Mercurial > pylearn

comparison algorithms/daa.py @ 487:94a4c5b7293b

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
DAA code more generic: Can now choose activation function and reconstruction cost.
author Joseph Turian <turian@gmail.com>
date Tue, 28 Oct 2008 02:21:50 -0400
parents 5ccb1662f9f6
children e06666ac32d5
comparison
equal deleted inserted replaced
486:5ccb1662f9f6 487:94a4c5b7293b
2 import theano 2 import theano
3 from theano import tensor as T 3 from theano import tensor as T
4 from theano.tensor import nnet as NN 4 from theano.tensor import nnet as NN
5 import numpy as N 5 import numpy as N
6 6
7 from pylearn import cost as cost
8
7 class DenoisingAA(T.RModule): 9 class DenoisingAA(T.RModule):
8 10
9 def __init__(self, input = None, regularize = True, tie_weights = True): 11 def __init__(self, input = None, regularize = True, tie_weights = True,
12 activation_function=NN.sigmoid, reconstruction_cost_function=cost.cross_entropy):
13 """
14 @param reconstruction_cost: Should return one cost per example (row)
15 """
10 super(DenoisingAA, self).__init__() 16 super(DenoisingAA, self).__init__()
11 17
12 # MODEL CONFIGURATION 18 # MODEL CONFIGURATION
13 self.regularize = regularize 19 self.regularize = regularize
14 self.tie_weights = tie_weights 20 self.tie_weights = tie_weights
21 self.activation_function = activation_function
22 self.reconstruction_cost_function = reconstruction_cost_function
15 23
16 # ACQUIRE/MAKE INPUT 24 # ACQUIRE/MAKE INPUT
17 if not input: 25 if not input:
18 input = T.matrix('input') 26 input = T.matrix('input')
19 self.input = theano.External(input) 27 self.input = theano.External(input)
112 if seed is not None: 120 if seed is not None:
113 obj.seed(seed) 121 obj.seed(seed)
114 obj.__hide__ = ['params'] 122 obj.__hide__ = ['params']
115 123
116 def build_regularization(self): 124 def build_regularization(self):
125 """
126 @todo: Why do we need this function?
127 """
117 return T.zero() # no regularization! 128 return T.zero() # no regularization!
118 129
119 130
120 class SigmoidXEDenoisingAA(DenoisingAA): 131 class SigmoidXEDenoisingAA(DenoisingAA):
132 """
133 @todo: Merge this into the above.
134 """
121 135
122 def build_corrupted_input(self): 136 def build_corrupted_input(self):
123 self.noise_level = theano.Member(T.scalar()) 137 self.noise_level = theano.Member(T.scalar())
124 return self.random.binomial(T.shape(self.input), 1, 1 - self.noise_level) * self.input 138 return self.random.binomial(T.shape(self.input), 1, 1 - self.noise_level) * self.input
125 139
126 def hid_activation_function(self, activation): 140 def hid_activation_function(self, activation):
127 return NN.sigmoid(activation) 141 return self.activation_function(activation)
128 142
129 def out_activation_function(self, activation): 143 def out_activation_function(self, activation):
130 return NN.sigmoid(activation) 144 return self.activation_function(activation)
131 145
132 def build_reconstruction_costs(self, output): 146 def build_reconstruction_costs(self, output):
133 reconstruction_cost_matrix = -(self.input * T.log(output) + (1 - self.input) * T.log(1 - output)) 147 return self.reconstruction_cost_function(self.input, output)
134 return T.sum(reconstruction_cost_matrix, axis=1)
135 148
136 def build_regularization(self): 149 def build_regularization(self):
137 self.l2_coef = theano.Member(T.scalar()) 150 self.l2_coef = theano.Member(T.scalar())
138 if self.tie_weights: 151 if self.tie_weights:
139 return self.l2_coef * T.sum(self.w1 * self.w1) 152 return self.l2_coef * T.sum(self.w1 * self.w1)