Mercurial > pylearn

annotate denoising_aa.py @ 295:7380376816e5

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
started a test for datasets where one field has a variable length. Not obvious, all tests requires a matrix as a reference
author Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
date Fri, 06 Jun 2008 17:11:25 -0400
parents 9e96fe8b955c
children eded3cb54930
rev   line source
210
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
1 """
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
2 A denoising auto-encoder
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
3 """
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
4
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
5 import theano
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
6 from theano.formula import *
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
7 from learner import *
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
8 from theano import tensor as t
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
9 from nnet_ops import *
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
10 import math
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
11 from misc import *
233
9e96fe8b955c moved the function from misc.py that have dependency on theano in misc_theano.py
Frederic Bastien <bastienf@iro.umontreal.ca>
parents: 218
diff changeset
12 from misc_theano import *
210
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
13 from theano.tensor_random import binomial
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
14
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
15 def hiding_corruption_formula(seed,average_fraction_hidden):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
16 """
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
17 Return a formula for the corruption process, in which a random
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
18 subset of the input numbers are hidden (mapped to 0).
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
19
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
20 @param seed: seed of the random generator
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
21 @type seed: anything that numpy.random.RandomState accepts
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
22
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
23 @param average_fraction_hidden: the probability with which each
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
24 input number is hidden (set to 0).
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
25 @type average_fraction_hidden: 0 <= real number <= 1
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
26 """
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
27 class HidingCorruptionFormula(Formulas):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
28 x = t.matrix()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
29 corrupted_x = x * binomial(seed,x,1,fraction_sampled)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
30
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
31 return HidingCorruptionFormula()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
32
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
33 def squash_affine_formula(squash_function=sigmoid):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
34 """
218
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
35 Simply does: squash_function(b + xW)
210
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
36 By convention prefix the parameters by _
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
37 """
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
38 class SquashAffineFormula(Formulas):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
39 x = t.matrix() # of dimensions minibatch_size x n_inputs
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
40 _b = t.row() # of dimensions 1 x n_outputs
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
41 _W = t.matrix() # of dimensions n_inputs x n_outputs
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
42 a = _b + t.dot(x,_W) # of dimensions minibatch_size x n_outputs
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
43 y = squash_function(a)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
44 return SquashAffineFormula()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
45
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
46 def gradient_descent_update_formula():
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
47 class GradientDescentUpdateFormula(Formula):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
48 param = t.matrix()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
49 learning_rate = t.scalar()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
50 cost = t.column() # cost of each example in a minibatch
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
51 param_update = t.add_inplace(param, -learning_rate*t.sgrad(cost))
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
52 return gradient_descent_update_formula()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
53
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
54 def probabilistic_classifier_loss_formula():
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
55 class ProbabilisticClassifierLossFormula(Formulas):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
56 a = t.matrix() # of dimensions minibatch_size x n_classes, pre-softmax output
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
57 target_class = t.ivector() # dimension (minibatch_size)
218
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
58 nll, probability_predictions = crossentropy_softmax_1hot(a, target_class) # defined in nnet_ops.py
210
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
59 return ProbabilisticClassifierLossFormula()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
60
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
61 def binomial_cross_entropy_formula():
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
62 class BinomialCrossEntropyFormula(Formulas):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
63 a = t.matrix() # pre-sigmoid activations, minibatch_size x dim
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
64 p = sigmoid(a) # model prediction
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
65 q = t.matrix() # target binomial probabilities, minibatch_size x dim
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
66 # using the identity softplus(a) - softplus(-a) = a,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
67 # we obtain that q log(p) + (1-q) log(1-p) = q a - softplus(a)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
68 nll = -t.sum(q*a - softplus(-a))
218
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
69 # next line was missing... hope it's all correct above
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
70 return BinomialCrossEntropyFormula()
210
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
71
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
72 def squash_affine_autoencoder_formula(hidden_squash=t.tanh,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
73 reconstruction_squash=sigmoid,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
74 share_weights=True,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
75 reconstruction_nll_formula=binomial_cross_entropy_formula(),
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
76 update_formula=gradient_descent_update_formula):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
77 if share_weights:
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
78 autoencoder = squash_affine_formula(hidden_squash).rename(a='code_a') + \
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
79 squash_affine_formula(reconstruction_squash).rename(x='hidden',y='reconstruction',_b='_c') + \
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
80 reconstruction_nll_formula
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
81 else:
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
82 autoencoder = squash_affine_formula(hidden_squash).rename(a='code_a',_W='_W1') + \
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
83 squash_affine_formula(reconstruction_squash).rename(x='hidden',y='reconstruction',_b='_c',_W='_W2') + \
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
84 reconstruction_nll_formula
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
85 autoencoder = autoencoder + [update_formula().rename(cost = 'nll',
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
86 param = p)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
87 for p in autoencoder.get_all('_.*')]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
88 return autoencoder
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
89
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
90
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
91 # @todo: try other corruption formulae. The above is the default one.
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
92 # not quite used in the ICML paper... (had a fixed number of 0s).
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
93
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
94 class DenoisingAutoEncoder(LearningAlgorithm):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
95
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
96 def __init__(self,n_inputs,n_hidden_per_layer,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
97 learning_rate=0.1,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
98 max_n_epochs=100,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
99 L1_regularizer=0,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
100 init_range=1.,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
101 corruption_formula = hiding_corruption_formula(),
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
102 autoencoder = squash_affine_autoencoder_formula(),
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
103 minibatch_size=None,linker = "c|py"):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
104 for name,val in locals().items():
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
105 if val is not self: self.__setattribute__(name,val)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
106 self.denoising_autoencoder_formula = corruption_formula + autoencoder.rename(x='corrupted_x')
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
107
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
108 def __call__(self, training_set=None):
218
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
109 """ Allocate and optionnaly train a model"""
210
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
110 model = DenoisingAutoEncoderModel(self)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
111 if training_set:
218
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
112 print 'DenoisingAutoEncoder(): what do I do if training_set????'
