|
364
|
1 #!/usr/bin/python
|
|
|
2 # coding: utf-8
|
|
|
3
|
|
|
4 import numpy
|
|
|
5 import theano
|
|
|
6 import time
|
|
|
7 import datetime
|
|
|
8 import theano.tensor as T
|
|
|
9 import sys
|
|
|
10 import pickle
|
|
|
11
|
|
|
12 from jobman import DD
|
|
|
13 import jobman, jobman.sql
|
|
|
14 from copy import copy
|
|
|
15
|
|
|
16 from stacked_convolutional_dae_uit import CSdA
|
|
|
17
|
|
|
18 from ift6266.utils.seriestables import *
|
|
|
19
|
|
|
20 buffersize=1000
|
|
|
21
|
|
|
22 default_series = { \
|
|
|
23 'reconstruction_error' : DummySeries(),
|
|
|
24 'training_error' : DummySeries(),
|
|
|
25 'validation_error' : DummySeries(),
|
|
|
26 'test_error' : DummySeries(),
|
|
|
27 'params' : DummySeries()
|
|
|
28 }
|
|
|
29
|
|
|
30 def itermax(iter, max):
|
|
|
31 for i,it in enumerate(iter):
|
|
|
32 if i >= max:
|
|
|
33 break
|
|
|
34 yield it
|
|
|
35 def get_conv_shape(kernels,imgshp,batch_size,max_pool_layers):
|
|
|
36 # Returns the dimension at the output of the convoluational net
|
|
|
37 # and a list of Image and kernel shape for every
|
|
|
38 # Convolutional layer
|
|
|
39 conv_layers=[]
|
|
|
40 init_layer = [ [ kernels[0][0],1,kernels[0][1],kernels[0][2] ],\
|
|
|
41 [ batch_size , 1, imgshp[0], imgshp[1] ],
|
|
|
42 max_pool_layers[0] ]
|
|
|
43 conv_layers.append(init_layer)
|
|
|
44
|
|
|
45 conv_n_out = int((32-kernels[0][2]+1)/max_pool_layers[0][0])
|
|
|
46
|
|
|
47 for i in range(1,len(kernels)):
|
|
|
48 layer = [ [ kernels[i][0],kernels[i-1][0],kernels[i][1],kernels[i][2] ],\
|
|
|
49 [ batch_size, kernels[i-1][0],conv_n_out,conv_n_out ],
|
|
|
50 max_pool_layers[i] ]
|
|
|
51 conv_layers.append(layer)
|
|
|
52 conv_n_out = int( (conv_n_out - kernels[i][2]+1)/max_pool_layers[i][0])
|
|
|
53 conv_n_out=kernels[-1][0]*conv_n_out**2
|
|
|
54 return conv_n_out,conv_layers
|
|
|
55
|
|
|
56
|
|
|
57
|
|
|
58
|
|
|
59
|
|
|
60 class CSdASgdOptimizer:
|
|
|
61 def __init__(self, dataset, hyperparameters, n_ins, n_outs,
|
|
|
62 examples_per_epoch, series=default_series, max_minibatches=None):
|
|
|
63 self.dataset = dataset
|
|
|
64 self.hp = hyperparameters
|
|
|
65 self.n_ins = n_ins
|
|
|
66 self.n_outs = n_outs
|
|
|
67 self.parameters_pre=[]
|
|
|
68
|
|
|
69 self.max_minibatches = max_minibatches
|
|
|
70 print "CSdASgdOptimizer, max_minibatches =", max_minibatches
|
|
|
71
|
|
|
72 self.ex_per_epoch = examples_per_epoch
|
|
|
73 self.mb_per_epoch = examples_per_epoch / self.hp.minibatch_size
|
|
|
74
|
|
|
75 self.series = series
|
|
|
76
|
|
|
77 self.rng = numpy.random.RandomState(1234)
|
|
|
78 self.init_classifier()
|
|
|
79
|
|
|
80 sys.stdout.flush()
|
|
|
81
|
|
|
82 def init_classifier(self):
|
|
|
83 print "Constructing classifier"
|
|
|
84
|
|
|
