Mercurial > ift6266
annotate baseline/mlp/mlp_nist.py @ 629:75458692efba
make a makefile rule for the aistats2011_cameraready.pdf target. It use new generic rule behind.
| author | Frederic Bastien <nouiz@nouiz.org> |
|---|---|
| date | Thu, 17 Mar 2011 14:34:52 -0400 |
| parents | 868f82777839 |
| children |
| rev | line source |
|---|---|
| 110 | 1 """ |
| 2 This tutorial introduces the multilayer perceptron using Theano. | |
| 3 | |
| 4 A multilayer perceptron is a logistic regressor where | |
| 5 instead of feeding the input to the logistic regression you insert a | |
| 6 intermidiate layer, called the hidden layer, that has a nonlinear | |
| 7 activation function (usually tanh or sigmoid) . One can use many such | |
| 8 hidden layers making the architecture deep. The tutorial will also tackle | |
| 9 the problem of MNIST digit classification. | |
| 10 | |
| 11 .. math:: | |
| 12 | |
| 13 f(x) = G( b^{(2)} + W^{(2)}( s( b^{(1)} + W^{(1)} x))), | |
| 14 | |
| 15 References: | |
| 16 | |
| 17 - textbooks: "Pattern Recognition and Machine Learning" - | |
| 18 Christopher M. Bishop, section 5 | |
| 19 | |
| 20 TODO: recommended preprocessing, lr ranges, regularization ranges (explain | |
| 21 to do lr first, then add regularization) | |
| 22 | |
| 23 """ | |
| 24 __docformat__ = 'restructedtext en' | |
| 25 | |
|
355
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
26 import sys |
| 110 | 27 import pdb |
| 28 import numpy | |
| 29 import pylab | |
| 30 import theano | |
| 31 import theano.tensor as T | |
| 32 import time | |
| 33 import theano.tensor.nnet | |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
34 import pylearn |
|
304
1e4bf5a5b46d
added type 2 adaptive learning configurable learning weight + versionning
xaviermuller
parents:
237
diff
changeset
|
35 import theano,pylearn.version,ift6266 |
| 110 | 36 from pylearn.io import filetensor as ft |
| 322 | 37 from ift6266 import datasets |
| 110 | 38 |
| 39 data_path = '/data/lisa/data/nist/by_class/' | |
| 40 | |
| 41 class MLP(object): | |
| 42 """Multi-Layer Perceptron Class | |
| 43 | |
| 44 A multilayer perceptron is a feedforward artificial neural network model | |
| 45 that has one layer or more of hidden units and nonlinear activations. | |
| 46 Intermidiate layers usually have as activation function thanh or the | |
| 47 sigmoid function while the top layer is a softamx layer. | |
| 48 """ | |
| 49 | |
| 50 | |
| 51 | |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
52 def __init__(self, input, n_in, n_hidden, n_out,learning_rate,detection_mode): |
| 110 | 53 """Initialize the parameters for the multilayer perceptron |
| 54 | |
| 55 :param input: symbolic variable that describes the input of the | |
| 56 architecture (one minibatch) | |
| 57 | |
| 58 :param n_in: number of input units, the dimension of the space in | |
| 59 which the datapoints lie | |
| 60 | |
| 61 :param n_hidden: number of hidden units | |
| 62 | |
| 63 :param n_out: number of output units, the dimension of the space in | |
| 64 which the labels lie | |
| 65 | |
| 66 """ | |
| 67 | |
| 68 # initialize the parameters theta = (W1,b1,W2,b2) ; note that this | |
| 69 # example contains only one hidden layer, but one can have as many | |
| 70 # layers as he/she wishes, making the network deeper. The only | |
| 71 # problem making the network deep this way is during learning, | |
| 72 # backpropagation being unable to move the network from the starting | |
| 73 # point towards; this is where pre-training helps, giving a good | |
| 74 # starting point for backpropagation, but more about this in the | |
| 75 # other tutorials | |
| 76 | |
| 77 # `W1` is initialized with `W1_values` which is uniformely sampled | |
| 78 # from -6./sqrt(n_in+n_hidden) and 6./sqrt(n_in+n_hidden) | |
| 79 # the output of uniform if converted using asarray to dtype | |
| 80 # theano.config.floatX so that the code is runable on GPU | |
| 81 W1_values = numpy.asarray( numpy.random.uniform( \ | |
| 82 low = -numpy.sqrt(6./(n_in+n_hidden)), \ | |
| 83 high = numpy.sqrt(6./(n_in+n_hidden)), \ | |
| 84 size = (n_in, n_hidden)), dtype = theano.config.floatX) | |
| 85 # `W2` is initialized with `W2_values` which is uniformely sampled | |
| 86 # from -6./sqrt(n_hidden+n_out) and 6./sqrt(n_hidden+n_out) | |
| 87 # the output of uniform if converted using asarray to dtype | |
| 88 # theano.config.floatX so that the code is runable on GPU | |
| 89 W2_values = numpy.asarray( numpy.random.uniform( | |
| 90 low = -numpy.sqrt(6./(n_hidden+n_out)), \ | |
| 91 high= numpy.sqrt(6./(n_hidden+n_out)),\ | |
| 92 size= (n_hidden, n_out)), dtype = theano.config.floatX) | |
| 93 | |
| 94 self.W1 = theano.shared( value = W1_values ) | |
| 95 self.b1 = theano.shared( value = numpy.zeros((n_hidden,), | |
| 96 dtype= theano.config.floatX)) | |
| 97 self.W2 = theano.shared( value = W2_values ) | |
| 98 self.b2 = theano.shared( value = numpy.zeros((n_out,), | |
| 99 dtype= theano.config.floatX)) | |
| 100 | |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
101 #include the learning rate in the classifer so |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
102 #we can modify it on the fly when we want |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
103 lr_value=learning_rate |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
104 self.lr=theano.shared(value=lr_value) |
| 110 | 105 # symbolic expression computing the values of the hidden layer |
| 106 self.hidden = T.tanh(T.dot(input, self.W1)+ self.b1) | |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
107 |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
108 |
| 110 | 109 |
| 110 # symbolic expression computing the values of the top layer | |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
111 if(detection_mode==0): |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
