Mercurial > ift6266

comparison deep/stacked_dae/v_sylvain/sgd_optimization.py @ 310:a5abd5c8b4b0

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Petits changements pour pouvoir utiliser le GPU
author SylvainPL <sylvain.pannetier.lebeuf@umontreal.ca>
date Thu, 01 Apr 2010 13:44:04 -0400
parents fe5d428c2acc
children 403b9e6ecfaa
comparison
equal deleted inserted replaced
309:60cacb9a70e4 310:a5abd5c8b4b0
16 from copy import copy 16 from copy import copy
17 17
18 from stacked_dae import SdA 18 from stacked_dae import SdA
19 19
20 from ift6266.utils.seriestables import * 20 from ift6266.utils.seriestables import *
21
22 #For test purpose only
23 buffersize=1000
21 24
22 default_series = { \ 25 default_series = { \
23 'reconstruction_error' : DummySeries(), 26 'reconstruction_error' : DummySeries(),
24 'training_error' : DummySeries(), 27 'training_error' : DummySeries(),
25 'validation_error' : DummySeries(), 28 'validation_error' : DummySeries(),
201 parameters_finetune=[] 204 parameters_finetune=[]
202 205
203 while (epoch < num_finetune) and (not done_looping): 206 while (epoch < num_finetune) and (not done_looping):
204 epoch = epoch + 1 207 epoch = epoch + 1
205 208
206 for x,y in dataset.train(minibatch_size): 209 for x,y in dataset.train(minibatch_size,bufsize=buffersize):
207 minibatch_index += 1 210 minibatch_index += 1
208 if special == 0: 211 if special == 0:
209 cost_ij = self.classifier.finetune(x,y) 212 cost_ij = self.classifier.finetune(x,y)
210 elif special == 1: 213 elif special == 1:
211 cost_ij = self.classifier.finetune2(x,y) 214 cost_ij = self.classifier.finetune2(x,y)
215 218
216 if (total_mb_index+1) % validation_frequency == 0: 219 if (total_mb_index+1) % validation_frequency == 0:
217 #minibatch_index += 1 220 #minibatch_index += 1
218 #The validation set is always NIST (we want the model to be good on NIST) 221 #The validation set is always NIST (we want the model to be good on NIST)
219 if ind_test == 0 | ind_test == 20: 222 if ind_test == 0 | ind_test == 20:
220 iter=dataset_test.valid(minibatch_size) 223 iter=dataset_test.valid(minibatch_size,bufsize=buffersize)
221 else: 224 else:
222 iter = dataset.valid(minibatch_size) 225 iter = dataset.valid(minibatch_size,bufsize=buffersize)
223 if self.max_minibatches: 226 if self.max_minibatches:
224 iter = itermax(iter, self.max_minibatches) 227 iter = itermax(iter, self.max_minibatches)
225 validation_losses = [validate_model(x,y) for x,y in iter] 228 validation_losses = [validate_model(x,y) for x,y in iter]
226 this_validation_loss = numpy.mean(validation_losses) 229 this_validation_loss = numpy.mean(validation_losses)
227 230
245 best_validation_loss = this_validation_loss 248 best_validation_loss = this_validation_loss
246 best_iter = total_mb_index 249 best_iter = total_mb_index
247 parameters_finetune=[copy(x.value) for x in self.classifier.params] 250 parameters_finetune=[copy(x.value) for x in self.classifier.params]
248 251
249 # test it on the test set 252 # test it on the test set
250 iter = dataset.test(minibatch_size) 253 iter = dataset.test(minibatch_size,bufsize=buffersize)
251 if self.max_minibatches: 254 if self.max_minibatches:
252 iter = itermax(iter, self.max_minibatches) 255 iter = itermax(iter, self.max_minibatches)
253 test_losses = [test_model(x,y) for x,y in iter] 256 test_losses = [test_model(x,y) for x,y in iter]
254 test_score = numpy.mean(test_losses) 257 test_score = numpy.mean(test_losses)
255 258
256 #test it on the second test set 259 #test it on the second test set
257 iter2 = dataset_test.test(minibatch_size) 260 iter2 = dataset_test.test(minibatch_size,bufsize=buffersize)
258 if self.max_minibatches: 261 if self.max_minibatches:
259 iter2 = itermax(iter2, self.max_minibatches) 262 iter2 = itermax(iter2, self.max_minibatches)
260 test_losses2 = [test_model(x,y) for x,y in iter2] 263 test_losses2 = [test_model(x,y) for x,y in iter2]
261 test_score2 = numpy.mean(test_losses2) 264 test_score2 = numpy.mean(test_losses2)
262 265
332 #self.parameters_pre=pickle.load('params_pretrain.txt') 335 #self.parameters_pre=pickle.load('params_pretrain.txt')
333 f = open(which) 336 f = open(which)
334 self.parameters_pre=pickle.load(f) 337 self.parameters_pre=pickle.load(f)
335 f.close() 338 f.close()
336 for idx,x in enumerate(self.parameters_pre): 339 for idx,x in enumerate(self.parameters_pre):
337 self.classifier.params[idx].value=copy(x) 340 if x.dtype=='float64':
338 341 self.classifier.params[idx].value=theano._asarray(copy(x),dtype=theano.config.floatX)
339 342 else:
340 343 self.classifier.params[idx].value=copy(x)
341 344
342 345
346
347
348