85 n_ins,convlayers = get_conv_shape(self.hp.kernels,self.hp.imgshp,self.hp.minibatch_size,self.hp.max_pool_layers)
|
|
|
86
|
|
|
87 self.classifier = CSdA(n_ins_mlp = n_ins,
|
|
|
88 batch_size = self.hp.minibatch_size,
|
|
|
89 conv_hidden_layers_sizes = convlayers,
|
|
|
90 mlp_hidden_layers_sizes = self.hp.mlp_size,
|
|
|
91 corruption_levels = self.hp.corruption_levels,
|
|
|
92 rng = self.rng,
|
|
|
93 n_out = self.n_outs,
|
|
|
94 pretrain_lr = self.hp.pretraining_lr,
|
|
|
95 finetune_lr = self.hp.finetuning_lr)
|
|
|
96
|
|
|
97
|
|
|
98
|
|
|
99 #theano.printing.pydotprint(self.classifier.pretrain_functions[0], "function.graph")
|
|
|
100
|
|
|
101 sys.stdout.flush()
|
|
|
102
|
|
|
103 def train(self):
|
|
|
104 self.pretrain(self.dataset)
|
|
|
105 self.finetune(self.dataset)
|
|
|
106
|
|
|
107 def pretrain(self,dataset):
|
|
|
108 print "STARTING PRETRAINING, time = ", datetime.datetime.now()
|
|
|
109 sys.stdout.flush()
|
|
|
110
|
|
|
111 un_fichier=int(819200.0/self.hp.minibatch_size) #Number of batches in a P07 file
|
|
|
112
|
|
|
113 start_time = time.clock()
|
|
|
114 ## Pre-train layer-wise
|
|
|
115 for i in xrange(self.classifier.n_layers):
|
|
|
116 # go through pretraining epochs
|
|
|
117 for epoch in xrange(self.hp.pretraining_epochs_per_layer):
|
|
|
118 # go through the training set
|
|
|
119 batch_index=0
|
|
|
120 count=0
|
|
|
121 num_files=0
|
|
|
122 for x,y in dataset.train(self.hp.minibatch_size):
|
|
|
123 if x.shape[0] != self.hp.minibatch_size:
|
|
|
124 continue
|
|
|
125 c = self.classifier.pretrain_functions[i](x)
|
|
|
126 count +=1
|
|
|
127
|
|
|
128 self.series["reconstruction_error"].append((epoch, batch_index), c)
|
|
|
129 batch_index+=1
|
|
|
130
|
|
|
131 #if batch_index % 100 == 0:
|
|
|
132 # print "100 batches"
|
|
|
133
|
|
|
134 # useful when doing tests
|
|
|
135 if self.max_minibatches and batch_index >= self.max_minibatches:
|
|
|
136 break
|
|
|
137
|
|
|
138 #When we pass through the data only once (the case with P07)
|
|
|
139 #There is approximately 800*1024=819200 examples per file (1k per example and files are 800M)
|
|
|
140 if self.hp.pretraining_epochs_per_layer == 1 and count%un_fichier == 0:
|
|
|
141 print 'Pre-training layer %i, epoch %d, cost '%(i,num_files),c
|
|
|
142 num_files+=1
|
|
|
143 sys.stdout.flush()
|
|
|
144 self.series['params'].append((num_files,), self.classifier.all_params)
|
|
|
145
|
|
|
146 #When NIST is used
|
|
|
147 if self.hp.pretraining_epochs_per_layer > 1:
|
|
|
148 print 'Pre-training layer %i, epoch %d, cost '%(i,epoch),c
|
|
|
149 sys.stdout.flush()
|
|
|
150
|
|
|
151 self.series['params'].append((epoch,), self.classifier.all_params)
|
|
|
152 end_time = time.clock()
|
|
|
153
|
|
|
154 print ('Pretraining took %f minutes' %((end_time-start_time)/60.))