112 self.p_y_given_x= T.nnet.softmax(T.dot(self.hidden, self.W2)+self.b2) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
113 else: |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
114 self.p_y_given_x= T.nnet.sigmoid(T.dot(self.hidden, self.W2)+self.b2) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
115 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
116 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
117 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
118 # self.y_out_sig= T.sigmoid(T.dot(self.hidden, self.W2)+self.b2) |
| 110 | 119 |
| 120 # compute prediction as class whose probability is maximal in | |
| 121 # symbolic form | |
| 122 self.y_pred = T.argmax( self.p_y_given_x, axis =1) | |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
123 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
124 # self.y_pred_sig = T.argmax( self.y_out_sig, axis =1) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
125 |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
126 |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
127 |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
128 |
| 110 | 129 |
| 130 # L1 norm ; one regularization option is to enforce L1 norm to | |
| 131 # be small | |
| 132 self.L1 = abs(self.W1).sum() + abs(self.W2).sum() | |
| 133 | |
| 134 # square of L2 norm ; one regularization option is to enforce | |
| 135 # square of L2 norm to be small | |
| 136 self.L2_sqr = (self.W1**2).sum() + (self.W2**2).sum() | |
| 137 | |
| 138 | |
| 139 | |
| 140 def negative_log_likelihood(self, y): | |
| 141 """Return the mean of the negative log-likelihood of the prediction | |
| 142 of this model under a given target distribution. | |
| 143 | |
| 144 .. math:: | |
| 145 | |
| 146 \frac{1}{|\mathcal{D}|}\mathcal{L} (\theta=\{W,b\}, \mathcal{D}) = | |
| 147 \frac{1}{|\mathcal{D}|}\sum_{i=0}^{|\mathcal{D}|} \log(P(Y=y^{(i)}|x^{(i)}, W,b)) \\ | |
| 148 \ell (\theta=\{W,b\}, \mathcal{D}) | |
| 149 | |
| 150 | |
| 151 :param y: corresponds to a vector that gives for each example the | |
| 152 :correct label | |
| 153 """ | |
| 154 return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]),y]) | |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
155 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
156 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
157 def cross_entropy(self, y): |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
158 return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]),y]+T.sum(T.log(1-self.p_y_given_x), axis=1)-T.log(1-self.p_y_given_x)[T.arange(y.shape[0]),y]) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
159 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
160 |
| 110 | 161 |
| 162 | |
| 163 | |
| 164 | |
| 165 def errors(self, y): | |
| 166 """Return a float representing the number of errors in the minibatch | |
| 167 over the total number of examples of the minibatch | |
| 168 """ | |
| 169 | |
| 170 # check if y has same dimension of y_pred | |
| 171 if y.ndim != self.y_pred.ndim: | |
| 172 raise TypeError('y should have the same shape as self.y_pred', | |
| 173 ('y', target.type, 'y_pred', self.y_pred.type)) | |
| 174 # check if y is of the correct datatype | |
| 175 if y.dtype.startswith('int'): | |
| 176 # the T.neq operator returns a vector of 0s and 1s, where 1 | |
| 177 # represents a mistake in prediction | |
| 178 return T.mean(T.neq(self.y_pred, y)) | |
| 179 else: | |
| 180 raise NotImplementedError() | |
| 181 | |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
182 def mlp_get_nist_error(model_name='/u/mullerx/ift6266h10_sandbox_db/xvm_final_lr1_p073/8/best_model.npy.npz', |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
183 data_set=0): |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
184 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
185 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
186 |
| 404 | 187 |
| 188 | |
| 189 | |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
190 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
191 # load the data set and create an mlp based on the dimensions of the model |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
192 model=numpy.load(model_name) |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
193 W1=model['W1'] |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
194 W2=model['W2'] |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
195 b1=model['b1'] |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
196 b2=model['b2'] |
| 404 | 197 |
| 198 total_error_count=0.0 | |
| 199 total_exemple_count=0.0 | |
| 200 | |
| 201 nb_error_count=0.0 | |
| 202 nb_exemple_count=0.0 | |
| 203 | |
| 204 char_error_count=0.0 | |
| 205 char_exemple_count=0.0 | |
| 206 | |
| 207 min_error_count=0.0 | |
| 208 min_exemple_count=0.0 | |
| 209 | |
| 210 maj_error_count=0.0 | |
| 211 maj_exemple_count=0.0 | |
| 212 | |
|
414
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
213 vtotal_error_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
214 vtotal_exemple_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
215 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
216 vnb_error_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
217 vnb_exemple_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
218 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
219 vchar_error_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
220 vchar_exemple_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
221 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
222 vmin_error_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
223 vmin_exemple_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
224 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
225 vmaj_error_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
