/
tempered_dbn.py
269 lines (224 loc) · 9.26 KB
/
tempered_dbn.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
import numpy
import copy
import time
import pickle
from collections import OrderedDict
import theano
import theano.tensor as T
from theano.printing import Print
from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
from theano import function, shared
from pylearn2.training_algorithms import default
from pylearn2.utils import serial
from pylearn2.base import Block
from pylearn2.models.model import Model
from pylearn2.space import VectorSpace
from utils import cost as costmod
from utils import rbm_utils
from utils import sharedX, floatX, npy_floatX
class TemperedDBN(Model, Block):
def validate_flags(self, flags):
flags.setdefault('train_on_samples', False)
flags.setdefault('sample_data', False)
flags.setdefault('pretrain', False)
flags.setdefault('order', 'swap_sample_learn')
flags.setdefault('use_dtneg', False)
if len(flags.keys()) != 5:
raise notimplementederror('one or more flags are currently not implemented.')
def __init__(self, rbms=None, max_updates=1e6, flags={}):
Model.__init__(self)
Block.__init__(self)
self.jobman_channel = None
self.jobman_state = {}
self.validate_flags(flags)
self.register_names_to_del(['jobman_channel'])
# dump initialization parameters to object
for (k,v) in locals().iteritems():
if k!='self': setattr(self,k,v)
# validate that RBMs have the same number of units.
for (rbm1, rbm2) in zip(rbms[:-1], rbms[1:]):
assert rbm1.n_h == rbm2.n_v
assert rbm1.batch_size == rbm2.batch_size
#assert rbm1.flags['enable_centering']
#assert rbm2.flags['enable_centering']
self.rbms = rbms
self.depth = len(rbms)
self.rng = self.rbms[0].rng
# configure input-space (necessary evil)
self.input_space = VectorSpace(self.rbms[0].n_v)
self.output_space = VectorSpace(self.rbms[-1].n_h)
self.batches_seen = 0 # incremented on every batch
self.examples_seen = 0 # incremented on every training example
self.batch_size = self.rbms[0].batch_size
self.cpu_time = 0
self.init_train_sequence()
self.do_theano()
def init_parameters_from_data(self, x):
"""
Override default model initialization, using training data statistics.
"""
for i, rbm in enumerate(self.rbms):
if i == 0:
rbm.init_parameters_from_data(x)
else:
new_x = numpy.ones((1, rbm.n_v)) * 0.5
rbm.init_parameters_from_data(new_x)
def uncenter(self):
for rbm in self.rbms:
rbm.uncenter()
def recenter(self):
for rbm in self.rbms:
rbm.recenter()
def do_theano(self):
init_names = dir(self)
###### All fields you don't want to get pickled should be created below this line
self.build_swap_funcs()
self.build_inference_func(sample=True)
self.build_inference_func(sample=False)
###### All fields you don't want to get pickled should be created above this line
final_names = dir(self)
self.register_names_to_del( [ name for name in (final_names) if name not in init_names ])
def init_train_sequence(self):
self.rbms[0].flags['learn'] = True
for rbm in self.rbms[1:]:
rbm.flags['learn'] = False if self.flags['pretrain'] else True
def build_swap_funcs(self):
self.swap_funcs = []
self.swap_ratios = []
for idx, (rbm1, rbm2) in enumerate(zip(self.rbms[:-1], self.rbms[1:])):
#logp_old1 = -rbm1.free_energy_h(rbm1.neg_h)
rbm1_negh = rbm1.sample_h_given_v(rbm1.neg_v)
logp_old1 = -rbm1.free_energy_h(rbm1_negh)
logp_old2 = -rbm2.free_energy_v(rbm2.neg_v)
logp_new1 = -rbm1.free_energy_h(rbm2.neg_v)
logp_new2 = -rbm2.free_energy_v(rbm1_negh)
logr = logp_new1 + logp_new2 - logp_old1 - logp_old2
r = T.minimum(1, T.exp(logr))
swap = rbm1.theano_rng.binomial(n=1, p=r, size=(self.batch_size,), dtype=floatX)
self.swap_funcs += [theano.function([], [swap, rbm1_negh])]
self.swap_ratios += [0.]
