Mercurial > pylearn

comparison mlp.py @ 134:3f4e5c9bdc5e

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
author Yoshua Bengio <bengioy@iro.umontreal.ca>
date Fri, 09 May 2008 17:38:57 -0400
parents b4657441dd65
children ae5651a3696b
comparison
equal deleted inserted replaced
133:b4657441dd65 134:3f4e5c9bdc5e
66 - 'parameters' = [b1, W1, b2, W2] 66 - 'parameters' = [b1, W1, b2, W2]
67 - 'regularization_term' 67 - 'regularization_term'
68 68
69 """ 69 """
70 70
71 def __init__(self,n_hidden,n_classes,learning_rate,max_n_epochs,init_range=1.,n_inputs=None,minibatch_size=None): 71 def __init__(self,n_hidden,n_classes,learning_rate,max_n_epochs,L2_regularizer=0,init_range=1.,n_inputs=None,minibatch_size=None):
72 self._n_inputs = n_inputs 72 self._n_inputs = n_inputs
73 self._n_outputs = n_classes 73 self._n_outputs = n_classes
74 self._n_hidden = n_hidden 74 self._n_hidden = n_hidden
75 self._init_range = init_range 75 self._init_range = init_range
76 self._max_n_epochs = max_n_epochs 76 self._max_n_epochs = max_n_epochs
77 self._minibatch_size = minibatch_size 77 self._minibatch_size = minibatch_size
78 self.learning_rate = learning_rate # this is the float 78 self.learning_rate = learning_rate # this is the float
79 self.L2_regularizer = L2_regularizer
79 self._learning_rate = t.scalar('learning_rate') # this is the symbol 80 self._learning_rate = t.scalar('learning_rate') # this is the symbol
80 self._input = t.matrix('input') # n_examples x n_inputs 81 self._input = t.matrix('input') # n_examples x n_inputs
81 self._target = t.ivector('target') # n_examples x n_outputs 82 self._target = t.imatrix('target') # n_examples x 1
83 self._target_vector = self._target[:,0]
82 self._L2_regularizer = t.scalar('L2_regularizer') 84 self._L2_regularizer = t.scalar('L2_regularizer')
83 self._W1 = t.matrix('W1') 85 self._W1 = t.matrix('W1')
84 self._W2 = t.matrix('W2') 86 self._W2 = t.matrix('W2')
85 self._b1 = t.row('b1') 87 self._b1 = t.row('b1')
86 self._b2 = t.row('b2') 88 self._b2 = t.row('b2')
87 self._regularization_term = self._L2_regularizer * (t.sum(self._W1*self._W1) + t.sum(self._W2*self._W2)) 89 self._regularization_term = self._L2_regularizer * (t.sum(self._W1*self._W1) + t.sum(self._W2*self._W2))
88 self._output_activations =self._b2+t.dot(t.tanh(self._b1+t.dot(self._input,self._W1.T)),self._W2.T) 90 self._output_activations =self._b2+t.dot(t.tanh(self._b1+t.dot(self._input,self._W1.T)),self._W2.T)
89 self._nll,self._output = crossentropy_softmax_1hot(self._output_activations,self._target) 91 self._nll,self._output = crossentropy_softmax_1hot(self._output_activations,self._target_vector)
90 self._output_class = t.argmax(self._output,1) 92 self._output_class, self._max_output = t.argmax(self._output,1)
91 self._class_error = self._output_class != self._target 93 self._class_error = t.neq(self._output_class,self._target_vector)
92 self._minibatch_criterion = self._nll + self._regularization_term / t.shape(self._input)[0] 94 self._minibatch_criterion = self._nll + self._regularization_term / t.shape(self._input)[0]
93 OnlineGradientTLearner.__init__(self) 95 OnlineGradientTLearner.__init__(self)
94 96
95 def attributeNames(self): 97 def attributeNames(self):
96 return ["parameters","b1","W2","b2","W2", "L2_regularizer","regularization_term"] 98 return ["parameters","b1","W2","b2","W2", "L2_regularizer","regularization_term"]
97 99
98 def parameterAttributes(self): 100 def parameterAttributes(self):
99 return ["b1","W1", "b2", "W2"] 101 return ["b1","W1", "b2", "W2"]
100 102
101 def useInputAttributes(self):
102 return self.parameterAttributes()
103
104 def useOutputAttributes(self):
105 return []
106
107 def updateInputAttributes(self):
108 return self.parameterAttributes() + ["L2_regularizer"]
109
110 def updateMinibatchInputFields(self): 103 def updateMinibatchInputFields(self):
111 return ["input","target"] 104 return ["input","target"]
112 105
113 def updateEndOutputAttributes(self): 106 def updateEndOutputAttributes(self):
114 return ["regularization_term"] 107 return ["regularization_term"]
124 117
125 def allocate(self,minibatch): 118 def allocate(self,minibatch):
126 minibatch_n_inputs = minibatch["input"].shape[1] 119 minibatch_n_inputs = minibatch["input"].shape[1]
127 if not self._n_inputs: 120 if not self._n_inputs:
128 self._n_inputs = minibatch_n_inputs 121 self._n_inputs = minibatch_n_inputs
129 self.b1 = numpy.zeros(self._n_hidden) 122 self.b1 = numpy.zeros((1,self._n_hidden))
130 self.b2 = numpy.zeros(self._n_outputs) 123 self.b2 = numpy.zeros((1,self._n_outputs))
131 self.forget() 124 self.forget()
132 elif self._n_inputs!=minibatch_n_inputs: 125 elif self._n_inputs!=minibatch_n_inputs:
133 # if the input changes dimension on the fly, we resize and forget everything 126 # if the input changes dimension on the fly, we resize and forget everything
134 self.forget() 127 self.forget()
135 128