226 vmaj_exemple_count=0.0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
227 |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
228 nbc_error_count=0.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
229 vnbc_error_count=0.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
230 |
| 404 | 231 |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
232 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
233 if data_set==0: |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
234 print 'using nist' |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
235 dataset=datasets.nist_all() |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
236 elif data_set==1: |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
237 print 'using p07' |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
238 dataset=datasets.nist_P07() |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
239 elif data_set==2: |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
240 print 'using pnist' |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
241 dataset=datasets.PNIST07() |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
242 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
243 |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
244 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
245 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
246 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
247 #get the test error |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
248 #use a batch size of 1 so we can get the sub-class error |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
249 #without messing with matrices (will be upgraded later) |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
250 test_score=0 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
251 temp=0 |
| 404 | 252 for xt,yt in dataset.test(1): |
| 253 | |
| 254 total_exemple_count = total_exemple_count +1 | |
| 255 #get activation for layer 1 | |
| 256 a0=numpy.dot(numpy.transpose(W1),numpy.transpose(xt[0])) + b1 | |
| 257 #add non linear function to layer 1 activation | |
| 258 a0_out=numpy.tanh(a0) | |
| 259 | |
| 260 #get activation for output layer | |
| 261 a1= numpy.dot(numpy.transpose(W2),a0_out) + b2 | |
| 262 #add non linear function for output activation (softmax) | |
| 263 a1_exp = numpy.exp(a1) | |
| 264 sum_a1=numpy.sum(a1_exp) | |
| 265 a1_out=a1_exp/sum_a1 | |
| 266 | |
| 267 predicted_class=numpy.argmax(a1_out) | |
| 268 wanted_class=yt[0] | |
| 269 if(predicted_class!=wanted_class): | |
| 270 total_error_count = total_error_count +1 | |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
271 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
272 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
273 if(not(predicted_class==wanted_class or ( (((predicted_class+26)==wanted_class) or ((predicted_class-26)==wanted_class)) and wanted_class>9) )): |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
274 nbc_error_count = nbc_error_count +1 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
275 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
276 |
| 404 | 277 #treat digit error |
| 278 if(wanted_class<10): | |
| 279 nb_exemple_count=nb_exemple_count + 1 | |
| 280 predicted_class=numpy.argmax(a1_out[0:10]) | |
| 281 if(predicted_class!=wanted_class): | |
| 282 nb_error_count = nb_error_count +1 | |
| 283 | |
| 284 if(wanted_class>9): | |
| 285 char_exemple_count=char_exemple_count + 1 | |
| 286 predicted_class=numpy.argmax(a1_out[10:62])+10 | |
| 287 if((predicted_class!=wanted_class) and ((predicted_class+26)!=wanted_class) and ((predicted_class-26)!=wanted_class)): | |
| 288 char_error_count = char_error_count +1 | |
| 405 | 289 |
| 290 #minuscule | |
| 291 if(wanted_class>9 and wanted_class<36): | |
| 292 maj_exemple_count=maj_exemple_count + 1 | |
| 293 predicted_class=numpy.argmax(a1_out[10:35])+10 | |
| 294 if(predicted_class!=wanted_class): | |
| 295 maj_error_count = maj_error_count +1 | |
| 296 #majuscule | |
| 297 if(wanted_class>35): | |
| 298 min_exemple_count=min_exemple_count + 1 | |
| 299 predicted_class=numpy.argmax(a1_out[36:62])+36 | |
| 300 if(predicted_class!=wanted_class): | |
| 301 min_error_count = min_error_count +1 | |
| 404 | 302 |
| 303 | |
|
414
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
304 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
305 vtest_score=0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
306 vtemp=0 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
307 for xt,yt in dataset.valid(1): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
308 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
309 vtotal_exemple_count = vtotal_exemple_count +1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
310 #get activation for layer 1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
311 a0=numpy.dot(numpy.transpose(W1),numpy.transpose(xt[0])) + b1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
312 #add non linear function to layer 1 activation |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
313 a0_out=numpy.tanh(a0) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
314 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
315 #get activation for output layer |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
316 a1= numpy.dot(numpy.transpose(W2),a0_out) + b2 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
317 #add non linear function for output activation (softmax) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
318 a1_exp = numpy.exp(a1) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
319 sum_a1=numpy.sum(a1_exp) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
320 a1_out=a1_exp/sum_a1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
321 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
322 predicted_class=numpy.argmax(a1_out) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
