class DNN_Dropout(nnet):
def __init__(self, numpy_rng, theano_rng=None, n_ins=784,
hidden_layers_sizes=[500, 500], n_outs=10,
activation = T.nnet.sigmoid, input_dropout_factor = 0,
dropout_factor = [0.2,0.2,0.2,0.2,0.2,0.2,0.2],
adv_activation = None, max_col_norm = None,
l1_reg = None, l2_reg = None):
super(DNN_Dropout, self).__init__()
self.layers = []
self.dropout_layers = []
self.n_layers = len(hidden_layers_sizes)
self.max_col_norm = max_col_norm
self.l1_reg = l1_reg
self.l2_reg = l2_reg
self.input_dropout_factor = input_dropout_factor
self.dropout_factor = dropout_factor
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# allocate symbolic variables for the data
self.x = T.matrix('x')
self.y = T.ivector('y')
for i in xrange(self.n_layers):
# construct the sigmoidal layer
if i == 0:
input_size = n_ins
layer_input = self.x
if input_dropout_factor > 0.0:
dropout_layer_input = _dropout_from_layer(theano_rng, self.x, input_dropout_factor)
else:
dropout_layer_input = self.x
else:
input_size = hidden_layers_sizes[i - 1]
layer_input = (1 - self.dropout_factor[i - 1]) * self.layers[-1].output
dropout_layer_input = self.dropout_layers[-1].dropout_output
if not adv_activation is None:
dropout_layer = DropoutHiddenLayer(rng=numpy_rng,
input=dropout_layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i] * adv_activation['pool_size'],
activation= activation,
adv_activation_method = adv_activation['method'],
pool_size = adv_activation['pool_size'],
pnorm_order = adv_activation['pnorm_order'],
dropout_factor=self.dropout_factor[i])
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i] * adv_activation['pool_size'],
activation=activation,
adv_activation_method = adv_activation['method'],
pool_size = adv_activation['pool_size'],
pnorm_order = adv_activation['pnorm_order'],
W=dropout_layer.W, b=dropout_layer.b)
else:
dropout_layer = DropoutHiddenLayer(rng=numpy_rng,
input=dropout_layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation= activation,
dropout_factor=self.dropout_factor[i])
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i] ,
activation= activation,
W=dropout_layer.W, b=dropout_layer.b)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
self.dropout_layers.append(dropout_layer)
self.params.extend(dropout_layer.params)
self.delta_params.extend(dropout_layer.delta_params)
# We now need to add a logistic layer on top of the MLP
self.dropout_logLayer = LogisticRegression(
input=self.dropout_layers[-1].dropout_output,
n_in=hidden_layers_sizes[-1], n_out=n_outs)
self.logLayer = LogisticRegression(
input=(1 - self.dropout_factor[-1]) * self.layers[-1].output,
n_in=hidden_layers_sizes[-1], n_out=n_outs,
W=self.dropout_logLayer.W, b=self.dropout_logLayer.b)
self.dropout_layers.append(self.dropout_logLayer)
self.layers.append(self.logLayer)
self.params.extend(self.dropout_logLayer.params)
self.delta_params.extend(self.dropout_logLayer.delta_params)
# compute the cost
self.finetune_cost = self.dropout_logLayer.negative_log_likelihood(self.y)
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction();
self.features = self.layers[-2].output;
self.features_dim = self.layers[-2].n_out
if self.l1_reg is not None:
self.__l1Regularization__();
if self.l2_reg is not None:
self.__l2Regularization__();
def save(self,filename,start_layer = 0,max_layer_num = -1,withfinal=True):
nnet_dict = {}
if max_layer_num == -1:
max_layer_num = self.n_layers
for i in range(start_layer, max_layer_num):
dict_a = str(i) + ' W'
if i == 0:
nnet_dict[dict_a] = _array2string((1.0 - self.input_dropout_factor) * (
self.layers[i].params[0].get_value()))
else:
nnet_dict[dict_a] = _array2string((1.0 - self.dropout_factor[i - 1])* (
self.layers[i].params[0].get_value()))
dict_a = str(i) + ' b'
nnet_dict[dict_a] = _array2string(self.layers[i].params[1].get_value())
if withfinal:
dict_a = 'logreg W'
nnet_dict[dict_a] = _array2string((1.0 - self.dropout_factor[-1])* (
self.logLayer.params[0].get_value()))
dict_a = 'logreg b'
nnet_dict[dict_a] = _array2string(self.logLayer.params[1].get_value())
with open(filename, 'wb') as fp:
json.dump(nnet_dict, fp, indent=2, sort_keys = True)
fp.flush()
def load(self,filename,start_layer = 0,max_layer_num = -1,withfinal=True):
nnet_dict = {}
if max_layer_num == -1:
max_layer_num = self.n_layers
with open(filename, 'rb') as fp:
nnet_dict = json.load(fp)
for i in xrange(max_layer_num):
dict_key = str(i) + ' W'
self.layers[i].params[0].set_value(numpy.asarray(_string2array(nnet_dict[dict_key]),
dtype=theano.config.floatX))
dict_key = str(i) + ' b'
self.layers[i].params[1].set_value(numpy.asarray(_string2array(nnet_dict[dict_key]),
dtype=theano.config.floatX))
if withfinal:
dict_key = 'logreg W'
self.logLayer.params[0].set_value(numpy.asarray(_string2array(nnet_dict[dict_key]),
dtype=theano.config.floatX))
dict_key = 'logreg b'
self.logLayer.params[1].set_value(numpy.asarray(_string2array(nnet_dict[dict_key]),
dtype=theano.config.floatX))
def __init__(self, numpy_rng, theano_rng=None, n_ins=784,
hidden_layers_sizes=[500, 500], n_outs=10,
activation = T.nnet.sigmoid, input_dropout_factor = 0,
dropout_factor = [0.2,0.2,0.2,0.2,0.2,0.2,0.2],
adv_activation = None, max_col_norm = None,
l1_reg = None, l2_reg = None):
super(DNN_Dropout, self).__init__()
self.layers = []
self.dropout_layers = []
self.n_layers = len(hidden_layers_sizes)
self.max_col_norm = max_col_norm
self.l1_reg = l1_reg
self.l2_reg = l2_reg
self.input_dropout_factor = input_dropout_factor
self.dropout_factor = dropout_factor
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# allocate symbolic variables for the data
self.x = T.matrix('x')
self.y = T.ivector('y')
for i in xrange(self.n_layers):
# construct the sigmoidal layer
if i == 0:
input_size = n_ins
layer_input = self.x
if input_dropout_factor > 0.0:
dropout_layer_input = _dropout_from_layer(theano_rng, self.x, input_dropout_factor)
else:
dropout_layer_input = self.x
else:
input_size = hidden_layers_sizes[i - 1]
layer_input = (1 - self.dropout_factor[i - 1]) * self.layers[-1].output
dropout_layer_input = self.dropout_layers[-1].dropout_output
if not adv_activation is None:
dropout_layer = DropoutHiddenLayer(rng=numpy_rng,