|
|
|
155 self.hp.update({'pretraining_time': end_time-start_time})
|
|
|
156
|
|
|
157 sys.stdout.flush()
|
|
|
158
|
|
|
159 #To be able to load them later for tests on finetune
|
|
|
160 self.parameters_pre=[copy(x.value) for x in self.classifier.params]
|
|
|
161 f = open('params_pretrain.txt', 'w')
|
|
|
162 pickle.dump(self.parameters_pre,f)
|
|
|
163 f.close()
|
|
|
164 def finetune(self,dataset,dataset_test,num_finetune,ind_test,special=0,decrease=0):
|
|
|
165
|
|
|
166 if special != 0 and special != 1:
|
|
|
167 sys.exit('Bad value for variable special. Must be in {0,1}')
|
|
|
168 print "STARTING FINETUNING, time = ", datetime.datetime.now()
|
|
|
169
|
|
|
170 minibatch_size = self.hp.minibatch_size
|
|
|
171 if ind_test == 0 or ind_test == 20:
|
|
|
172 nom_test = "NIST"
|
|
|
173 nom_train="P07"
|
|
|
174 else:
|
|
|
175 nom_test = "P07"
|
|
|
176 nom_train = "NIST"
|
|
|
177
|
|
|
178
|
|
|
179 # create a function to compute the mistakes that are made by the model
|
|
|
180 # on the validation set, or testing set
|
|
|
181 test_model = \
|
|
|
182 theano.function(
|
|
|
183 [self.classifier.x,self.classifier.y], self.classifier.errors)
|
|
|
184 # givens = {
|
|
|
185 # self.classifier.x: ensemble_x,
|
|
|
186 # self.classifier.y: ensemble_y]})
|
|
|
187
|
|
|
188 validate_model = \
|
|
|
189 theano.function(
|
|
|
190 [self.classifier.x,self.classifier.y], self.classifier.errors)
|
|
|
191 # givens = {
|
|
|
192 # self.classifier.x: ,
|
|
|
193 # self.classifier.y: ]})
|
|
|
194 # early-stopping parameters
|
|
|
195 patience = 10000 # look as this many examples regardless
|
|
|
196 patience_increase = 2. # wait this much longer when a new best is
|
|
|
197 # found
|
|
|
198 improvement_threshold = 0.995 # a relative improvement of this much is
|
|
|
199 # considered significant
|
|
|
200 validation_frequency = min(self.mb_per_epoch, patience/2)
|
|
|
201 # go through this many
|
|
|
202 # minibatche before checking the network
|
|
|
203 # on the validation set; in this case we
|
|
|
204 # check every epoch
|
|
|
205 if self.max_minibatches and validation_frequency > self.max_minibatches:
|
|
|
206 validation_frequency = self.max_minibatches / 2
|
|
|
207 best_params = None
|
|
|
208 best_validation_loss = float('inf')
|
|
|
209 test_score = 0.
|
|
|
210 start_time = time.clock()
|
|
|
211
|
|
|
212 done_looping = False
|
|
|
213 epoch = 0
|
|
|
214
|
|
|
215 total_mb_index = 0
|
|
|
216 minibatch_index = 0
|
|
|
217 parameters_finetune=[]
|
|
|
218 learning_rate = self.hp.finetuning_lr
|
|
|
219
|
|
|
220
|
|
|
221 while (epoch < num_finetune) and (not done_looping):
|
|
|
222 epoch = epoch + 1
|
|
|
223
|
|
|
224 for x,y in dataset.train(minibatch_size,bufsize=buffersize):
|
|
|
225
|
|
|
226 minibatch_index += 1
|
|
|
227
|
|
|
228 if x.shape[0] != self.hp.minibatch_size:
|
|
|
229 print 'bim'
|
|
|
230 continue
|
|
|
231
|
|
|
232 cost_ij = self.classifier.finetune(x,y)#,learning_rate)
|
|
|
233 total_mb_index += 1
|
|
|
234
|
|
|
235 self.series["training_error"].append((epoch, minibatch_index), cost_ij)
|
|
|
236
|
|
|
237 if (total_mb_index+1) % validation_frequency == 0:
|
|
|
238 #minibatch_index += 1
|
|
|
239 #The validation set is always NIST (we want the model to be good on NIST)
|
|
|
240
|
|
|
241 iter=dataset_test.valid(minibatch_size,bufsize=buffersize)
|
|
|
242
|
|
|
243
|
|
|
244 if self.max_minibatches:
|
|
|
245 iter = itermax(iter, self.max_minibatches)
|
|
|
246
|
|
|
247 validation_losses = []
|
|
|
248
|
|
|
249 for x,y in iter:
|
|
|
250 if x.shape[0] != self.hp.minibatch_size:
|
|
|
251 print 'bim'
|
|
|
252 continue
|
|
|
253 validation_losses.append(validate_model(x,y))
|
|
|
254
|
|
|
255 this_validation_loss = numpy.mean(validation_losses)
|
|
|
256
|
|
|
257 self.series["validation_error"].\
|
|
|
258 append((epoch, minibatch_index), this_validation_loss*100.)