323 wanted_class=yt[0] |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
324 if(predicted_class!=wanted_class): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
325 vtotal_error_count = vtotal_error_count +1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
326 |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
327 if(not(predicted_class==wanted_class or ( (((predicted_class+26)==wanted_class) or ((predicted_class-26)==wanted_class)) and wanted_class>9) )): |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
328 vnbc_error_count = nbc_error_count +1 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
329 |
|
414
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
330 #treat digit error |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
331 if(wanted_class<10): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
332 vnb_exemple_count=vnb_exemple_count + 1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
333 predicted_class=numpy.argmax(a1_out[0:10]) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
334 if(predicted_class!=wanted_class): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
335 vnb_error_count = vnb_error_count +1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
336 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
337 if(wanted_class>9): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
338 vchar_exemple_count=vchar_exemple_count + 1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
339 predicted_class=numpy.argmax(a1_out[10:62])+10 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
340 if((predicted_class!=wanted_class) and ((predicted_class+26)!=wanted_class) and ((predicted_class-26)!=wanted_class)): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
341 vchar_error_count = vchar_error_count +1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
342 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
343 #minuscule |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
344 if(wanted_class>9 and wanted_class<36): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
345 vmaj_exemple_count=vmaj_exemple_count + 1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
346 predicted_class=numpy.argmax(a1_out[10:35])+10 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
347 if(predicted_class!=wanted_class): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
348 vmaj_error_count = vmaj_error_count +1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
349 #majuscule |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
350 if(wanted_class>35): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
351 vmin_exemple_count=vmin_exemple_count + 1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
352 predicted_class=numpy.argmax(a1_out[36:62])+36 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
353 if(predicted_class!=wanted_class): |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
354 vmin_error_count = vmin_error_count +1 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
355 |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
356 |
| 404 | 357 print (('total error = %f') % ((total_error_count/total_exemple_count)*100.0)) |
| 358 print (('number error = %f') % ((nb_error_count/nb_exemple_count)*100.0)) | |
| 359 print (('char error = %f') % ((char_error_count/char_exemple_count)*100.0)) | |
| 405 | 360 print (('min error = %f') % ((min_error_count/min_exemple_count)*100.0)) |
| 361 print (('maj error = %f') % ((maj_error_count/maj_exemple_count)*100.0)) | |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
362 print (('36 error = %f') % ((nbc_error_count/total_exemple_count)*100.0)) |
|
414
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
363 |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
364 print (('valid total error = %f') % ((vtotal_error_count/vtotal_exemple_count)*100.0)) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
365 print (('valid number error = %f') % ((vnb_error_count/vnb_exemple_count)*100.0)) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
366 print (('valid char error = %f') % ((vchar_error_count/vchar_exemple_count)*100.0)) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
367 print (('valid min error = %f') % ((vmin_error_count/vmin_exemple_count)*100.0)) |
|
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
368 print (('valid maj error = %f') % ((vmaj_error_count/vmaj_exemple_count)*100.0)) |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
369 print (('valid 36 error = %f') % ((vnbc_error_count/vtotal_exemple_count)*100.0)) |
|
414
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
370 |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
371 print (('num total = %d,%d') % (total_exemple_count,total_error_count)) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
372 print (('num nb = %d,%d') % (nb_exemple_count,nb_error_count)) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
373 print (('num min = %d,%d') % (min_exemple_count,min_error_count)) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
374 print (('num maj = %d,%d') % (maj_exemple_count,maj_error_count)) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
375 print (('num char = %d,%d') % (char_exemple_count,char_error_count)) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
376 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
377 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
378 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
379 total_error_count/=total_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
380 nb_error_count/=nb_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
381 char_error_count/=char_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
382 min_error_count/=min_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