input=dropout_layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i] * adv_activation['pool_size'],
activation= activation,
adv_activation_method = adv_activation['method'],
pool_size = adv_activation['pool_size'],
pnorm_order = adv_activation['pnorm_order'],
dropout_factor=self.dropout_factor[i])
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i] * adv_activation['pool_size'],
activation=activation,
adv_activation_method = adv_activation['method'],
pool_size = adv_activation['pool_size'],
pnorm_order = adv_activation['pnorm_order'],
W=dropout_layer.W, b=dropout_layer.b)
else:
dropout_layer = DropoutHiddenLayer(rng=numpy_rng,
input=dropout_layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation= activation,
dropout_factor=self.dropout_factor[i])
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i] ,
activation= activation,
W=dropout_layer.W, b=dropout_layer.b)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
self.dropout_layers.append(dropout_layer)
self.params.extend(dropout_layer.params)
self.delta_params.extend(dropout_layer.delta_params)
# We now need to add a logistic layer on top of the MLP
self.dropout_logLayer = LogisticRegression(
input=self.dropout_layers[-1].dropout_output,
n_in=hidden_layers_sizes[-1], n_out=n_outs)
self.logLayer = LogisticRegression(
input=(1 - self.dropout_factor[-1]) * self.layers[-1].output,
n_in=hidden_layers_sizes[-1], n_out=n_outs,
W=self.dropout_logLayer.W, b=self.dropout_logLayer.b)
self.dropout_layers.append(self.dropout_logLayer)
self.layers.append(self.logLayer)
self.params.extend(self.dropout_logLayer.params)
self.delta_params.extend(self.dropout_logLayer.delta_params)
# compute the cost
self.finetune_cost = self.dropout_logLayer.negative_log_likelihood(self.y)
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction();
self.features = self.layers[-2].output;
self.features_dim = self.layers[-2].n_out
if self.l1_reg is not None:
self.__l1Regularization__();
if self.l2_reg is not None:
self.__l2Regularization__();
def __init__(self, numpy_rng, theano_rng, batch_size, n_outs,conv_layer_configs, hidden_layer_configs,
use_fast=False,hidden_activation = T.nnet.sigmoid,l1_reg=None,l2_reg=None,
max_col_norm=None,input_dropout_factor=0.0):
super(DropoutCNN, self).__init__(conv_layer_configs,hidden_layer_configs,l1_reg,l2_reg,max_col_norm)
self.input_dropout_factor = input_dropout_factor;
self.dropout_layers = []
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
for i in xrange(self.conv_layer_num): # construct the convolution layer
if i == 0: #is_input layer
conv_input = self.x
if self.input_dropout_factor > 0.0:
dropout_conv_input = _dropout_from_layer(theano_rng, self.x,self.input_dropout_factor)
else:
dropout_conv_input = self.x;
else:
conv_input = (1-conv_layer_configs[i-1]['dropout_factor'])*self.layers[-1].output #output of previous layer
dropout_conv_input = self.dropout_layers[-1].dropout_output;
config = conv_layer_configs[i]
conv_activation= parse_activation(config['activation'])
dropout_conv_layer = DropoutConvLayer(numpy_rng=numpy_rng, input=dropout_conv_input,
input_shape=config['input_shape'],filter_shape=config['filter_shape'],poolsize=config['poolsize'],
activation = conv_activation, use_fast = use_fast,dropout_factor=conv_layer_configs[i]['dropout_factor'])
conv_layer = ConvLayer(numpy_rng=numpy_rng, input=conv_input,input_shape=config['input_shape'],
filter_shape=config['filter_shape'],poolsize=config['poolsize'],activation = conv_activation,
use_fast = use_fast, W = dropout_conv_layer.W, b = dropout_conv_layer.b)
self.dropout_layers.append(dropout_conv_layer);
self.layers.append(conv_layer)
self.conv_layers.append(conv_layer)
if config['update']==True: # only few layers of convolution layer are considered for updation
self.params.extend(dropout_conv_layer.params)
self.delta_params.extend(dropout_conv_layer.delta_params)
hidden_layers = hidden_layer_configs['hidden_layers'];
self.conv_output_dim = config['output_shape'][1] * config['output_shape'][2] * config['output_shape'][3]
adv_activation_configs = hidden_layer_configs['adv_activation']
#flattening the last convolution output layer
self.dropout_features = self.dropout_layers[-1].dropout_output.flatten(2);
self.features = self.conv_layers[-1].output.flatten(2);
self.features_dim = self.conv_output_dim;
self.dropout_layers = [];
self.dropout_factor = hidden_layer_configs['dropout_factor'];
for i in xrange(self.hidden_layer_num): # construct the hidden layer
if i == 0: # is first sigmoidal layer
input_size = self.conv_output_dim
dropout_layer_input = self.dropout_features
layer_input = self.features
else:
input_size = hidden_layers[i - 1] # number of hidden neurons in previous layers
dropout_layer_input = self.dropout_layers[-1].dropout_output
layer_input = (1 - self.dropout_factor[i-1]) * self.layers[-1].output
if adv_activation_configs is None:
dropout_sigmoid_layer = DropoutHiddenLayer(rng=numpy_rng, input=layer_input,n_in=input_size,
n_out = hidden_layers[i], activation=hidden_activation,
dropout_factor = self.dropout_factor[i]);
sigmoid_layer = HiddenLayer(rng=numpy_rng, input=layer_input,n_in=input_size,
n_out = hidden_layers[i], activation=hidden_activation,
W=dropout_sigmoid_layer.W, b=dropout_sigmoid_layer.b);
else:
dropout_sigmoid_layer = DropoutHiddenLayer(rng=numpy_rng, input=layer_input,n_in=input_size,
n_out = hidden_layers[i]*adv_activation_configs['pool_size'], activation=hidden_activation,
adv_activation_method = adv_activation_configs['method'],
pool_size = adv_activation_configs['pool_size'],
pnorm_order = adv_activation_configs['pnorm_order'],
dropout_factor = self.dropout_factor[i]);
sigmoid_layer = HiddenLayer(rng=numpy_rng, input=layer_input,n_in=input_size,
n_out = hidden_layers[i]*adv_activation_configs['pool_size'], activation=hidden_activation,
adv_activation_method = adv_activation_configs['method'],
pool_size = adv_activation_configs['pool_size'],
pnorm_order = adv_activation_configs['pnorm_order'],
W=dropout_sigmoid_layer.W, b=dropout_sigmoid_layer.b);
self.layers.append(sigmoid_layer)
self.dropout_layers.append(dropout_sigmoid_layer)
self.mlp_layers.append(sigmoid_layer)