|
|
|
259
|
|
|
260 print('epoch %i, minibatch %i, validation error on NIST : %f %%' % \
|
|
|
261 (epoch, minibatch_index+1, \
|
|
|
262 this_validation_loss*100.))
|
|
|
263
|
|
|
264
|
|
|
265 # if we got the best validation score until now
|
|
|
266 if this_validation_loss < best_validation_loss:
|
|
|
267
|
|
|
268 #improve patience if loss improvement is good enough
|
|
|
269 if this_validation_loss < best_validation_loss * \
|
|
|
270 improvement_threshold :
|
|
|
271 patience = max(patience, total_mb_index * patience_increase)
|
|
|
272
|
|
|
273 # save best validation score, iteration number and parameters
|
|
|
274 best_validation_loss = this_validation_loss
|
|
|
275 best_iter = total_mb_index
|
|
|
276 parameters_finetune=[copy(x.value) for x in self.classifier.params]
|
|
|
277
|
|
|
278 # test it on the test set
|
|
|
279 iter = dataset.test(minibatch_size,bufsize=buffersize)
|
|
|
280 if self.max_minibatches:
|
|
|
281 iter = itermax(iter, self.max_minibatches)
|
|
|
282 test_losses = []
|
|
|
283 test_losses2 = []
|
|
|
284 for x,y in iter:
|
|
|
285 if x.shape[0] != self.hp.minibatch_size:
|
|
|
286 print 'bim'
|
|
|
287 continue
|
|
|
288 test_losses.append(test_model(x,y))
|
|
|
289
|
|
|
290 test_score = numpy.mean(test_losses)
|
|
|
291
|
|
|
292 #test it on the second test set
|
|
|
293 iter2 = dataset_test.test(minibatch_size,bufsize=buffersize)
|
|
|
294 if self.max_minibatches:
|
|
|
295 iter2 = itermax(iter2, self.max_minibatches)
|
|
|
296 for x,y in iter2:
|
|
|
297 if x.shape[0] != self.hp.minibatch_size:
|
|
|
298 continue
|
|
|
299 test_losses2.append(test_model(x,y))
|
|
|
300
|
|
|
301 test_score2 = numpy.mean(test_losses2)
|
|
|
302
|
|
|
303 self.series["test_error"].\
|
|
|
304 append((epoch, minibatch_index), test_score*100.)
|
|
|
305
|
|
|
306 print((' epoch %i, minibatch %i, test error on dataset %s (train data) of best '
|
|
|
307 'model %f %%') %
|
|
|
308 (epoch, minibatch_index+1,nom_train,
|
|
|
309 test_score*100.))
|
|
|
310
|
|
|
311 print((' epoch %i, minibatch %i, test error on dataset %s of best '
|
|
|
312 'model %f %%') %
|
|
|
313 (epoch, minibatch_index+1,nom_test,
|
|
|
314 test_score2*100.))
