Mercurial > ift6266

comparison baseline/mlp/mlp_nist.py @ 338:fca22114bb23

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
added async save, restart from old model and independant error calculation based on Arnaud's iterator
author xaviermuller
date Sat, 17 Apr 2010 12:42:48 -0400
parents 1763c64030d1
children 22efb4968054
comparison
equal deleted inserted replaced
337:8d116d4a7593 338:fca22114bb23
161 # represents a mistake in prediction 161 # represents a mistake in prediction
162 return T.mean(T.neq(self.y_pred, y)) 162 return T.mean(T.neq(self.y_pred, y))
163 else: 163 else:
164 raise NotImplementedError() 164 raise NotImplementedError()
165 165
166 def mlp_get_nist_error(model_name='/u/mullerx/ift6266h10_sandbox_db/xvm_final_lr1_p073/8/best_model.npy.npz',
167 data_set=0):
168
169
170
171 # allocate symbolic variables for the data
172 x = T.fmatrix() # the data is presented as rasterized images
173 y = T.lvector() # the labels are presented as 1D vector of
174 # [long int] labels
175
176 # load the data set and create an mlp based on the dimensions of the model
177 model=numpy.load(model_name)
178 W1=model['W1']
179 W2=model['W2']
180 b1=model['b1']
181 b2=model['b2']
182 nb_hidden=b1.shape[0]
183 input_dim=W1.shape[0]
184 nb_targets=b2.shape[0]
185 learning_rate=0.1
186
187
188 if data_set==0:
189 dataset=datasets.nist_all()
190 elif data_set==1:
191 dataset=datasets.nist_P07()
192
193
194 classifier = MLP( input=x,\
195 n_in=input_dim,\
196 n_hidden=nb_hidden,\
197 n_out=nb_targets,
198 learning_rate=learning_rate)
199
200
201 #overwrite weights with weigths from model
202 classifier.W1.value=W1
203 classifier.W2.value=W2
204 classifier.b1.value=b1
205 classifier.b2.value=b2
206
207
208 cost = classifier.negative_log_likelihood(y) \
209 + 0.0 * classifier.L1 \
210 + 0.0 * classifier.L2_sqr
211
212 # compiling a theano function that computes the mistakes that are made by
213 # the model on a minibatch
214 test_model = theano.function([x,y], classifier.errors(y))
215
216
217
218 #get the test error
219 #use a batch size of 1 so we can get the sub-class error
220 #without messing with matrices (will be upgraded later)
221 test_score=0
222 temp=0
223 for xt,yt in dataset.test(20):
224 test_score += test_model(xt,yt)
225 temp = temp+1
226 test_score /= temp
227
228
229 return test_score*100
230
231
232
233
234
166 235
167 def mlp_full_nist( verbose = 1,\ 236 def mlp_full_nist( verbose = 1,\
168 adaptive_lr = 0,\ 237 adaptive_lr = 0,\
169 data_set=0,\ 238 data_set=0,\
170 learning_rate=0.01,\ 239 learning_rate=0.01,\
172 L2_reg = 0.0001,\ 241 L2_reg = 0.0001,\
173 nb_max_exemples=1000000,\ 242 nb_max_exemples=1000000,\
174 batch_size=20,\ 243 batch_size=20,\
175 nb_hidden = 30,\ 244 nb_hidden = 30,\
176 nb_targets = 62, 245 nb_targets = 62,
177 tau=1e6,\ 246 tau=1e6,\
178 lr_t2_factor=0.5): 247 lr_t2_factor=0.5,\
179 248 init_model=0,\
180 249 channel=0):
250
251
252 if channel!=0:
253 channel.save()
181 configuration = [learning_rate,nb_max_exemples,nb_hidden,adaptive_lr] 254 configuration = [learning_rate,nb_max_exemples,nb_hidden,adaptive_lr]
182 255
183 #save initial learning rate if classical adaptive lr is used 256 #save initial learning rate if classical adaptive lr is used
184 initial_lr=learning_rate 257 initial_lr=learning_rate
185 max_div_count=3 258 max_div_count=1000