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
113 # copied from mlp_factory_approach:
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
114 if len(trainset) == sys.maxint:
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
115 raise NotImplementedError('Learning from infinite streams is not supported')
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
116 nval = int(self.validation_portion * len(trainset))
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
117 nmin = len(trainset) - nval
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
118 assert nmin >= 0
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
119 minset = trainset[:nmin] #real training set for minimizing loss
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
120 valset = trainset[nmin:] #validation set for early stopping
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
121 best = model
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
122 for stp in self.early_stopper():
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
123 model.update(
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
124 minset.minibatches([input, target], minibatch_size=min(32,
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
125 len(trainset))))
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
126 #print 'mlp.__call__(), we did an update'
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
127 if stp.set_score:
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
128 stp.score = model(valset, ['loss_01'])
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
129 if (stp.score < stp.best_score):
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
130 best = copy.copy(model)
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
131 model = best
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
132 # end of the copy from mlp_factory_approach
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
133
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
134 return model
df3fae88ab46 small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 211
diff changeset
135
210
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
136
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
137 def compile(self, inputs, outputs):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
138 return theano.function(inputs,outputs,unpack_single=False,linker=self.linker)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
139
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
140 class DenoisingAutoEncoderModel(LearnerModel):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
141 def __init__(self,learning_algorithm,params):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
142 self.learning_algorithm=learning_algorithm
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
143 self.params=params
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
144 v = learning_algorithm.v
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
145 self.update_fn = learning_algorithm.compile(learning_algorithm.denoising_autoencoder_formula.inputs,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
146 learning_algorithm.denoising_autoencoder_formula.outputs)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
147
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
148 def update(self, training_set, train_stats_collector=None):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
149
211
bd728c83faff in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents: 210
diff changeset
150 print 'dont update you crazy frog!'
210
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
151
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
152 # old stuff
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
153
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
154 # self._learning_rate = t.scalar('learning_rate') # this is the symbol
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
155 # self.L1_regularizer = L1_regularizer
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
156 # self._L1_regularizer = t.scalar('L1_regularizer')
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
157 # self._input = t.matrix('input') # n_examples x n_inputs
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
158 # self._W = t.matrix('W')
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
159 # self._b = t.row('b')
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
160 # self._c = t.row('b')
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
161 # self._regularization_term = self._L1_regularizer * t.sum(t.abs(self._W))
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
162 # self._corrupted_input = corruption_process(self._input)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
163 # self._hidden = t.tanh(self._b + t.dot(self._input, self._W.T))
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
164 # self._reconstruction_activations =self._c+t.dot(self._hidden,self._W)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
165 # self._nll,self._output = crossentropy_softmax_1hot(Print("output_activations")(self._output_activations),self._target_vector)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
166 # self._output_class = t.argmax(self._output,1)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
167 # self._class_error = t.neq(self._output_class,self._target_vector)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
168 # self._minibatch_criterion = self._nll + self._regularization_term / t.shape(self._input)[0]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
169 # OnlineGradientTLearner.__init__(self)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
170
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
171 # def attributeNames(self):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
172 # return ["parameters","b1","W2","b2","W2", "L2_regularizer","regularization_term"]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
173
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
174 # def parameterAttributes(self):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
175 # return ["b1","W1", "b2", "W2"]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
176
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
177 # def updateMinibatchInputFields(self):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
178 # return ["input","target"]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
179
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
180 # def updateEndOutputAttributes(self):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
181 # return ["regularization_term"]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
182
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
183 # def lossAttribute(self):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
184 # return "minibatch_criterion"
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
185
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
186 # def defaultOutputFields(self, input_fields):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
187 # output_fields = ["output", "output_class",]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
188 # if "target" in input_fields:
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
189 # output_fields += ["class_error", "nll"]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
190 # return output_fields
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
191
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
192 # def allocate(self,minibatch):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
193 # minibatch_n_inputs = minibatch["input"].shape[1]
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
194 # if not self._n_inputs:
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
195 # self._n_inputs = minibatch_n_inputs
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
196 # self.b1 = numpy.zeros((1,self._n_hidden))
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
197 # self.b2 = numpy.zeros((1,self._n_outputs))
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
198 # self.forget()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
199 # elif self._n_inputs!=minibatch_n_inputs:
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
200 # # if the input changes dimension on the fly, we resize and forget everything
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
201 # self.forget()
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
202
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
203 # def forget(self):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
204 # if self._n_inputs:
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
205 # r = self._init_range/math.sqrt(self._n_inputs)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
206 # self.W1 = numpy.random.uniform(low=-r,high=r,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
207 # size=(self._n_hidden,self._n_inputs))
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
208 # r = self._init_range/math.sqrt(self._n_hidden)
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
209 # self.W2 = numpy.random.uniform(low=-r,high=r,
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
210 # size=(self._n_outputs,self._n_hidden))
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
211 # self.b1[:]=0
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
212 # self.b2[:]=0
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
213 # self._n_epochs=0
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
214
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
215 # def isLastEpoch(self):
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
216 # self._n_epochs +=1
ffd50efefb70 work in progress denoising auto-encoder
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
diff changeset
217 # return self._n_epochs>=self._max_n_epochs