if config['update']==True: # only few layers of hidden layer are considered for updation
self.params.extend(dropout_sigmoid_layer.params)
self.delta_params.extend(dropout_sigmoid_layer.delta_params)
self.dropout_logLayer = LogisticRegression(input=self.dropout_layers[-1].dropout_output,n_in=hidden_layers[-1],n_out=n_outs)
self.logLayer = LogisticRegression(
input=(1 - self.dropout_factor[-1]) * self.layers[-1].output,
n_in=hidden_layers[-1],n_out=n_outs,
W=self.dropout_logLayer.W, b=self.dropout_logLayer.b)
self.dropout_layers.append(self.dropout_logLayer)
self.layers.append(self.logLayer)
self.params.extend(self.dropout_logLayer.params)
self.delta_params.extend(self.dropout_logLayer.delta_params)
self.finetune_cost = self.dropout_logLayer.negative_log_likelihood(self.y)
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction()
#regularization
if self.l1_reg is not None:
self.__l1Regularization__(self.hidden_layer_num*2);
if self.l2_reg is not None:
self.__l2Regularization__(self.hidden_layer_num*2);
def __init__(self, numpy_rng, theano_rng, batch_size, n_outs,conv_layer_configs, hidden_layer_configs,
hidden_activation = T.nnet.sigmoid,l1_reg=None,l2_reg=None,max_col_norm=None):
super(CNN3D, self).__init__(conv_layer_configs, hidden_layer_configs,l1_reg,l2_reg,max_col_norm)
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
for i in xrange(self.conv_layer_num): # construct the convolution layer
if i == 0: #is_input layer
input = self.x
else:
input = self.layers[-1].output #output of previous layer
config = conv_layer_configs[i]
conv_activation = parse_activation(config['activation']);
conv_layer = ConvLayer(numpy_rng=numpy_rng, input=input,input_shape=config['input_shape'],
filter_shape=config['filter_shape'],poolsize=config['poolsize'],
activation = conv_activation)
self.layers.append(conv_layer)
self.conv_layers.append(conv_layer)
if config['update']==True: # only few layers of convolution layer are considered for updation
self.params.extend(conv_layer.params)
self.delta_params.extend(conv_layer.delta_params)
hidden_layers = hidden_layer_configs['hidden_layers'];
self.conv_output_dim = numpy.prod(config['output_shape'][1:])
adv_activation_configs = hidden_layer_configs['adv_activation']
#flattening the last convolution output layer
self.features = self.conv_layers[-1].output.flatten(2);
self.features_dim = self.conv_output_dim;
for i in xrange(self.hidden_layer_num): # construct the hidden layer
if i == 0: # is first sigmoidla layer
input_size = self.conv_output_dim
layer_input = self.features
else:
input_size = hidden_layers[i - 1] # number of hidden neurons in previous layers
layer_input = self.layers[-1].output
if adv_activation_configs is None:
sigmoid_layer = HiddenLayer(rng=numpy_rng, input=layer_input,n_in=input_size,
n_out = hidden_layers[i], activation=hidden_activation);
else:
sigmoid_layer = HiddenLayer(rng=numpy_rng, input=layer_input,n_in=input_size,
n_out = hidden_layers[i]*adv_activation_configs['pool_size'], activation=hidden_activation,
adv_activation_method = adv_activation_configs['method'],
pool_size = adv_activation_configs['pool_size'],
pnorm_order = adv_activation_configs['pnorm_order']);
self.layers.append(sigmoid_layer)
self.mlp_layers.append(sigmoid_layer)
if config['update']==True: # only few layers of hidden layer are considered for updation
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
self.logLayer = LogisticRegression(input=self.layers[-1].output,n_in=hidden_layers[-1],n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction()
#regularization
if self.l1_reg is not None:
self.__l1Regularization__(self.hidden_layer_num*2);
if self.l2_reg is not None:
self.__l2Regularization__(self.hidden_layer_num*2);
def __init__(self,
numpy_rng,
theano_rng=None,
n_ins=784,
hidden_layers_sizes=[500, 500],
n_outs=10,
first_layer_gb=True,
pretrainedLayers=None,
activation=T.nnet.sigmoid):
"""This class is made to support a variable number of layers.
:type numpy_rng: numpy.random.RandomState
:param numpy_rng: numpy random number generator used to draw initial
weights
:type theano_rng: theano.tensor.shared_randomstreams.RandomStreams
:param theano_rng: Theano random generator; if None is given one is
generated based on a seed drawn from `rng`
:type n_ins: int
:param n_ins: dimension of the input to the DBN
:type hidden_layers_sizes: list of ints
:param hidden_layers_sizes: intermediate layers size, must contain
at least one value
:type n_outs: int
:param n_outs: dimension of the output of the network
:type first_layer_gb: bool
:param first_layer_gb: wether first layer is gausian-bernolli or
bernolli-bernolli
"""
super(DBN, self).__init__()
self.layers = []
self.rbm_layers = []
self.n_layers = len(hidden_layers_sizes)
if pretrainedLayers == None:
self.nPreTrainLayers = n_layers
else:
self.nPreTrainLayers = pretrainedLayers
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2**30))
# allocate symbolic variables for the data
self.x = T.matrix('x') # the data is presented as rasterized images
self.y = T.ivector('y') # the labels are presented as 1D vector
# of [int] labels
# The DBN is an MLP, for which all weights of intermediate
# layers are shared with a different RBM. We will first
# construct the DBN as a deep multilayer perceptron, and when
# constructing each sigmoidal layer we also construct an RBM
# that shares weights with that layer. During pretraining we
# will train these RBMs (which will lead to chainging the
# weights of the MLP as well) During finetuning we will finish
# training the DBN by doing stochastic gradient descent on the
# MLP.
for i in xrange(self.n_layers):
# construct the sigmoidal layer
# the size of the input is either the number of hidden
# units of the layer below or the input size if we are on
# the first layer
# the input to this layer is either the activation of the
# hidden layer below or the input of the DBN if you are on
# the first layer
if i == 0:
input_size = n_ins
layer_input = self.x
else:
input_size = hidden_layers_sizes[i - 1]
layer_input = self.layers[-1].output
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=activation)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
