch.nn.DataParallel(self.energy_MLP, device_ids=range(torch.cuda.device_count()))
cudnn.benchmark = True
for e in range(epochs):
num_correct = 0
loss_val = 0
for _, (dat, label, seg) in enumerate(fce_loader):
# consider adding VAT loss
if useVAT:
vat_loss = VATLoss(xi=10.0, eps=1.0, ip=1)
lds = vat_loss(self.energy_MLP, dat)
# noise model probs:
noise_logpdf = self.noise_logpdf(dat).view(-1,
1) # torch.tensor( self.noise_dist.logpdf( dat ).astype(np.float32) ).view(-1,1)
# pass to correct device:
if use_cuda:
dat = dat.to(self.device)
seg = seg.to(self.device)
label = label.to(self.device)
# dat, seg = dat.cuda(), seg.cuda()
# get ebm log pdf
ebm_logpdf = self.compute_ebm_logpdf(dat, seg, logNorm, augment=augment).view(-1, 1)
# define logits
logits = torch.cat((ebm_logpdf - noise_logpdf, noise_logpdf - ebm_logpdf), 1)
logits.to(self.device)
# compute accuracy:
num_correct += (logits.argmax(1) == label.argmax(1)).sum().item()
# define loss
loss = loss_criterion(torch.sigmoid(logits), label)
if useVAT:
loss += 1 * lds
loss_val += loss.item()
# take gradient step
self.energy_MLP.zero_grad()
optimizer.zero_grad()
# compute gradients
loss.backward()
# update parameters
optimizer.step()
# print some statistics
if self.verbose:
print('epoch {} \tloss: {}\taccuracy: {}'.format(e, np.round(loss_val, 4),
np.round(num_correct / (2 * n), 3)))
if num_correct / (2 * n) > cutoff:
# stop training
if self.verbose:
print('epoch {}\taccuracy: {}'.format(e, np.round(num_correct / (2 * n), 3)))
print('cutoff value satisfied .. stopping training\n----------\n')
break
self.ebm_norm = logNorm.item()
def reset_noise(self):
self.noise_samples = self.sample_noise(self.noise_samples.shape[0])
def pretrain_flow_model(self, epochs=50, lr=1e-4):
"""
pertraining of flow model using MLE
"""
optimizer = optim.Adam(self.flow_model.parameters(), lr=1e-4, weight_decay=1e-5) # todo tune WD
# print("number of params: ", sum(p.numel() for p in model_flow.parameters()))
dset = SimpleDataset(self.data.astype(np.float32), device=self.device)
train_loader = DataLoader(dset, shuffle=True, batch_size=128)
# run optimization
loss_vals = []
use_cuda = torch.cuda.is_available()
if use_cuda:
self.flow_model.to(self.device)
self.flow_model = torch.nn.DataParallel(self.flow_model, device_ids=range(torch.cuda.device_count()))
# self.flow_model.to( self.device )
cudnn.benchmark = True
print("using gpus! " + str(self.device))
self.flow_model.train()
for e in range(epochs):
loss_val = 0
for _, dat in enumerate(train_loader):
if use_cuda:
dat = dat.cuda()
dat = Variable(dat)
zs, prior_logprob, log_det = self.flow_model(dat)
logprob = prior_logprob + log_det
loss = - torch.sum(logprob) # NLL
# print(loss.item())
loss_val += loss.item()
#
self.flow_model.zero_grad()
optimizer.zero_grad()
# compute gradients
loss.backward()
# update parameters
optimizer.step()
if self.verbose:
print('epoch {}/{} \tloss: {}'.format(e, epochs, loss_val))
loss_vals.append(loss_val)
def train_flow_fce(self, epochs=50, lr=1e-4, objConstant=-1.0, cutoff=None):
"""
FCE training of EBM model
"""
if self.verbose:
print('Training flow contrastive noise for FCE')
if cutoff is None:
cutoff = 0. # basically only stop for perfect misclassification
# noise data already sampled during EBM training
n = self.data.shape[0]
self.reset_noise()
# define classification labels
y = np.array([0] * n + [1] * n)
# define
dat_fce = ContrastiveConditionalDataset(np.vstack((self.data, self.noise_samples)).astype(np.float32),
to_one_hot(y)[0].astype(np.float32),
np.vstack((self.segments, self.segments)), device=self.device)
fce_loader = DataLoader(dat_fce, shuffle=True, batch_size=128)