186 259
187 260
188 total_validation_error_list = [] 261 total_validation_error_list = []
189 total_train_error_list = [] 262 total_train_error_list = []
190 learning_rate_list=[] 263 learning_rate_list=[]
213 n_hidden=nb_hidden,\ 286 n_hidden=nb_hidden,\
214 n_out=nb_targets, 287 n_out=nb_targets,
215 learning_rate=learning_rate) 288 learning_rate=learning_rate)
216 289
217 290
291 # check if we want to initialise the weights with a previously calculated model
292 # dimensions must be consistent between old model and current configuration!!!!!! (nb_hidden and nb_targets)
293 if init_model!=0:
294 old_model=numpy.load(init_model)
295 classifier.W1.value=old_model['W1']
296 classifier.W2.value=old_model['W2']
297 classifier.b1.value=old_model['b1']
298 classifier.b2.value=old_model['b2']
218 299
219 300
220 # the cost we minimize during training is the negative log likelihood of 301 # the cost we minimize during training is the negative log likelihood of
221 # the model plus the regularization terms (L1 and L2); cost is expressed 302 # the model plus the regularization terms (L1 and L2); cost is expressed
222 # here symbolically 303 # here symbolically
301 382
302 383
303 #train model 384 #train model
304 cost_ij = train_model(x,y) 385 cost_ij = train_model(x,y)
305 386
306 if (minibatch_index+1) % validation_frequency == 0: 387 if (minibatch_index) % validation_frequency == 0:
307 #save the current learning rate 388 #save the current learning rate
308 learning_rate_list.append(classifier.lr.value) 389 learning_rate_list.append(classifier.lr.value)
309 divergence_flag_list.append(divergence_flag) 390 divergence_flag_list.append(divergence_flag)
391
392 #save temp results to check during training
393 numpy.savez('temp_results.npy',config=configuration,total_validation_error_list=total_validation_error_list,\
394 learning_rate_list=learning_rate_list, divergence_flag_list=divergence_flag_list)
310 395
311 # compute the validation error 396 # compute the validation error
312 this_validation_loss = 0. 397 this_validation_loss = 0.
313 temp=0 398 temp=0
314 for xv,yv in dataset.valid(1): 399 for xv,yv in dataset.valid(1):
391 476
392 #check if we have seen enough exemples 477 #check if we have seen enough exemples
393 #force one epoch at least 478 #force one epoch at least
394 if epoch>0 and minibatch_index*batch_size>nb_max_exemples: 479 if epoch>0 and minibatch_index*batch_size>nb_max_exemples:
395 break 480 break
481
482
483
396 484
397 485
398 time_n= time_n + batch_size 486 time_n= time_n + batch_size
399 minibatch_index = minibatch_index + 1 487 minibatch_index = minibatch_index + 1
400 488
425 nb_hidden=state.nb_hidden,\ 513 nb_hidden=state.nb_hidden,\
426 adaptive_lr=state.adaptive_lr,\ 514 adaptive_lr=state.adaptive_lr,\
427 tau=state.tau,\ 515 tau=state.tau,\
428 verbose = state.verbose,\ 516 verbose = state.verbose,\
429 lr_t2_factor=state.lr_t2_factor, 517 lr_t2_factor=state.lr_t2_factor,
430 data_set=state.data_set) 518 data_set=state.data_set,
519 channel=channel)
431 state.train_error=train_error 520 state.train_error=train_error
432 state.validation_error=validation_error 521 state.validation_error=validation_error
433 state.test_error=test_error 522 state.test_error=test_error
434 state.nb_exemples=nb_exemples 523 state.nb_exemples=nb_exemples
435 state.time=time 524 state.time=time