# the parameters of the sigmoid_layers are parameters of the DBN.
# The visible biases in the RBM are parameters of those RBMs,
# but not of the DBN.
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
# Construct an RBM that shared weights with this layer
# the first layer could be Gaussian-Bernoulli RBM
# other layers are Bernoulli-Bernoulli RBMs
if i == 0 and first_layer_gb:
rbm_layer = GBRBM(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b,
activation=activation)
else:
rbm_layer = RBM(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b,
activation=activation)
self.rbm_layers.append(rbm_layer)
# We now need to add a logistic layer on top of the MLP
self.logLayer = LogisticRegression(input=self.layers[-1].output,
n_in=hidden_layers_sizes[-1],
n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
# compute the cost for second phase of training, defined as the
# negative log likelihood of the logistic regression (output) layer
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
# compute the gradients with respect to the model parameters
# symbolic variable that points to the number of errors made on the
# minibatch given by self.x and self.y
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction()
self.features = self.layers[-2].output
self.features_dim = self.layers[-2].n_out
class CNN(CNNBase):
""" Instantiation of Convolution neural network ... """
def __init__(self, numpy_rng, theano_rng, batch_size, n_outs,conv_layer_configs, hidden_layer_configs,
use_fast=False,hidden_activation = T.nnet.sigmoid,l1_reg=None,l2_reg=None,max_col_norm=None):
super(CNN, self).__init__(conv_layer_configs, hidden_layer_configs,l1_reg,l2_reg,max_col_norm)
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
for i in xrange(self.conv_layer_num): # construct the convolution layer
if i == 0: #is_input layer
input = self.x
else:
input = self.layers[-1].output #output of previous layer
config = conv_layer_configs[i]
conv_activation= parse_activation(config['activation'])
conv_layer = ConvLayer(numpy_rng=numpy_rng, input=input,input_shape=config['input_shape'],
filter_shape=config['filter_shape'],poolsize=config['poolsize'],
activation = conv_activation, use_fast = use_fast)
self.layers.append(conv_layer)
self.conv_layers.append(conv_layer)
if config['update']==True: # only few layers of convolution layer are considered for updation
self.params.extend(conv_layer.params)
self.delta_params.extend(conv_layer.delta_params)
hidden_layers = hidden_layer_configs['hidden_layers'];
self.conv_output_dim = config['output_shape'][1] * config['output_shape'][2] * config['output_shape'][3]
adv_activation_configs = hidden_layer_configs['adv_activation']
#flattening the last convolution output layer
self.features = self.conv_layers[-1].output.flatten(2);
self.features_dim = self.conv_output_dim;
for i in xrange(self.hidden_layer_num): # construct the hidden layer
if i == 0: # is first sigmoidla layer
input_size = self.conv_output_dim
layer_input = self.features
else:
input_size = hidden_layers[i - 1] # number of hidden neurons in previous layers
layer_input = self.layers[-1].output
if adv_activation_configs is None:
sigmoid_layer = HiddenLayer(rng=numpy_rng, input=layer_input,n_in=input_size,
n_out = hidden_layers[i], activation=hidden_activation);
else:
sigmoid_layer = HiddenLayer(rng=numpy_rng, input=layer_input,n_in=input_size,
n_out = hidden_layers[i]*adv_activation_configs['pool_size'], activation=hidden_activation,
adv_activation_method = adv_activation_configs['method'],
pool_size = adv_activation_configs['pool_size'],
pnorm_order = adv_activation_configs['pnorm_order']);
self.layers.append(sigmoid_layer)
self.mlp_layers.append(sigmoid_layer)
if config['update']==True: # only few layers of hidden layer are considered for updation
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
self.logLayer = LogisticRegression(input=self.layers[-1].output,n_in=hidden_layers[-1],n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction()
#regularization
if self.l1_reg is not None:
self.__l1Regularization__(self.hidden_layer_num*2);
if self.l2_reg is not None:
self.__l2Regularization__(self.hidden_layer_num*2);
def save_mlp2dict(self,withfinal=True,max_layer_num=-1):
if max_layer_num == -1:
max_layer_num = self.hidden_layer_num
mlp_dict = {}
for i in range(max_layer_num):
dict_a = str(i) +' W'
mlp_dict[dict_a] = _array2string(self.mlp_layers[i].params[0].get_value())
dict_a = str(i) + ' b'
mlp_dict[dict_a] = _array2string(self.mlp_layers[i].params[1].get_value())
if withfinal:
dict_a = 'logreg W'
mlp_dict[dict_a] = _array2string(self.logLayer.params[0].get_value())
dict_a = 'logreg b'
mlp_dict[dict_a] = _array2string(self.logLayer.params[1].get_value())
return mlp_dict
def __init__(self,
numpy_rng,
theano_rng,
batch_size,
n_outs,
conv_layer_configs,
hidden_layer_configs,
hidden_activation=T.nnet.sigmoid,
l1_reg=None,
l2_reg=None,
max_col_norm=None):
super(CNN3D, self).__init__(conv_layer_configs, hidden_layer_configs,
l1_reg, l2_reg, max_col_norm)
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2**30))
for i in xrange(
self.conv_layer_num): # construct the convolution layer
if i == 0: #is_input layer
input = self.x
else:
input = self.layers[-1].output #output of previous layer
config = conv_layer_configs[i]
conv_activation = parse_activation(config['activation'])
conv_layer = ConvLayer(numpy_rng=numpy_rng,
input=input,
input_shape=config['input_shape'],
filter_shape=config['filter_shape'],
poolsize=config['poolsize'],
activation=conv_activation)
self.layers.append(conv_layer)
self.conv_layers.append(conv_layer)
if config[
'update'] == True: # only few layers of convolution layer are considered for updation
self.params.extend(conv_layer.params)
self.delta_params.extend(conv_layer.delta_params)
hidden_layers = hidden_layer_configs['hidden_layers']
self.conv_output_dim = numpy.prod(config['output_shape'][1:])
adv_activation_configs = hidden_layer_configs['adv_activation']