383 maj_error_count/=maj_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
384 nbc_error_count/=total_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
385 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
386 vtotal_error_count/=vtotal_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
387 vnb_error_count/=vnb_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
388 vchar_error_count/=vchar_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
389 vmin_error_count/=vmin_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
390 vmaj_error_count/=vmaj_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
391 vnbc_error_count/=vtotal_exemple_count |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
392 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
393 |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
394 |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
395 return (total_error_count,nb_error_count,char_error_count,min_error_count,maj_error_count,nbc_error_count,\ |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
396 vtotal_error_count,vnb_error_count,vchar_error_count,vmin_error_count,vmaj_error_count,vnbc_error_count) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
397 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
398 def jobman_get_error(state,channel): |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
399 (all_t_error,nb_t_error,char_t_error,min_t_error,maj_t_error,nbc_t_error, |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
400 all_v_error,nb_v_error,char_v_error,min_v_error,maj_v_error,nbc_v_error)=mlp_get_nist_error(data_set=state.data_set,\ |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
401 model_name=state.model_name) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
402 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
403 state.all_t_error=all_t_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
404 state.nb_t_error=nb_t_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
405 state.char_t_error=char_t_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
406 state.min_t_error=min_t_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
407 state.maj_t_error=maj_t_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
408 state.nbc_t_error=nbc_t_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
409 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
410 state.all_v_error=all_v_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
411 state.nb_v_error=nb_v_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
412 state.char_v_error=char_v_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
413 state.min_v_error=min_v_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
414 state.maj_v_error=maj_v_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
415 state.nbc_v_error=nbc_v_error*100.0 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
416 |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
417 pylearn.version.record_versions(state,[theano,ift6266,pylearn]) |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
418 return channel.COMPLETE |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
419 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
420 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
421 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
422 |
| 110 | 423 |
|
304
1e4bf5a5b46d
added type 2 adaptive learning configurable learning weight + versionning
xaviermuller
parents:
237
diff
changeset
|
424 def mlp_full_nist( verbose = 1,\ |
|
145
8ceaaf812891
changed adaptive lr flag from bool to int for jobman issues
XavierMuller
parents:
143
diff
changeset
|
425 adaptive_lr = 0,\ |
| 322 | 426 data_set=0,\ |
| 110 | 427 learning_rate=0.01,\ |
| 428 L1_reg = 0.00,\ | |
| 429 L2_reg = 0.0001,\ | |
| 430 nb_max_exemples=1000000,\ | |
| 431 batch_size=20,\ | |
| 322 | 432 nb_hidden = 30,\ |
|
212
e390b0454515
added classic lr time decay and py code to calculate the error based on a saved model
xaviermuller
parents:
169
diff
changeset
|
433 nb_targets = 62, |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
434 tau=1e6,\ |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
435 lr_t2_factor=0.5,\ |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
436 init_model=0,\ |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
437 channel=0,\ |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
438 detection_mode=0): |
| 110 | 439 |
| 440 | |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
441 if channel!=0: |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
442 channel.save() |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
443 configuration = [learning_rate,nb_max_exemples,nb_hidden,adaptive_lr] |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
444 |
|
212
e390b0454515
added classic lr time decay and py code to calculate the error based on a saved model
xaviermuller
parents:
169
diff
changeset
|
445 #save initial learning rate if classical adaptive lr is used |
|
e390b0454515
added classic lr time decay and py code to calculate the error based on a saved model
xaviermuller
parents:
169
diff
changeset
|
446 initial_lr=learning_rate |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
447 max_div_count=1000 |
|
414
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
448 optimal_test_error=0 |
| 323 | 449 |
|
212
e390b0454515
added classic lr time decay and py code to calculate the error based on a saved model
xaviermuller
parents:
169
diff
changeset
|
450 |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
451 total_validation_error_list = [] |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
452 total_train_error_list = [] |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
453 learning_rate_list=[] |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
454 best_training_error=float('inf'); |
