0
|
1 """
|
|
2 This tutorial introduces logistic regression using Theano and conjugate
|
|
3 gradient descent.
|
|
4
|
|
5 Logistic regression is a probabilistic, linear classifier. It is parametrized
|
|
6 by a weight matrix :math:`W` and a bias vector :math:`b`. Classification is
|
|
7 done by projecting data points onto a set of hyperplanes, the distance to
|
|
8 which is used to determine a class membership probability.
|
|
9
|
|
10 Mathematically, this can be written as:
|
|
11
|
|
12 .. math::
|
|
13 P(Y=i|x, W,b) &= softmax_i(W x + b) \\
|
|
14 &= \frac {e^{W_i x + b_i}} {\sum_j e^{W_j x + b_j}}
|
|
15
|
|
16
|
|
17 The output of the model or prediction is then done by taking the argmax of
|
|
18 the vector whose i'th element is P(Y=i|x).
|
|
19
|
|
20 .. math::
|
|
21
|
|
22 y_{pred} = argmax_i P(Y=i|x,W,b)
|
|
23
|
|
24
|
|
25 This tutorial presents a stochastic gradient descent optimization method
|
|
26 suitable for large datasets, and a conjugate gradient optimization method
|
|
27 that is suitable for smaller datasets.
|
|
28
|
|
29
|
|
30 References:
|
|
31
|
|
32 - textbooks: "Pattern Recognition and Machine Learning" -
|
|
33 Christopher M. Bishop, section 4.3.2
|
|
34
|
|
35
|
|
36 """
|
|
37 __docformat__ = 'restructedtext en'
|
|
38
|
|
39
|
|
40 import numpy, cPickle, gzip
|
|
41
|
|
42 import time
|
|
43
|
|
44 import theano
|
|
45 import theano.tensor as T
|
|
46 import theano.tensor.nnet
|
|
47
|
|
48
|
|
49 class LogisticRegression(object):
|
|
50 """Multi-class Logistic Regression Class
|
|
51
|
|
52 The logistic regression is fully described by a weight matrix :math:`W`
|
|
53 and bias vector :math:`b`. Classification is done by projecting data
|
|
54 points onto a set of hyperplanes, the distance to which is used to
|
|
55 determine a class membership probability.
|
|
56 """
|
|
57
|
|
58
|
|
59
|
|
60
|
|
61 def __init__(self, input, n_in, n_out):
|
|
62 """ Initialize the parameters of the logistic regression
|
|
63
|
|
64 :param input: symbolic variable that describes the input of the
|
|
65 architecture ( one minibatch)
|
|
66
|
|
67 :param n_in: number of input units, the dimension of the space in
|
|
68 which the datapoint lies
|
|
69
|
|
70 :param n_out: number of output units, the dimension of the space in
|
|
71 which the target lies
|
|
72
|
|
73 """
|
|
74
|
|
75 # initialize theta = (W,b) with 0s; W gets the shape (n_in, n_out),
|
|
76 # while b is a vector of n_out elements, making theta a vector of
|
|
77 # n_in*n_out + n_out elements
|
|
78 self.theta = theano.shared( value = numpy.zeros(n_in*n_out+n_out) )
|
|
79 # W is represented by the fisr n_in*n_out elements of theta
|
|
80 self.W = self.theta[0:n_in*n_out].reshape((n_in,n_out))
|
|
81 # b is the rest (last n_out elements)
|
|
82 self.b = self.theta[n_in*n_out:n_in*n_out+n_out]
|
|
83
|
|
84
|
|
85 # compute vector of class-membership probabilities in symbolic form
|
|
86 self.p_y_given_x = T.nnet.softmax(T.dot(input, self.W)+self.b)
|
|
87
|
|
88 # compute prediction as class whose probability is maximal in
|
|
89 # symbolic form
|
|
90 self.y_pred=T.argmax(self.p_y_given_x, axis=1)
|
|
91
|
|
92
|
|
93
|
|
94
|
|
95
|
|
96 def negative_log_likelihood(self, y):
|
|
97 """Return the negative log-likelihood of the prediction of this model
|
|
98 under a given target distribution.