#flattening the last convolution output layer
self.features = self.conv_layers[-1].output.flatten(2)
self.features_dim = self.conv_output_dim
for i in xrange(self.hidden_layer_num): # construct the hidden layer
if i == 0: # is first sigmoidla layer
input_size = self.conv_output_dim
layer_input = self.features
else:
input_size = hidden_layers[
i - 1] # number of hidden neurons in previous layers
layer_input = self.layers[-1].output
if adv_activation_configs is None:
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers[i],
activation=hidden_activation)
else:
sigmoid_layer = HiddenLayer(
rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers[i] *
adv_activation_configs['pool_size'],
activation=hidden_activation,
adv_activation_method=adv_activation_configs['method'],
pool_size=adv_activation_configs['pool_size'],
pnorm_order=adv_activation_configs['pnorm_order'])
self.layers.append(sigmoid_layer)
self.mlp_layers.append(sigmoid_layer)
if config[
'update'] == True: # only few layers of hidden layer are considered for updation
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
self.logLayer = LogisticRegression(input=self.layers[-1].output,
n_in=hidden_layers[-1],
n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction()
#regularization
if self.l1_reg is not None:
self.__l1Regularization__(self.hidden_layer_num * 2)
if self.l2_reg is not None:
self.__l2Regularization__(self.hidden_layer_num * 2)
class DBN(nnet):
"""Deep Belief Network
A deep belief network is obtained by stacking several RBMs on top of each
other. The hidden layer of the RBM at layer `i` becomes the input of the
RBM at layer `i+1`. The first layer RBM gets as input the input of the
network, and the hidden layer of the last RBM represents the output. When
used for classification, the DBN is treated as a MLP, by adding a logistic
regression layer on top.
"""
def __init__(self,
numpy_rng,
theano_rng=None,
n_ins=784,
hidden_layers_sizes=[500, 500],
n_outs=10,
first_layer_gb=True,
pretrainedLayers=None,
activation=T.nnet.sigmoid):
"""This class is made to support a variable number of layers.
:type numpy_rng: numpy.random.RandomState
:param numpy_rng: numpy random number generator used to draw initial
weights
:type theano_rng: theano.tensor.shared_randomstreams.RandomStreams
:param theano_rng: Theano random generator; if None is given one is
generated based on a seed drawn from `rng`
:type n_ins: int
:param n_ins: dimension of the input to the DBN
:type hidden_layers_sizes: list of ints
:param hidden_layers_sizes: intermediate layers size, must contain
at least one value
:type n_outs: int
:param n_outs: dimension of the output of the network
:type first_layer_gb: bool
:param first_layer_gb: wether first layer is gausian-bernolli or
bernolli-bernolli
"""
super(DBN, self).__init__()
self.layers = []
self.rbm_layers = []
self.n_layers = len(hidden_layers_sizes)
if pretrainedLayers == None:
self.nPreTrainLayers = n_layers
else:
self.nPreTrainLayers = pretrainedLayers
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2**30))
# allocate symbolic variables for the data
self.x = T.matrix('x') # the data is presented as rasterized images
self.y = T.ivector('y') # the labels are presented as 1D vector
# of [int] labels
# The DBN is an MLP, for which all weights of intermediate
# layers are shared with a different RBM. We will first
# construct the DBN as a deep multilayer perceptron, and when
# constructing each sigmoidal layer we also construct an RBM
# that shares weights with that layer. During pretraining we
# will train these RBMs (which will lead to chainging the
# weights of the MLP as well) During finetuning we will finish
# training the DBN by doing stochastic gradient descent on the
# MLP.
for i in xrange(self.n_layers):
# construct the sigmoidal layer
# the size of the input is either the number of hidden
# units of the layer below or the input size if we are on
# the first layer
# the input to this layer is either the activation of the
# hidden layer below or the input of the DBN if you are on
# the first layer
if i == 0:
input_size = n_ins
layer_input = self.x
else:
input_size = hidden_layers_sizes[i - 1]
layer_input = self.layers[-1].output
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=activation)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
# the parameters of the sigmoid_layers are parameters of the DBN.
# The visible biases in the RBM are parameters of those RBMs,
# but not of the DBN.
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
# Construct an RBM that shared weights with this layer
# the first layer could be Gaussian-Bernoulli RBM
# other layers are Bernoulli-Bernoulli RBMs
if i == 0 and first_layer_gb:
rbm_layer = GBRBM(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b,
activation=activation)
else:
rbm_layer = RBM(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b,
activation=activation)
self.rbm_layers.append(rbm_layer)
# We now need to add a logistic layer on top of the MLP
self.logLayer = LogisticRegression(input=self.layers[-1].output,
n_in=hidden_layers_sizes[-1],
n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
# compute the cost for second phase of training, defined as the
# negative log likelihood of the logistic regression (output) layer
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
# compute the gradients with respect to the model parameters
# symbolic variable that points to the number of errors made on the
# minibatch given by self.x and self.y
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction()
self.features = self.layers[-2].output
self.features_dim = self.layers[-2].n_out
def pretraining_functions(self, train_set_x, batch_size, weight_cost):
'''Generates a list of functions, for performing one step of
gradient descent at a given layer. The function will require
as input the minibatch index, and to train an RBM you just
need to iterate, calling the corresponding function on all
minibatch indexes.