| 323 | 455 divergence_flag_list=[] |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
456 |
| 322 | 457 if data_set==0: |
| 378 | 458 print 'using nist' |
| 322 | 459 dataset=datasets.nist_all() |
| 323 | 460 elif data_set==1: |
| 378 | 461 print 'using p07' |
| 323 | 462 dataset=datasets.nist_P07() |
|
349
22efb4968054
added pnist support, will check in code for data set iterator later
xaviermuller
parents:
338
diff
changeset
|
463 elif data_set==2: |
| 378 | 464 print 'using pnist' |
|
349
22efb4968054
added pnist support, will check in code for data set iterator later
xaviermuller
parents:
338
diff
changeset
|
465 dataset=datasets.PNIST07() |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
466 |
|
212
e390b0454515
added classic lr time decay and py code to calculate the error based on a saved model
xaviermuller
parents:
169
diff
changeset
|
467 |
| 110 | 468 |
| 469 | |
| 470 ishape = (32,32) # this is the size of NIST images | |
| 471 | |
| 472 # allocate symbolic variables for the data | |
| 473 x = T.fmatrix() # the data is presented as rasterized images | |
| 474 y = T.lvector() # the labels are presented as 1D vector of | |
| 475 # [long int] labels | |
| 476 | |
| 322 | 477 |
| 110 | 478 # construct the logistic regression class |
| 322 | 479 classifier = MLP( input=x,\ |
| 110 | 480 n_in=32*32,\ |
| 481 n_hidden=nb_hidden,\ | |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
482 n_out=nb_targets, |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
483 learning_rate=learning_rate, |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
484 detection_mode=detection_mode) |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
485 |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
486 |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
487 # check if we want to initialise the weights with a previously calculated model |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
488 # dimensions must be consistent between old model and current configuration!!!!!! (nb_hidden and nb_targets) |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
489 if init_model!=0: |
| 378 | 490 print 'using old model' |
| 491 print init_model | |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
492 old_model=numpy.load(init_model) |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
493 classifier.W1.value=old_model['W1'] |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
494 classifier.W2.value=old_model['W2'] |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
495 classifier.b1.value=old_model['b1'] |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
496 classifier.b2.value=old_model['b2'] |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
497 |
| 110 | 498 |
| 499 # the cost we minimize during training is the negative log likelihood of | |
| 500 # the model plus the regularization terms (L1 and L2); cost is expressed | |
| 501 # here symbolically | |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
502 if(detection_mode==0): |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
503 cost = classifier.negative_log_likelihood(y) \ |
| 110 | 504 + L1_reg * classifier.L1 \ |
| 505 + L2_reg * classifier.L2_sqr | |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
506 else: |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
507 cost = classifier.cross_entropy(y) \ |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
508 + L1_reg * classifier.L1 \ |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
509 + L2_reg * classifier.L2_sqr |
|
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
510 |
| 110 | 511 |
| 512 # compiling a theano function that computes the mistakes that are made by | |
| 513 # the model on a minibatch | |
| 514 test_model = theano.function([x,y], classifier.errors(y)) | |
| 515 | |
| 516 # compute the gradient of cost with respect to theta = (W1, b1, W2, b2) | |
| 517 g_W1 = T.grad(cost, classifier.W1) | |
| 518 g_b1 = T.grad(cost, classifier.b1) | |
| 519 g_W2 = T.grad(cost, classifier.W2) | |
| 520 g_b2 = T.grad(cost, classifier.b2) | |
| 521 | |
| 522 # specify how to update the parameters of the model as a dictionary | |
| 523 updates = \ | |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
524 { classifier.W1: classifier.W1 - classifier.lr*g_W1 \ |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
525 , classifier.b1: classifier.b1 - classifier.lr*g_b1 \ |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
526 , classifier.W2: classifier.W2 - classifier.lr*g_W2 \ |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
527 , classifier.b2: classifier.b2 - classifier.lr*g_b2 } |
| 110 | 528 |
| 529 # compiling a theano function `train_model` that returns the cost, but in | |
| 530 # the same time updates the parameter of the model based on the rules | |
| 531 # defined in `updates` | |
| 532 train_model = theano.function([x, y], cost, updates = updates ) | |
| 322 | 533 |
| 534 | |
| 535 | |
| 110 | 536 |
| 537 | |
| 538 | |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
539 |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
540 |
| 110 | 541 |
| 542 #conditions for stopping the adaptation: | |
| 323 | 543 #1) we have reached nb_max_exemples (this is rounded up to be a multiple of the train size so we always do at least 1 epoch) |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
544 #2) validation error is going up twice in a row(probable overfitting) |
| 110 | 545 |
| 546 # This means we no longer stop on slow convergence as low learning rates stopped | |
| 323 | 547 # too fast but instead we will wait for the valid error going up 3 times in a row |
| 548 # We save the curb of the validation error so we can always go back to check on it | |
| 549 # and we save the absolute best model anyway, so we might as well explore | |
| 550 # a bit when diverging | |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