|
|
99
|
|
100 .. math::
|
|
101
|
|
102 \frac{1}{|\mathcal{D}|}\mathcal{L} (\theta=\{W,b\}, \mathcal{D}) =
|
|
103 \frac{1}{|\mathcal{D}|}\sum_{i=0}^{|\mathcal{D}|} \log(P(Y=y^{(i)}|x^{(i)}, W,b)) \\
|
|
104 \ell (\theta=\{W,b\}, \mathcal{D})
|
|
105
|
|
106
|
|
107 :param y: corresponds to a vector that gives for each example the
|
|
108 :correct label
|
|
109 """
|
|
110 return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]),y])
|
|
111
|
|
112
|
|
113
|
|
114
|
|
115
|
|
116 def errors(self, y):
|
|
117 """Return a float representing the number of errors in the minibatch
|
|
118 over the total number of examples of the minibatch
|
|
119 """
|
|
120
|
|
121 # check if y has same dimension of y_pred
|
|
122 if y.ndim != self.y_pred.ndim:
|
|
123 raise TypeError('y should have the same shape as self.y_pred',
|
|
124 ('y', target.type, 'y_pred', self.y_pred.type))
|
|
125 # check if y is of the correct datatype
|
|
126 if y.dtype.startswith('int'):
|
|
127 # the T.neq operator returns a vector of 0s and 1s, where 1
|
|
128 # represents a mistake in prediction
|
|
129 return T.mean(T.neq(self.y_pred, y))
|
|
130 else:
|
|
131 raise NotImplementedError()
|
|
132
|
|
133
|
|
134
|
|
135
|
|
136
|
|
137
|
|
138
|
|
139 def cg_optimization_mnist( n_iter=50 ):
|
|
140 """Demonstrate conjugate gradient optimization of a log-linear model
|
|
141
|
|
142 This is demonstrated on MNIST.
|
|
143
|
|
144 :param n_iter: number of iterations ot run the optimizer
|
|
145
|
|
146 """
|
|
147
|
|
148 # Load the dataset
|
|
149 f = gzip.open('mnist.pkl.gz','rb')
|
|
150 train_set, valid_set, test_set = cPickle.load(f)
|
|
151 f.close()
|
|
152
|
|
153 # make minibatches of size 20
|
|
154 batch_size = 20 # sized of the minibatch
|
|
155
|
|
156 # Dealing with the training set
|
|
157 # get the list of training images (x) and their labels (y)
|
|
158 (train_set_x, train_set_y) = train_set
|
|
159 # initialize the list of training minibatches with empty list
|
|
160 train_batches = []
|
|
161 for i in xrange(0, len(train_set_x), batch_size):
|
|
162 # add to the list of minibatches the minibatch starting at
|
|
163 # position i, ending at position i+batch_size
|
|
164 # a minibatch is a pair ; the first element of the pair is a list
|
|
165 # of datapoints, the second element is the list of corresponding
|
|
166 # labels
|
|
167 train_batches = train_batches + \
|
|
168 [(train_set_x[i:i+batch_size], train_set_y[i:i+batch_size])]
|
|
169
|
|
170 # Dealing with the validation set
|
|
171 (valid_set_x, valid_set_y) = valid_set
|
|
172 # initialize the list of validation minibatches
|
|
173 valid_batches = []
|
|
174 for i in xrange(0, len(valid_set_x), batch_size):
|
|
175 valid_batches = valid_batches + \
|
|
176 [(valid_set_x[i:i+batch_size], valid_set_y[i:i+batch_size])]
|
|
177
|
|
178 # Dealing with the testing set
|
|
179 (test_set_x, test_set_y) = test_set
|
|
180 # initialize the list of testing minibatches
|
|
181 test_batches = []
|
|
182 for i in xrange(0, len(test_set_x), batch_size):
|
|
183 test_batches = test_batches + \
|
|
184 [(test_set_x[i:i+batch_size], test_set_y[i:i+batch_size])]
|
|
185
|
|
186
|
|
187 ishape = (28,28) # this is the size of MNIST images
|
|
188 n_in = 28*28 # number of input units
|
|
189 n_out = 10 # number of output units
|
|
190 # allocate symbolic variables for the data
|
|
191 x = T.fmatrix() # the data is presented as rasterized images