:type train_set_x: theano.tensor.TensorType
:param train_set_x: Shared var. that contains all datapoints used
for training the RBM
:type batch_size: int
:param batch_size: size of a [mini]batch
:param weight_cost: weigth cost
'''
# index to a [mini]batch
index = T.lscalar('index') # index to a minibatch
momentum = T.scalar('momentum')
learning_rate = T.scalar('lr') # learning rate to use
# number of batches
n_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
# begining of a batch, given `index`
batch_begin = index * batch_size
# ending of a batch given `index`
batch_end = batch_begin + batch_size
pretrain_fns = []
for rbm in self.rbm_layers:
# get the cost and the updates list
# using CD-k here (persisent=None,k=1) for training each RBM.
r_cost, fe_cost, updates = rbm.get_cost_updates(
batch_size, learning_rate, momentum, weight_cost)
# compile the theano function
fn = theano.function(
inputs=[
index,
theano.Param(learning_rate, default=0.0001),
theano.Param(momentum, default=0.5)
],
outputs=[r_cost, fe_cost],
updates=updates,
givens={self.x: train_set_x[batch_begin:batch_end]})
# append function to the list of functions
pretrain_fns.append(fn)
return pretrain_fns
def __init__(self, numpy_rng, theano_rng=None, n_ins=784,
hidden_layers_sizes=[500, 500], n_outs=10,
first_layer_gb = True,pretrainedLayers=None,activation=T.nnet.sigmoid):
"""This class is made to support a variable number of layers.
:type numpy_rng: numpy.random.RandomState
:param numpy_rng: numpy random number generator used to draw initial
weights
:type theano_rng: theano.tensor.shared_randomstreams.RandomStreams
:param theano_rng: Theano random generator; if None is given one is
generated based on a seed drawn from `rng`
:type n_ins: int
:param n_ins: dimension of the input to the DBN
:type hidden_layers_sizes: list of ints
:param hidden_layers_sizes: intermediate layers size, must contain
at least one value
:type n_outs: int
:param n_outs: dimension of the output of the network
:type first_layer_gb: bool
:param first_layer_gb: wether first layer is gausian-bernolli or
bernolli-bernolli
"""
super(DBN, self).__init__()
self.layers = []
self.rbm_layers = []
self.n_layers = len(hidden_layers_sizes)
if pretrainedLayers == None:
self.nPreTrainLayers = n_layers
else :
self.nPreTrainLayers = pretrainedLayers
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# allocate symbolic variables for the data
self.x = T.matrix('x') # the data is presented as rasterized images
self.y = T.ivector('y') # the labels are presented as 1D vector
# of [int] labels
# The DBN is an MLP, for which all weights of intermediate
# layers are shared with a different RBM. We will first
# construct the DBN as a deep multilayer perceptron, and when
# constructing each sigmoidal layer we also construct an RBM
# that shares weights with that layer. During pretraining we
# will train these RBMs (which will lead to chainging the
# weights of the MLP as well) During finetuning we will finish
# training the DBN by doing stochastic gradient descent on the
# MLP.
for i in xrange(self.n_layers):
# construct the sigmoidal layer
# the size of the input is either the number of hidden
# units of the layer below or the input size if we are on
# the first layer
# the input to this layer is either the activation of the
# hidden layer below or the input of the DBN if you are on
# the first layer
if i == 0:
input_size = n_ins
layer_input = self.x
else:
input_size = hidden_layers_sizes[i - 1]
layer_input = self.layers[-1].output
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=activation)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
# the parameters of the sigmoid_layers are parameters of the DBN.
# The visible biases in the RBM are parameters of those RBMs,
# but not of the DBN.
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
# Construct an RBM that shared weights with this layer
# the first layer could be Gaussian-Bernoulli RBM
# other layers are Bernoulli-Bernoulli RBMs
if i == 0 and first_layer_gb:
rbm_layer = GBRBM(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b,
activation=activation)
else:
rbm_layer = RBM(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b,
activation=activation)
self.rbm_layers.append(rbm_layer)
# We now need to add a logistic layer on top of the MLP
self.logLayer = LogisticRegression(
input=self.layers[-1].output,
n_in=hidden_layers_sizes[-1],
n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
# compute the cost for second phase of training, defined as the
# negative log likelihood of the logistic regression (output) layer
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
# compute the gradients with respect to the model parameters
# symbolic variable that points to the number of errors made on the
# minibatch given by self.x and self.y
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction();
self.features = self.layers[-2].output;
self.features_dim = self.layers[-2].n_out
class DBN(nnet):
"""Deep Belief Network
A deep belief network is obtained by stacking several RBMs on top of each
other. The hidden layer of the RBM at layer `i` becomes the input of the
RBM at layer `i+1`. The first layer RBM gets as input the input of the
network, and the hidden layer of the last RBM represents the output. When
used for classification, the DBN is treated as a MLP, by adding a logistic
regression layer on top.
"""
def __init__(self, numpy_rng, theano_rng=None, n_ins=784,
hidden_layers_sizes=[500, 500], n_outs=10,
first_layer_gb = True,pretrainedLayers=None,activation=T.nnet.sigmoid):
"""This class is made to support a variable number of layers.
:type numpy_rng: numpy.random.RandomState
:param numpy_rng: numpy random number generator used to draw initial
weights
:type theano_rng: theano.tensor.shared_randomstreams.RandomStreams
:param theano_rng: Theano random generator; if None is given one is
generated based on a seed drawn from `rng`
:type n_ins: int
:param n_ins: dimension of the input to the DBN
:type hidden_layers_sizes: list of ints
:param hidden_layers_sizes: intermediate layers size, must contain
at least one value
:type n_outs: int
:param n_outs: dimension of the output of the network
:type first_layer_gb: bool
:param first_layer_gb: wether first layer is gausian-bernolli or
bernolli-bernolli
"""
super(DBN, self).__init__()
self.layers = []
self.rbm_layers = []
self.n_layers = len(hidden_layers_sizes)
if pretrainedLayers == None:
self.nPreTrainLayers = n_layers
else :
self.nPreTrainLayers = pretrainedLayers
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# allocate symbolic variables for the data
self.x = T.matrix('x') # the data is presented as rasterized images
self.y = T.ivector('y') # the labels are presented as 1D vector
# of [int] labels
# The DBN is an MLP, for which all weights of intermediate
# layers are shared with a different RBM. We will first
# construct the DBN as a deep multilayer perceptron, and when
# constructing each sigmoidal layer we also construct an RBM
# that shares weights with that layer. During pretraining we
# will train these RBMs (which will lead to chainging the
# weights of the MLP as well) During finetuning we will finish
# training the DBN by doing stochastic gradient descent on the
# MLP.
for i in xrange(self.n_layers):
# construct the sigmoidal layer
# the size of the input is either the number of hidden
# units of the layer below or the input size if we are on
# the first layer
# the input to this layer is either the activation of the
# hidden layer below or the input of the DBN if you are on
# the first layer
if i == 0:
input_size = n_ins
layer_input = self.x
else:
input_size = hidden_layers_sizes[i - 1]
layer_input = self.layers[-1].output
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=activation)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