551 |
| 323 | 552 #approximate number of samples in the nist training set |
| 322 | 553 #this is just to have a validation frequency |
| 323 | 554 #roughly proportionnal to the original nist training set |
| 322 | 555 n_minibatches = 650000/batch_size |
| 556 | |
| 557 | |
| 323 | 558 patience =2*nb_max_exemples/batch_size #in units of minibatch |
| 110 | 559 validation_frequency = n_minibatches/4 |
| 560 | |
| 561 | |
| 562 | |
| 563 | |
| 322 | 564 |
| 110 | 565 best_validation_loss = float('inf') |
| 566 best_iter = 0 | |
| 567 test_score = 0. | |
| 568 start_time = time.clock() | |
|
212
e390b0454515
added classic lr time decay and py code to calculate the error based on a saved model
xaviermuller
parents:
169
diff
changeset
|
569 time_n=0 #in unit of exemples |
| 322 | 570 minibatch_index=0 |
| 571 epoch=0 | |
| 572 temp=0 | |
| 323 | 573 divergence_flag=0 |
| 322 | 574 |
|
212
e390b0454515
added classic lr time decay and py code to calculate the error based on a saved model
xaviermuller
parents:
169
diff
changeset
|
575 |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
576 |
|
355
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
577 |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
578 print 'starting training' |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
579 sys.stdout.flush() |
| 322 | 580 while(minibatch_index*batch_size<nb_max_exemples): |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
581 |
| 322 | 582 for x, y in dataset.train(batch_size): |
| 110 | 583 |
| 323 | 584 #if we are using the classic learning rate deacay, adjust it before training of current mini-batch |
| 322 | 585 if adaptive_lr==2: |
| 586 classifier.lr.value = tau*initial_lr/(tau+time_n) | |
| 587 | |
| 588 | |
| 589 #train model | |
| 590 cost_ij = train_model(x,y) | |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
591 if (minibatch_index) % validation_frequency == 0: |
| 322 | 592 #save the current learning rate |
| 593 learning_rate_list.append(classifier.lr.value) | |
| 323 | 594 divergence_flag_list.append(divergence_flag) |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
595 |
|
355
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
596 |
| 322 | 597 |
| 598 # compute the validation error | |
| 599 this_validation_loss = 0. | |
| 600 temp=0 | |
| 601 for xv,yv in dataset.valid(1): | |
| 602 # sum up the errors for each minibatch | |
| 323 | 603 this_validation_loss += test_model(xv,yv) |
| 322 | 604 temp=temp+1 |
| 605 # get the average by dividing with the number of minibatches | |
| 606 this_validation_loss /= temp | |
| 607 #save the validation loss | |
| 608 total_validation_error_list.append(this_validation_loss) | |
|
355
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
609 |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
610 print(('epoch %i, minibatch %i, learning rate %f current validation error %f ') % |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
611 (epoch, minibatch_index+1,classifier.lr.value, |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
612 this_validation_loss*100.)) |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
613 sys.stdout.flush() |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
614 |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
615 #save temp results to check during training |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
616 numpy.savez('temp_results.npy',config=configuration,total_validation_error_list=total_validation_error_list,\ |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
617 learning_rate_list=learning_rate_list, divergence_flag_list=divergence_flag_list) |
| 322 | 618 |
| 619 # if we got the best validation score until now | |
| 620 if this_validation_loss < best_validation_loss: | |
| 621 # save best validation score and iteration number | |
| 622 best_validation_loss = this_validation_loss | |
| 623 best_iter = minibatch_index | |
| 323 | 624 #reset divergence flag |
| 625 divergence_flag=0 | |
| 626 | |
| 627 #save the best model. Overwrite the current saved best model so | |
| 628 #we only keep the best | |
| 629 numpy.savez('best_model.npy', config=configuration, W1=classifier.W1.value, W2=classifier.W2.value, b1=classifier.b1.value,\ | |
| 630 b2=classifier.b2.value, minibatch_index=minibatch_index) | |
| 631 | |
| 322 | 632 # test it on the test set |
| 633 test_score = 0. | |
| 634 temp =0 | |
| 635 for xt,yt in dataset.test(batch_size): | |
| 636 test_score += test_model(xt,yt) | |
| 637 temp = temp+1 | |
| 638 test_score /= temp | |
|
355
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
639 |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
640 print(('epoch %i, minibatch %i, test error of best ' |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
641 'model %f %%') % |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
642 (epoch, minibatch_index+1, |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
643 test_score*100.)) |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
644 sys.stdout.flush() |
|
414
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
645 optimal_test_error=test_score |
| 322 | 646 |
| 647 # if the validation error is going up, we are overfitting (or oscillating) | |
| 323 | 648 # check if we are allowed to continue and if we will adjust the learning rate |
| 322 | 649 elif this_validation_loss >= best_validation_loss: |
| 323 | 650 |
| 651 | |
| 652 # In non-classic learning rate decay, we modify the weight only when | |
| 653 # validation error is going up | |
| 654 if adaptive_lr==1: | |
| 655 classifier.lr.value=classifier.lr.value*lr_t2_factor | |
| 656 | |
| 657 | |
| 658 #cap the patience so we are allowed to diverge max_div_count times | |
| 659 #if we are going up max_div_count in a row, we will stop immediatelty by modifying the patience | |