|
|
|
315
|
|
|
316 if patience <= total_mb_index:
|
|
|
317 done_looping = True
|
|
|
318 break #to exit the FOR loop
|
|
|
319
|
|
|
320 sys.stdout.flush()
|
|
|
321
|
|
|
322 # useful when doing tests
|
|
|
323 if self.max_minibatches and minibatch_index >= self.max_minibatches:
|
|
|
324 break
|
|
|
325
|
|
|
326 if decrease == 1:
|
|
|
327 learning_rate /= 2 #divide the learning rate by 2 for each new epoch
|
|
|
328
|
|
|
329 self.series['params'].append((epoch,), self.classifier.all_params)
|
|
|
330
|
|
|
331 if done_looping == True: #To exit completly the fine-tuning
|
|
|
332 break #to exit the WHILE loop
|
|
|
333
|
|
|
334 end_time = time.clock()
|
|
|
335 self.hp.update({'finetuning_time':end_time-start_time,\
|
|
|
336 'best_validation_error':best_validation_loss,\
|
|
|
337 'test_score':test_score,
|
|
|
338 'num_finetuning_epochs':epoch})
|
|
|
339
|
|
|
340 print(('\nOptimization complete with best validation score of %f %%,'
|
|
|
341 'with test performance %f %% on dataset %s ') %
|
|
|
342 (best_validation_loss * 100., test_score*100.,nom_train))
|
|
|
343 print(('The test score on the %s dataset is %f')%(nom_test,test_score2*100.))
|
|
|
344
|
|
|
345 print ('The finetuning ran for %f minutes' % ((end_time-start_time)/60.))
|
|
|
346
|
|
|
347 sys.stdout.flush()
|
|
|
348
|
|
|
349 #Save a copy of the parameters in a file to be able to get them in the future
|
|
|
350
|
|
|
351 if special == 1: #To keep a track of the value of the parameters
|
|
|
352 f = open('params_finetune_stanford.txt', 'w')
|
|
|
353 pickle.dump(parameters_finetune,f)
|
|
|
354 f.close()
|
|
|
355
|
|
|
356 elif ind_test == 0 | ind_test == 20: #To keep a track of the value of the parameters
|
|
|
357 f = open('params_finetune_P07.txt', 'w')
|
|
|
358 pickle.dump(parameters_finetune,f)
|
|
|
359 f.close()
|
|
|
360
|
|
|
361
|
|
|
362 elif ind_test== 1: #For the run with 2 finetunes. It will be faster.
|
|
|
363 f = open('params_finetune_NIST.txt', 'w')
|
|
|
364 pickle.dump(parameters_finetune,f)
|
|
|
365 f.close()
|
|
|
366
|
|
|
367 elif ind_test== 21: #To keep a track of the value of the parameters
|
|
|
368 f = open('params_finetune_P07_then_NIST.txt', 'w')
|
|
|
369 pickle.dump(parameters_finetune,f)
|
|
|
370 f.close()
|
|
|
371 #Set parameters like they where right after pre-train or finetune
|
|
|
372 def reload_parameters(self,which):
|
|
|
373
|
|
|
374 #self.parameters_pre=pickle.load('params_pretrain.txt')
|
|
|
375 f = open(which)
|
|
|
376 self.parameters_pre=pickle.load(f)
|
|
|
377 f.close()
|
|
|
378 for idx,x in enumerate(self.parameters_pre):
|
|
|
379 if x.dtype=='float64':
|
|
|
380 self.classifier.params[idx].value=theano._asarray(copy(x),dtype=theano.config.floatX)
|
|
|
381 else:
|
|
|
382 self.classifier.params[idx].value=copy(x)
|
|
|
383
|
|
|
384 def training_error(self,dataset):
|
|
|
385 # create a function to compute the mistakes that are made by the model
|
|
|
386 # on the validation set, or testing set
|
|
|
387 test_model = \
|
|
|
388 theano.function(
|
|
|
389 [self.classifier.x,self.classifier.y], self.classifier.errors)
|
|
|
390
|
|
|
391 iter2 = dataset.train(self.hp.minibatch_size,bufsize=buffersize)
|
|
|
392 train_losses2 = [test_model(x,y) for x,y in iter2]
|
|
|
393 train_score2 = numpy.mean(train_losses2)
|
|
|
394 print "Training error is: " + str(train_score2)
|