# the parameters of the sigmoid_layers are parameters of the DBN.
# The visible biases in the RBM are parameters of those RBMs,
# but not of the DBN.
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
# Construct an RBM that shared weights with this layer
# the first layer could be Gaussian-Bernoulli RBM
# other layers are Bernoulli-Bernoulli RBMs
if i == 0 and first_layer_gb:
rbm_layer = GBRBM(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b,
activation=activation)
else:
rbm_layer = RBM(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b,
activation=activation)
self.rbm_layers.append(rbm_layer)
# We now need to add a logistic layer on top of the MLP
self.logLayer = LogisticRegression(
input=self.layers[-1].output,
n_in=hidden_layers_sizes[-1],
n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
# compute the cost for second phase of training, defined as the
# negative log likelihood of the logistic regression (output) layer
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
# compute the gradients with respect to the model parameters
# symbolic variable that points to the number of errors made on the
# minibatch given by self.x and self.y
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction();
self.features = self.layers[-2].output;
self.features_dim = self.layers[-2].n_out
def pretraining_functions(self, train_set_x, batch_size, weight_cost):
'''Generates a list of functions, for performing one step of
gradient descent at a given layer. The function will require
as input the minibatch index, and to train an RBM you just
need to iterate, calling the corresponding function on all
minibatch indexes.
:type train_set_x: theano.tensor.TensorType
:param train_set_x: Shared var. that contains all datapoints used
for training the RBM
:type batch_size: int
:param batch_size: size of a [mini]batch
:param weight_cost: weigth cost
'''
# index to a [mini]batch
index = T.lscalar('index') # index to a minibatch
momentum = T.scalar('momentum')
learning_rate = T.scalar('lr') # learning rate to use
# number of batches
n_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
# begining of a batch, given `index`
batch_begin = index * batch_size
# ending of a batch given `index`
batch_end = batch_begin + batch_size
pretrain_fns = []
for rbm in self.rbm_layers:
# get the cost and the updates list
# using CD-k here (persisent=None,k=1) for training each RBM.
r_cost, fe_cost, updates = rbm.get_cost_updates(batch_size, learning_rate,
momentum, weight_cost)
# compile the theano function
fn = theano.function(inputs=[index,
theano.Param(learning_rate, default=0.0001),
theano.Param(momentum, default=0.5)],
outputs= [r_cost, fe_cost],
updates=updates,
givens={self.x: train_set_x[batch_begin:batch_end]})
# append function to the list of functions
pretrain_fns.append(fn)
return pretrain_fns
def __init__(self, numpy_rng, theano_rng=None, n_ins=784,
hidden_layers_sizes=[500, 500], n_outs=10,
corruption_levels=[0.1, 0.1],activation=T.nnet.sigmoid):
""" This class is made to support a variable number of layers.
:type numpy_rng: numpy.random.RandomState
:param numpy_rng: numpy random number generator used to draw initial
weights
:type theano_rng: theano.tensor.shared_randomstreams.RandomStreams
:param theano_rng: Theano random generator; if None is given one is
generated based on a seed drawn from `rng`
:type n_ins: int
:param n_ins: dimension of the input to the sdA
:type n_layers_sizes: list of ints
:param n_layers_sizes: intermediate layers size, must contain
at least one value
:type n_outs: int
:param n_outs: dimension of the output of the network
:type corruption_levels: list of float
:param corruption_levels: amount of corruption to use for each
layer
"""
super(SDA, self).__init__()
self.layers = []
self.dA_layers = []
self.n_layers = len(hidden_layers_sizes)
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# allocate symbolic variables for the data
self.x = T.matrix('x') # the data is presented as rasterized images
self.y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
# The SdA is an MLP, for which all weights of intermediate layers
# are shared with a different denoising autoencoders
# We will first construct the SdA as a deep multilayer perceptron,
# and when constructing each sigmoidal layer we also construct a
# denoising autoencoder that shares weights with that layer
# During pretraining we will train these autoencoders (which will
# lead to chainging the weights of the MLP as well)
# During finetunining we will finish training the SdA by doing
# stochastich gradient descent on the MLP
for i in xrange(self.n_layers):
# construct the sigmoidal layer
# the size of the input is either the number of hidden units of
# the layer below or the input size if we are on the first layer
if i == 0:
input_size = n_ins
else:
input_size = hidden_layers_sizes[i - 1]
# the input to this layer is either the activation of the hidden
# layer below or the input of the SdA if you are on the first
# layer
if i == 0:
layer_input = self.x
else:
layer_input = self.layers[-1].output
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=T.nnet.sigmoid)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
# its arguably a philosophical question...
# but we are going to only declare that the parameters of the
# sigmoid_layers are parameters of the StackedDAA
# the visible biases in the dA are parameters of those
# dA, but not the SdA
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
# Construct a denoising autoencoder that shared weights with this
# layer
dA_layer = dA(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
bhid=sigmoid_layer.b,
activation=T.nnet.sigmoid)
self.dA_layers.append(dA_layer)
# We now need to add a logistic layer on top of the MLP
self.logLayer = LogisticRegression(
input=self.layers[-1].output,
n_in=hidden_layers_sizes[-1], n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
# construct a function that implements one step of finetunining
# compute the cost for second phase of training,
# defined as the negative log likelihood
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
# compute the gradients with respect to the model parameters
# symbolic variable that points to the number of errors made on the
# minibatch given by self.x and self.y
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction();
self.features = self.layers[-2].output;
self.features_dim = self.layers[-2].n_out
class SDA(nnet):
"""Stacked denoising auto-encoder class (SdA)
A stacked denoising autoencoder model is obtained by stacking several
dAs. The hidden layer of the dA at layer `i` becomes the input of
the dA at layer `i+1`. The first layer dA gets as input the input of
the SdA, and the hidden layer of the last dA represents the output.
Note that after pretraining, the SdA is dealt with as a normal MLP,
the dAs are only used to initialize the weights.