# define optimizer
optimizer = optim.Adam(self.flow_model.parameters(), lr=lr, weight_decay=1e-5) # todo tune WD
use_cuda = torch.cuda.is_available()
self.flow_model.to(self.device)
self.flow_model.train()
# begin optimization
loss_criterion = nn.BCELoss()
for e in range(epochs):
num_correct = 0
loss_val = 0
for _, (dat, label, seg) in enumerate(fce_loader):
# pass to correct device:
if use_cuda:
dat = dat.to(self.device)
seg = seg.to(self.device)
label = label.to(self.device)
# noise model probs:
noise_logpdf = self.noise_logpdf(dat).view(-1,
1) # torch.tensor( self.noise_dist.logpdf( dat ).astype(np.float32) ).view(-1,1)
# get ebm model probs:
ebm_logpdf = self.compute_ebm_logpdf(dat, seg, self.ebm_norm).view(-1, 1)
# define logits
logits = torch.cat((ebm_logpdf - noise_logpdf, noise_logpdf - ebm_logpdf), 1)
logits *= objConstant
# compute accuracy:
num_correct += (logits.argmax(1) == label.argmax(1)).sum().item()
# define loss
loss = loss_criterion(torch.sigmoid(logits), label)
loss_mle = - torch.mean(noise_logpdf) # mle objective for training data
loss_val += (loss.item() + loss_mle.item()) # this is the jensen shannon
# take gradient step
self.flow_model.zero_grad()
optimizer.zero_grad()
# compute gradients
loss.backward()
# update parameters
optimizer.step()
# print some statistics
if self.verbose:
print('epoch {} \tloss: {}\taccuracy: {}'.format(e, np.round(loss_val, 4),
np.round(1 - num_correct / (2 * n), 3)))
if 1 - num_correct / (2 * n) < cutoff:
if self.verbose:
print('epoch {}\taccuracy: {}'.format(e, np.round(1 - num_correct / (2 * n), 3)))
print('cutoff value satisfied .. stopping training\n----------\n')
break
def unmixSamples(self, data, modelChoice):
"""
perform unmixing of samples
"""
if modelChoice == 'EBM':
# unmix using EBM:
if self.device == 'gpu':
recov = self.energy_MLP(torch.tensor(data.astype(np.float32))).detach().numpy()
else:
recov = self.energy_MLP(torch.tensor(data.astype(np.float32))).detach().cpu().numpy()
else:
# unmix using flow model
if self.device == 'cpu':
recov = self.flow_model(torch.tensor(data.astype(np.float32)))[0][-1].detach().numpy()
else:
recov = self.flow_model(torch.tensor(data.astype(np.float32)))[0][-1].detach().cpu().numpy()
return recov
### Virtual adversarial regularization loss
#
#
# this code has been shamelessly taken from:
# https://raw.githubusercontent.com/lyakaap/VAT-pytorch/master/vat.py
#
###
@contextlib.contextmanager
def _disable_tracking_bn_stats(model):
def switch_attr(m):
if hasattr(m, 'track_running_stats'):
m.track_running_stats ^= True
model.apply(switch_attr)
yield
model.apply(switch_attr)
def _l2_normalize(d):
d_reshaped = d.view(d.shape[0], -1, *(1 for _ in range(d.dim() - 2)))
d /= torch.norm(d_reshaped, dim=1, keepdim=True) + 1e-8
return d
class VATLoss(nn.Module):
def __init__(self, xi=10.0, eps=1.0, ip=1):
"""VAT loss
:param xi: hyperparameter of VAT (default: 10.0)
:param eps: hyperparameter of VAT (default: 1.0)
:param ip: iteration times of computing adv noise (default: 1)
"""
super(VATLoss, self).__init__()
self.xi = xi
self.eps = eps
self.ip = ip
def forward(self, model, x):
with torch.no_grad():
pred = F.softmax(model(x), dim=1)
# prepare random unit tensor
d = torch.rand(x.shape).sub(0.5).to(x.device)
d = _l2_normalize(d)
with _disable_tracking_bn_stats(model):
# calc adversarial direction
for _ in range(self.ip):
d.requires_grad_()
pred_hat = model(x + self.xi * d)
logp_hat = F.log_softmax(pred_hat, dim=1)
adv_distance = F.kl_div(logp_hat, pred, reduction='batchmean')
adv_distance.backward()
d = _l2_normalize(d.grad)
model.zero_grad()
# calc LDS
r_adv = d * self.eps
pred_hat = model(x + r_adv)
logp_hat = F.log_softmax(pred_hat, dim=1)
lds = F.kl_div(logp_hat, pred, reduction='batchmean')
return lds