Mercurial > pylearn

comparison denoising_aa.py @ 218:df3fae88ab46

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
small debugging
author Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
date Fri, 23 May 2008 12:22:54 -0400
parents bd728c83faff
children 9e96fe8b955c
comparison
equal deleted inserted replaced
217:44dd9b6448c5 218:df3fae88ab46
29 29
30 return HidingCorruptionFormula() 30 return HidingCorruptionFormula()
31 31
32 def squash_affine_formula(squash_function=sigmoid): 32 def squash_affine_formula(squash_function=sigmoid):
33 """ 33 """
34 Simply does: squash_function(b + xW)
34 By convention prefix the parameters by _ 35 By convention prefix the parameters by _
35 """ 36 """
36 class SquashAffineFormula(Formulas): 37 class SquashAffineFormula(Formulas):
37 x = t.matrix() # of dimensions minibatch_size x n_inputs 38 x = t.matrix() # of dimensions minibatch_size x n_inputs
38 _b = t.row() # of dimensions 1 x n_outputs 39 _b = t.row() # of dimensions 1 x n_outputs
51 52
52 def probabilistic_classifier_loss_formula(): 53 def probabilistic_classifier_loss_formula():
53 class ProbabilisticClassifierLossFormula(Formulas): 54 class ProbabilisticClassifierLossFormula(Formulas):
54 a = t.matrix() # of dimensions minibatch_size x n_classes, pre-softmax output 55 a = t.matrix() # of dimensions minibatch_size x n_classes, pre-softmax output
55 target_class = t.ivector() # dimension (minibatch_size) 56 target_class = t.ivector() # dimension (minibatch_size)
56 nll, probability_predictions = crossentropy_softmax_1hot(a, target_class) 57 nll, probability_predictions = crossentropy_softmax_1hot(a, target_class) # defined in nnet_ops.py
57 return ProbabilisticClassifierLossFormula() 58 return ProbabilisticClassifierLossFormula()
58 59
59 def binomial_cross_entropy_formula(): 60 def binomial_cross_entropy_formula():
60 class BinomialCrossEntropyFormula(Formulas): 61 class BinomialCrossEntropyFormula(Formulas):
61 a = t.matrix() # pre-sigmoid activations, minibatch_size x dim 62 a = t.matrix() # pre-sigmoid activations, minibatch_size x dim
62 p = sigmoid(a) # model prediction 63 p = sigmoid(a) # model prediction
63 q = t.matrix() # target binomial probabilities, minibatch_size x dim 64 q = t.matrix() # target binomial probabilities, minibatch_size x dim
64 # using the identity softplus(a) - softplus(-a) = a, 65 # using the identity softplus(a) - softplus(-a) = a,
65 # we obtain that q log(p) + (1-q) log(1-p) = q a - softplus(a) 66 # we obtain that q log(p) + (1-q) log(1-p) = q a - softplus(a)
66 nll = -t.sum(q*a - softplus(-a)) 67 nll = -t.sum(q*a - softplus(-a))
68 # next line was missing... hope it's all correct above
69 return BinomialCrossEntropyFormula()
67 70
68 def squash_affine_autoencoder_formula(hidden_squash=t.tanh, 71 def squash_affine_autoencoder_formula(hidden_squash=t.tanh,
69 reconstruction_squash=sigmoid, 72 reconstruction_squash=sigmoid,
70 share_weights=True, 73 share_weights=True,
71 reconstruction_nll_formula=binomial_cross_entropy_formula(), 74 reconstruction_nll_formula=binomial_cross_entropy_formula(),
100 for name,val in locals().items(): 103 for name,val in locals().items():
101 if val is not self: self.__setattribute__(name,val) 104 if val is not self: self.__setattribute__(name,val)
102 self.denoising_autoencoder_formula = corruption_formula + autoencoder.rename(x='corrupted_x') 105 self.denoising_autoencoder_formula = corruption_formula + autoencoder.rename(x='corrupted_x')
103 106
104 def __call__(self, training_set=None): 107 def __call__(self, training_set=None):
108 """ Allocate and optionnaly train a model"""
105 model = DenoisingAutoEncoderModel(self) 109 model = DenoisingAutoEncoderModel(self)
106 if training_set: 110 if training_set:
107 print 'what do I do if training set????' 111 print 'DenoisingAutoEncoder(): what do I do if training_set????'
112 # copied from mlp_factory_approach:
113 if len(trainset) == sys.maxint:
114 raise NotImplementedError('Learning from infinite streams is not supported')
115 nval = int(self.validation_portion * len(trainset))
116 nmin = len(trainset) - nval
117 assert nmin >= 0
118 minset = trainset[:nmin] #real training set for minimizing loss
119 valset = trainset[nmin:] #validation set for early stopping
120 best = model
121 for stp in self.early_stopper():
122 model.update(
123 minset.minibatches([input, target], minibatch_size=min(32,
124 len(trainset))))
125 #print 'mlp.__call__(), we did an update'
126 if stp.set_score:
127 stp.score = model(valset, ['loss_01'])
128 if (stp.score < stp.best_score):
129 best = copy.copy(model)
130 model = best
131 # end of the copy from mlp_factory_approach
132
133 return model
134
108 135
109 def compile(self, inputs, outputs): 136 def compile(self, inputs, outputs):
110 return theano.function(inputs,outputs,unpack_single=False,linker=self.linker) 137 return theano.function(inputs,outputs,unpack_single=False,linker=self.linker)
111 138
112 class DenoisingAutoEncoderModel(LearnerModel): 139 class DenoisingAutoEncoderModel(LearnerModel):