Mercurial > ift6266
diff deep/stacked_dae/stacked_dae.py @ 186:d364a130b221
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 |
line wrap: on
line diff
--- a/deep/stacked_dae/stacked_dae.py Fri Feb 26 17:45:52 2010 -0500 +++ b/deep/stacked_dae/stacked_dae.py Mon Mar 01 11:45:25 2010 -0500 @@ -152,9 +152,12 @@ self.layers = [] self.pretrain_functions = [] self.params = [] + # MODIF: added this so we also get the b_primes + # (not used for finetuning... still using ".params") + self.all_params = [] self.n_layers = len(hidden_layers_sizes) - self.input_divider = numpy.asarray(input_divider, dtype=theano.config.floatX) + self.shared_divider = theano.shared(numpy.asarray(input_divider, dtype=theano.config.floatX)) if len(hidden_layers_sizes) < 1 : raiseException (' You must have at least one hidden layer ') @@ -196,6 +199,8 @@ corruption_level = corruption_levels[0],\ input = layer_input, \ shared_W = layer.W, shared_b = layer.b) + + self.all_params += dA_layer.params # Construct a function that trains this dA # compute gradients of layer parameters @@ -209,7 +214,7 @@ update_fn = theano.function([index], dA_layer.cost, \ updates = updates, givens = { - self.x : train_set_x[index*batch_size:(index+1)*batch_size] / self.input_divider}) + self.x : train_set_x[index*batch_size:(index+1)*batch_size] / self.shared_divider}) # collect this function into a list self.pretrain_functions += [update_fn] @@ -220,6 +225,7 @@ n_in = hidden_layers_sizes[-1], n_out = n_outs) self.params += self.logLayer.params + self.all_params += self.logLayer.params # construct a function that implements one step of finetunining # compute the cost, defined as the negative log likelihood @@ -234,7 +240,7 @@ self.finetune = theano.function([index], cost, updates = updates, givens = { - self.x : train_set_x[index*batch_size:(index+1)*batch_size]/self.input_divider, + self.x : train_set_x[index*batch_size:(index+1)*batch_size]/self.shared_divider, self.y : train_set_y[index*batch_size:(index+1)*batch_size]} ) # symbolic variable that points to the number of errors made on the