"""
def __init__(self, numpy_rng, theano_rng=None, n_ins=784,
hidden_layers_sizes=[500, 500], n_outs=10,
corruption_levels=[0.1, 0.1],activation=T.nnet.sigmoid):
""" This class is made to support a variable number of layers.
:type numpy_rng: numpy.random.RandomState
:param numpy_rng: numpy random number generator used to draw initial
weights
:type theano_rng: theano.tensor.shared_randomstreams.RandomStreams
:param theano_rng: Theano random generator; if None is given one is
generated based on a seed drawn from `rng`
:type n_ins: int
:param n_ins: dimension of the input to the sdA
:type n_layers_sizes: list of ints
:param n_layers_sizes: intermediate layers size, must contain
at least one value
:type n_outs: int
:param n_outs: dimension of the output of the network
:type corruption_levels: list of float
:param corruption_levels: amount of corruption to use for each
layer
"""
super(SDA, self).__init__()
self.layers = []
self.dA_layers = []
self.n_layers = len(hidden_layers_sizes)
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# allocate symbolic variables for the data
self.x = T.matrix('x') # the data is presented as rasterized images
self.y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
# The SdA is an MLP, for which all weights of intermediate layers
# are shared with a different denoising autoencoders
# We will first construct the SdA as a deep multilayer perceptron,
# and when constructing each sigmoidal layer we also construct a
# denoising autoencoder that shares weights with that layer
# During pretraining we will train these autoencoders (which will
# lead to chainging the weights of the MLP as well)
# During finetunining we will finish training the SdA by doing
# stochastich gradient descent on the MLP
for i in xrange(self.n_layers):
# construct the sigmoidal layer
# the size of the input is either the number of hidden units of
# the layer below or the input size if we are on the first layer
if i == 0:
input_size = n_ins
else:
input_size = hidden_layers_sizes[i - 1]
# the input to this layer is either the activation of the hidden
# layer below or the input of the SdA if you are on the first
# layer
if i == 0:
layer_input = self.x
else:
layer_input = self.layers[-1].output
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=T.nnet.sigmoid)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
# its arguably a philosophical question...
# but we are going to only declare that the parameters of the
# sigmoid_layers are parameters of the StackedDAA
# the visible biases in the dA are parameters of those
# dA, but not the SdA
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
# Construct a denoising autoencoder that shared weights with this
# layer
dA_layer = dA(numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
bhid=sigmoid_layer.b,
activation=T.nnet.sigmoid)
self.dA_layers.append(dA_layer)
# We now need to add a logistic layer on top of the MLP
self.logLayer = LogisticRegression(
input=self.layers[-1].output,
n_in=hidden_layers_sizes[-1], n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
# construct a function that implements one step of finetunining
# compute the cost for second phase of training,
# defined as the negative log likelihood
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
# compute the gradients with respect to the model parameters
# symbolic variable that points to the number of errors made on the
# minibatch given by self.x and self.y
self.errors = self.logLayer.errors(self.y)
self.output = self.logLayer.prediction();
self.features = self.layers[-2].output;
self.features_dim = self.layers[-2].n_out
def pretraining_functions(self, train_x, batch_size):
''' Generates a list of functions, each of them implementing one
step in trainnig the dA corresponding to the layer with same index.
The function will require as input the minibatch index, and to train
a dA you just need to iterate, calling the corresponding function on
all minibatch indexes.
:type train_x: theano.tensor.TensorType
:param train_x: Shared variable that contains all datapoints used
for training the dA
:type batch_size: int
:param batch_size: size of a [mini]batch
'''
# index to a [mini]batch
index = T.lscalar('index') # index to a minibatch
corruption_level = T.scalar('corruption') # % of corruption to use
learning_rate = T.scalar('lr') # learning rate to use
# number of batches
n_batches = train_x.get_value(borrow=True).shape[0] / batch_size
# begining of a batch, given `index`
batch_begin = index * batch_size
# ending of a batch given `index`
batch_end = batch_begin + batch_size
pretrain_fns = []
for dA in self.dA_layers:
# get the cost and the updates list
cost, updates = dA.get_cost_updates(corruption_level,
learning_rate)
# compile the theano function
fn = theano.function(inputs=[index,
theano.Param(corruption_level, default=0.2),
theano.Param(learning_rate, default=0.1)],
outputs=cost,
updates=updates,
givens={self.x: train_x[batch_begin:
batch_end]})
# append `fn` to the list of functions
pretrain_fns.append(fn)
return pretrain_fns
def __init__(self, numpy_rng, theano_rng=None, n_ins=784,
hidden_layers_sizes=[500, 500], n_outs=10,
activation = T.nnet.sigmoid, adv_activation = None,
max_col_norm = None, l1_reg = None, l2_reg = None):
super(DNN, self).__init__()
self.layers = []
self.n_layers = len(hidden_layers_sizes)
self.max_col_norm = max_col_norm
self.l1_reg = l1_reg
self.l2_reg = l2_reg
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# allocate symbolic variables for the data
self.x = T.matrix('x')
self.y = T.ivector('y')
for i in xrange(self.n_layers):
# construct the sigmoidal layer
if i == 0:
input_size = n_ins
layer_input = self.x
else:
input_size = hidden_layers_sizes[i - 1]
layer_input = self.layers[-1].output
if not adv_activation is None:
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i] * pool_size,
activation = activation,
adv_activation_method = adv_activation['method'],
pool_size = adv_activation['pool_size'],
pnorm_order = adv_activation['pnorm_order'])
else:
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=activation)
# add the layer to our list of layers
self.layers.append(sigmoid_layer)
self.params.extend(sigmoid_layer.params)
self.delta_params.extend(sigmoid_layer.delta_params)
# We now need to add a logistic layer on top of the MLP
self.logLayer = LogisticRegression(
input=self.layers[-1].output,
n_in=hidden_layers_sizes[-1], n_out=n_outs)
self.layers.append(self.logLayer)
self.params.extend(self.logLayer.params)
self.delta_params.extend(self.logLayer.delta_params)
# construct a function that implements one step of finetunining
# compute the cost for second phase of training,
# defined as the negative log likelihood
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
self.errors = self.logLayer.errors(self.y)
if self.l1_reg is not None:
self.__l1Regularization__();
if self.l2_reg is not None:
self.__l2Regularization__();
self.output = self.logLayer.prediction();
self.features = self.layers[-2].output;
self.features_dim = self.layers[-2].n_out