def build_inference_func(self, sample=False):
rval = []
layer_in = self.rbms[0].input
for rbm in self.rbms:
if sample:
layer_out = rbm.sample_h_given_v(layer_in)
else:
layer_out = rbm.h_given_v(layer_in)
rval += [layer_in]
layer_in = layer_out
rval += [layer_in]
func = theano.function([self.rbms[0].input], rval)
if sample:
self.inference_sample_func = func
else:
self.inference_func = func
def do_swap(self, i, alpha=0.99):
""" Perform swaps between samples of i-th and (i+1)-th RBM """
assert i+1 < len(self.rbms)
swap, rbm1_negh = self.swap_funcs[i]()
rbm1_negv = self.rbms[i].neg_v.get_value()
rbm2_negv = self.rbms[i+1].neg_v.get_value()
rbm1_negv[swap == 1] = self.rbms[i].sample_v_given_h_func(rbm2_negv)[swap == 1]
rbm2_negv[swap == 1] = rbm1_negh[swap == 1]
self.rbms[i].neg_v.set_value(rbm1_negv)
self.rbms[i+1].neg_v.set_value(rbm2_negv)
self.swap_ratios[i] = alpha * self.swap_ratios[i] + (1.-alpha) * swap.mean()
return swap.any()
def do_swaps(self):
swapped = numpy.zeros(self.depth-1)
for i in xrange(self.depth - 1):
iswapped = False
if len(self.rbms) == 2:
# always swap for 2-layer model
iswapped = self.do_swap(i)
else:
# When using > 2 layers, we swap RBMs (i,i+1) with even i, on
# even iterations, and RBMs (i,i+1) with odd i, on odd
# iterations.
if i % 2 == 0 and self.batches_seen % 2 == 0:
# swap even layers at even iterations
iswapped = self.do_swap(i)
elif i % 2 == 1 and self.batches_seen % 2 == 1:
# swap odd layers at odd iterations
iswapped = self.do_swap(i)
swapped[i] = iswapped
return swapped
def do_sample(self):
for rbm in self.rbms:
rbm.sample_func()
def do_learn(self, x):
for rbm in self.rbms:
if rbm.flags['learn']:
rbm.batch_train_func(x)
if self.flags['use_dtneg']:
x = rbm.neg_v.get_value()
if self.flags['train_on_samples']:
x = rbm.sample_h_given_v_func(x)
else:
x = rbm.h_given_v_func(x)
def train_batch(self, dataset, batch_size):
try:
x = dataset._iterator.next()
except StopIteration:
if hasattr(dataset._iterator._subset_iterator, 'shuffe'):
dataset._iterator._subset_iterator.shuffle()
else:
dataset._iterator._subset_iterator.reset()
x = dataset._iterator.next()
if self.flags['sample_data']:
x = self.rng.random_sample(x.shape) < x
t1 = time.time()
if self.flags['order'] == 'sample_swap_learn':
self.do_sample()
if not self.flags['pretrain']:
self.do_swaps()
self.do_learn(x.astype(floatX))
elif self.flags['order'] == 'swap_sample_learn':
if not self.flags['pretrain']:
self.do_swaps()
self.do_sample()
self.do_learn(x.astype(floatX))
else:
raise ValueError('Invalid setting for flags[order]')
self.cpu_time += time.time() - t1
self.increase_timers()
return self.batches_seen < self.max_updates
def increase_timers(self):
""" Synchronize various timers across all RBMs under simulation """
self.examples_seen += self.batch_size
self.batches_seen += 1
for rbm in self.rbms:
rbm.examples_seen += self.batch_size
rbm.batches_seen += 1
rbm.cpu_time = self.cpu_time
if self.batches_seen % 100 == 0:
print 'Swap ratios:', self.swap_ratios
def get_monitoring_channels(self, x, y=None):
chans = OrderedDict()
for i, rbm in enumerate(self.rbms):
ichans = rbm.get_monitoring_channels(x, y=y)
for (k,v) in ichans.iteritems():
chans['%s.%i' % (k,i)] = v
return chans
def __getstate__(self):
# save each RBM in a separate pickle file
self.rbm_fnames = []
for i, rbm in enumerate(self.rbms):
fname = '.rbm%i_e%i.pkl' % (i, self.batches_seen)
# Always log filename of most recently saved model
# TODO: move functionality to serial.save
rbm.fname = fname
serial.save(fname, rbm)
self.rbm_fnames.append(fname)
rval = super(TemperedDBN, self).__getstate__()
rval.pop('rbms')
return rval
def __setstate__(self, d):
self.__dict__.update(d)
self.rbms = []
for i, fname in enumerate(self.rbm_fnames):
fp = open(fname)
self.rbms += [serial.load(fname)]
fp.close()
d.pop('rbm_fnames')