Mercurial > ift6266

comparison deep/stacked_dae/stacked_dae.py @ 186:d364a130b221

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Ajout du code de base pour scalar_series. Modifications à stacked_dae: réglé un problème avec les input_divider (empêchait une optimisation), et ajouté utilisation des séries. Si j'avais pas déjà commité, aussi, j'ai enlevé l'histoire de réutilisation du pretraining: c'était compliqué (error prone) et ça créait des jobs beaucoup trop longues.
author fsavard
date Mon, 01 Mar 2010 11:45:25 -0500
parents b9ea8e2d071a
children 3632e6258642
comparison
equal deleted inserted replaced
185:b9ea8e2d071a 186:d364a130b221
150 update_locals(self, locals()) 150 update_locals(self, locals())
151 151
152 self.layers = [] 152 self.layers = []
153 self.pretrain_functions = [] 153 self.pretrain_functions = []
154 self.params = [] 154 self.params = []
155 # MODIF: added this so we also get the b_primes
156 # (not used for finetuning... still using ".params")
157 self.all_params = []
155 self.n_layers = len(hidden_layers_sizes) 158 self.n_layers = len(hidden_layers_sizes)
156 159
157 self.input_divider = numpy.asarray(input_divider, dtype=theano.config.floatX) 160 self.shared_divider = theano.shared(numpy.asarray(input_divider, dtype=theano.config.floatX))
158 161
159 if len(hidden_layers_sizes) < 1 : 162 if len(hidden_layers_sizes) < 1 :
160 raiseException (' You must have at least one hidden layer ') 163 raiseException (' You must have at least one hidden layer ')
161 164
162 165
194 # layer 197 # layer
195 dA_layer = dA(input_size, hidden_layers_sizes[i], \ 198 dA_layer = dA(input_size, hidden_layers_sizes[i], \
196 corruption_level = corruption_levels[0],\ 199 corruption_level = corruption_levels[0],\
197 input = layer_input, \ 200 input = layer_input, \
198 shared_W = layer.W, shared_b = layer.b) 201 shared_W = layer.W, shared_b = layer.b)
202
203 self.all_params += dA_layer.params
199 204
200 # Construct a function that trains this dA 205 # Construct a function that trains this dA
201 # compute gradients of layer parameters 206 # compute gradients of layer parameters
202 gparams = T.grad(dA_layer.cost, dA_layer.params) 207 gparams = T.grad(dA_layer.cost, dA_layer.params)
203 # compute the list of updates 208 # compute the list of updates
207 212
208 # create a function that trains the dA 213 # create a function that trains the dA
209 update_fn = theano.function([index], dA_layer.cost, \ 214 update_fn = theano.function([index], dA_layer.cost, \
210 updates = updates, 215 updates = updates,
211 givens = { 216 givens = {
212 self.x : train_set_x[index*batch_size:(index+1)*batch_size] / self.input_divider}) 217 self.x : train_set_x[index*batch_size:(index+1)*batch_size] / self.shared_divider})
213 # collect this function into a list 218 # collect this function into a list
214 self.pretrain_functions += [update_fn] 219 self.pretrain_functions += [update_fn]
215 220
216 221
217 # We now need to add a logistic layer on top of the MLP 222 # We now need to add a logistic layer on top of the MLP
218 self.logLayer = LogisticRegression(\ 223 self.logLayer = LogisticRegression(\
219 input = self.layers[-1].output,\ 224 input = self.layers[-1].output,\
220 n_in = hidden_layers_sizes[-1], n_out = n_outs) 225 n_in = hidden_layers_sizes[-1], n_out = n_outs)
221 226
222 self.params += self.logLayer.params 227 self.params += self.logLayer.params
228 self.all_params += self.logLayer.params
223 # construct a function that implements one step of finetunining 229 # construct a function that implements one step of finetunining
224 230
225 # compute the cost, defined as the negative log likelihood 231 # compute the cost, defined as the negative log likelihood
226 cost = self.logLayer.negative_log_likelihood(self.y) 232 cost = self.logLayer.negative_log_likelihood(self.y)
227 # compute the gradients with respect to the model parameters 233 # compute the gradients with respect to the model parameters
232 updates[param] = param - gparam*finetune_lr 238 updates[param] = param - gparam*finetune_lr
233 239
234 self.finetune = theano.function([index], cost, 240 self.finetune = theano.function([index], cost,
235 updates = updates, 241 updates = updates,
236 givens = { 242 givens = {
237 self.x : train_set_x[index*batch_size:(index+1)*batch_size]/self.input_divider, 243 self.x : train_set_x[index*batch_size:(index+1)*batch_size]/self.shared_divider,
238 self.y : train_set_y[index*batch_size:(index+1)*batch_size]} ) 244 self.y : train_set_y[index*batch_size:(index+1)*batch_size]} )
239 245
240 # symbolic variable that points to the number of errors made on the 246 # symbolic variable that points to the number of errors made on the
241 # minibatch given by self.x and self.y 247 # minibatch given by self.x and self.y
242 248