|
|
192 y = T.lvector() # the labels are presented as 1D vector of
|
|
193 # [long int] labels
|
|
194
|
|
195
|
|
196 # construct the logistic regression class
|
|
197 classifier = LogisticRegression( \
|
|
198 input=x.reshape((batch_size,28*28)), n_in=28*28, n_out=10)
|
|
199
|
|
200 # the cost we minimize during training is the negative log likelihood of
|
|
201 # the model in symbolic format
|
|
202 cost = classifier.negative_log_likelihood(y).mean()
|
|
203
|
|
204 # compile a theano function that computes the mistakes that are made by
|
|
205 # the model on a minibatch
|
|
206 test_model = theano.function([x,y], classifier.errors(y))
|
|
207 # compile a theano function that returns the gradient of the minibatch
|
|
208 # with respect to theta
|
|
209 batch_grad = theano.function([x, y], T.grad(cost, classifier.theta))
|
|
210 # compile a thenao function that returns the cost of a minibatch
|
|
211 batch_cost = theano.function([x, y], cost)
|
|
212
|
|
213 # creates a function that computes the average cost on the training set
|
|
214 def train_fn(theta_value):
|
|
215 classifier.theta.value = theta_value
|
|
216 cost = 0.
|
|
217 for x,y in train_batches :
|
|
218 cost += batch_cost(x,y)
|
|
219 return cost / len(train_batches)
|
|
220
|
|
221 # creates a function that computes the average gradient of cost with
|
|
222 # respect to theta
|
|
223 def train_fn_grad(theta_value):
|
|
224 classifier.theta.value = theta_value
|
|
225 grad = numpy.zeros(n_in * n_out + n_out)
|
|
226 for x,y in train_batches:
|
|
227 grad += batch_grad(x,y)
|
|
228 return grad/ len(train_batches)
|
|
229
|
|
230
|
|
231
|
|
232 validation_scores = [float('inf'), 0]
|
|
233
|
|
234 # creates the validation function
|
|
235 def callback(theta_value):
|
|
236 classifier.theta.value = theta_value
|
|
237 #compute the validation loss
|
|
238 this_validation_loss = 0.
|
|
239 for x,y in valid_batches:
|
|
240 this_validation_loss += test_model(x,y)
|
|
241
|
|
242 this_validation_loss /= len(valid_batches)
|
|
243
|
|
244 print('validation error %f %%' % (this_validation_loss*100.,))
|
|
245
|
|
246 # check if it is better then best validation score got until now
|
|
247 if this_validation_loss < validation_scores[0]:
|
|
248 # if so, replace the old one, and compute the score on the
|
|
249 # testing dataset
|
|
250 validation_scores[0] = this_validation_loss
|
|
251 test_score = 0.
|
|
252 for x,y in test_batches:
|
|
253 test_score += test_model(x,y)
|
|
254 validation_scores[1] = test_score / len(test_batches)
|
|
255
|
|
256 # using scipy conjugate gradient optimizer
|
|
257 import scipy.optimize
|
|
258 print ("Optimizing using scipy.optimize.fmin_cg...")
|
|
259 start_time = time.clock()
|
|
260 best_w_b = scipy.optimize.fmin_cg(
|
|
261 f=train_fn,
|
|
262 x0=numpy.zeros((n_in+1)*n_out, dtype=x.dtype),
|
|
263 fprime=train_fn_grad,
|
|
264 callback=callback,
|
|
265 disp=0,
|
|
266 maxiter=n_iter)
|
|
267 end_time = time.clock()
|
|
268 print(('Optimization complete with best validation score of %f %%, with '
|
|
269 'test performance %f %%') %
|
|
270 (validation_scores[0]*100., validation_scores[1]*100.))
|
|
271
|
|
272 print ('The code ran for %f minutes' % ((end_time-start_time)/60.))
|
|
273
|
|
274
|
|
275
|
|
276
|
|
277
|
|
278
|
|
279
|
|
280 if __name__ == '__main__':
|
|
281 cg_optimization_mnist()
|
|
282
|