| 660 divergence_flag = divergence_flag +1 | |
| 661 | |
| 662 | |
| 322 | 663 #calculate the test error at this point and exit |
| 664 # test it on the test set | |
| 665 test_score = 0. | |
| 666 temp=0 | |
| 667 for xt,yt in dataset.test(batch_size): | |
| 668 test_score += test_model(xt,yt) | |
| 669 temp=temp+1 | |
| 670 test_score /= temp | |
|
355
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
671 |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
672 print ' validation error is going up, possibly stopping soon' |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
673 print((' epoch %i, minibatch %i, test error of best ' |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
674 'model %f %%') % |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
675 (epoch, minibatch_index+1, |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
676 test_score*100.)) |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
677 sys.stdout.flush() |
| 322 | 678 |
| 679 | |
| 680 | |
| 323 | 681 # check early stop condition |
| 682 if divergence_flag==max_div_count: | |
| 683 minibatch_index=nb_max_exemples | |
| 684 print 'we have diverged, early stopping kicks in' | |
| 685 break | |
| 686 | |
| 687 #check if we have seen enough exemples | |
| 688 #force one epoch at least | |
| 689 if epoch>0 and minibatch_index*batch_size>nb_max_exemples: | |
| 322 | 690 break |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
691 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
692 |
|
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
693 |
| 322 | 694 |
| 695 | |
| 696 time_n= time_n + batch_size | |
| 323 | 697 minibatch_index = minibatch_index + 1 |
| 698 | |
| 699 # we have finished looping through the training set | |
| 322 | 700 epoch = epoch+1 |
| 110 | 701 end_time = time.clock() |
|
355
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
702 |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
703 print(('Optimization complete. Best validation score of %f %% ' |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
704 'obtained at iteration %i, with test performance %f %%') % |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
705 (best_validation_loss * 100., best_iter, test_score*100.)) |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
706 print ('The code ran for %f minutes' % ((end_time-start_time)/60.)) |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
707 print minibatch_index |
|
76b7182dd32e
added support for pnist in iterator. corrected a print bug in mlp
xaviermuller
parents:
349
diff
changeset
|
708 sys.stdout.flush() |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
709 |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
710 #save the model and the weights |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
711 numpy.savez('model.npy', config=configuration, W1=classifier.W1.value,W2=classifier.W2.value, b1=classifier.b1.value,b2=classifier.b2.value) |
|
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
712 numpy.savez('results.npy',config=configuration,total_train_error_list=total_train_error_list,total_validation_error_list=total_validation_error_list,\ |
| 323 | 713 learning_rate_list=learning_rate_list, divergence_flag_list=divergence_flag_list) |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
714 |
|
414
3dba84c0fbc1
saving test score from best validation score in db now
xaviermuller
parents:
405
diff
changeset
|
715 return (best_training_error*100.0,best_validation_loss * 100.,optimal_test_error*100.,best_iter*batch_size,(end_time-start_time)/60) |
| 110 | 716 |
| 717 | |
| 718 if __name__ == '__main__': | |
| 719 mlp_full_mnist() | |
| 720 | |
| 721 def jobman_mlp_full_nist(state,channel): | |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
722 (train_error,validation_error,test_error,nb_exemples,time)=mlp_full_nist(learning_rate=state.learning_rate,\ |
|
304
1e4bf5a5b46d
added type 2 adaptive learning configurable learning weight + versionning
xaviermuller
parents:
237
diff
changeset
|
723 nb_max_exemples=state.nb_max_exemples,\ |
|
1e4bf5a5b46d
added type 2 adaptive learning configurable learning weight + versionning
xaviermuller
parents:
237
diff
changeset
|
724 nb_hidden=state.nb_hidden,\ |
|
1e4bf5a5b46d
added type 2 adaptive learning configurable learning weight + versionning
xaviermuller
parents:
237
diff
changeset
|
725 adaptive_lr=state.adaptive_lr,\ |
|
1e4bf5a5b46d
added type 2 adaptive learning configurable learning weight + versionning
xaviermuller
parents:
237
diff
changeset
|
726 tau=state.tau,\ |
|
1e4bf5a5b46d
added type 2 adaptive learning configurable learning weight + versionning
xaviermuller
parents:
237
diff
changeset
|
727 verbose = state.verbose,\ |
| 324 | 728 lr_t2_factor=state.lr_t2_factor, |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
729 data_set=state.data_set, |
| 378 | 730 init_model=state.init_model, |
|
445
868f82777839
added jobman all test + val error and sigmoid output
xaviermuller
parents:
414
diff
changeset
|
731 detection_mode = state.detection_mode,\ |
|
338
fca22114bb23
added async save, restart from old model and independant error calculation based on Arnaud's iterator
xaviermuller
parents:
324
diff
changeset
|
732 channel=channel) |
|
143
f341a4efb44a
added adaptive lr, weight file save, traine error and error curves
XavierMuller
parents:
110
diff
changeset
|
733 state.train_error=train_error |
| 110 | 734 state.validation_error=validation_error |
| 735 state.test_error=test_error | |
| 736 state.nb_exemples=nb_exemples | |
| 737 state.time=time | |
|
304
1e4bf5a5b46d
added type 2 adaptive learning configurable learning weight + versionning
xaviermuller
parents:
237
diff
changeset
|
738 pylearn.version.record_versions(state,[theano,ift6266,pylearn]) |
| 110 | 739 return channel.COMPLETE |
| 740 | |
| 741 |