def __init__(self,
model_name='lr',
main_dir='lr/',
models_dir='models/',
data_dir='data/',
summary_dir='logs/',
loss_func='cross_entropy',
dataset='mnist',
learning_rate=0.01,
verbose=0,
num_epochs=10,
batch_size=10):
"""Constructor.
:param verbose: Level of verbosity. 0 - silent, 1 - print accuracy.
"""
SupervisedModel.__init__(self, model_name, main_dir, models_dir,
data_dir, summary_dir)
self._initialize_training_parameters(loss_func, learning_rate,
num_epochs, batch_size, dataset,
None)
self.verbose = verbose
# Computational graph nodes
self.input_data = None
self.input_labels = None
self.W_ = None
self.b_ = None
self.accuracy = None
def __init__(self, layers, original_shape, model_name='convnet', main_dir='convnet',
models_dir='models/', data_dir='data/', summary_dir='logs/',
loss_func='softmax_cross_entropy', num_epochs=10, batch_size=10, dataset='mnist',
opt='gradient_descent', learning_rate=0.01, momentum=0.5, dropout=0.5, verbose=1):
"""
:param layers: string used to build the model.
This string is a comma-separate specification of the layers of the network.
Supported values:
conv2d-FX-FY-Z-S: 2d convolution with Z feature maps as output and FX x FY filters. S is the strides size
maxpool-X: max pooling on the previous layer. X is the size of the max pooling
full-X: fully connected layer with X units
softmax: softmax layer
For example:
conv2d-5-5-32,maxpool-2,conv2d-5-5-64,maxpool-2,full-128,full-128,softmax
:param original_shape: original shape of the images in the dataset
:param dropout: Dropout parameter
:param verbose: Level of verbosity. 0 - silent, 1 - print accuracy.
"""
SupervisedModel.__init__(self, model_name, main_dir, models_dir, data_dir, summary_dir)
self._initialize_training_parameters(loss_func, learning_rate, num_epochs, batch_size,
dataset, opt, momentum)
self.layers = layers
self.original_shape = original_shape
self.dropout = dropout
self.verbose = verbose
self.W_vars = None
self.B_vars = None
self.accuracy = None
def __init__(
self, model_name='lr', main_dir='lr/', models_dir='models/',
data_dir='data/', summary_dir='logs/', loss_func='cross_entropy',
dataset='mnist', learning_rate=0.01, verbose=0, num_epochs=10,
batch_size=10):
"""Constructor.
:param verbose: Level of verbosity. 0 - silent, 1 - print accuracy.
"""
SupervisedModel.__init__(
self, model_name, main_dir, models_dir, data_dir, summary_dir)
self._initialize_training_parameters(
loss_func, learning_rate, num_epochs, batch_size, dataset, None)
self.verbose = verbose
# Computational graph nodes
self.input_data = None
self.input_labels = None
self.W_ = None
self.b_ = None
self.accuracy = None
def __init__(self,
layers,
original_shape,
model_name='convnet',
main_dir='convnet',
models_dir='models/',
data_dir='data/',
summary_dir='logs/',
loss_func='softmax_cross_entropy',
num_epochs=10,
batch_size=10,
dataset='mnist',
opt='gradient_descent',
learning_rate=0.01,
momentum=0.5,
dropout=0.5,
verbose=1):
"""Constructor.
:param layers: string used to build the model.
This string is a comma-separate specification of the layers.
Supported values:
conv2d-FX-FY-Z-S: 2d convolution with Z feature maps as output
and FX x FY filters. S is the strides size
maxpool-X: max pooling on the previous layer. X is the size of
the max pooling
full-X: fully connected layer with X units
softmax: softmax layer
For example:
conv2d-5-5-32,maxpool-2,conv2d-5-5-64,maxpool-2,full-128,full-128,softmax
:param original_shape: original shape of the images in the dataset
:param dropout: Dropout parameter
:param verbose: Level of verbosity. 0 - silent, 1 - print accuracy.
"""
SupervisedModel.__init__(self, model_name, main_dir, models_dir,
data_dir, summary_dir)
self._initialize_training_parameters(loss_func, learning_rate,
num_epochs, batch_size, dataset,
opt, momentum)
self.layers = layers
self.original_shape = original_shape
self.dropout = dropout
self.verbose = verbose
self.W_vars = None
self.B_vars = None
self.accuracy = None
def pretrain(self, train_set, validation_set=None):
self.do_pretrain = True
def set_params_func(rbmmachine, rbmgraph):
params = rbmmachine.get_model_parameters(graph=rbmgraph)
self.encoding_w_.append(params['W'])
self.encoding_b_.append(params['bh_'])
return SupervisedModel.pretrain_procedure(self, self.rbms, self.rbm_graphs, set_params_func=set_params_func,
train_set=train_set, validation_set=validation_set)
def pretrain(self, train_set, validation_set=None):
self.do_pretrain = True
def set_params_func(autoenc, autoencgraph):
params = autoenc.get_model_parameters(graph=autoencgraph)
self.encoding_w_.append(params['enc_w'])
self.encoding_b_.append(params['enc_b'])
return SupervisedModel.pretrain_procedure(self, self.autoencoders, self.autoencoder_graphs,
set_params_func=set_params_func, train_set=train_set,
validation_set=validation_set)
def pretrain(self, train_set, validation_set=None):
"""Perform Unsupervised pretraining of the autoencoder."""
self.do_pretrain = True
def set_params_func(autoenc, autoencgraph):
params = autoenc.get_model_parameters(graph=autoencgraph)
self.encoding_w_.append(params['enc_w'])
self.encoding_b_.append(params['enc_b'])
return SupervisedModel.pretrain_procedure(
self, self.autoencoders, self.autoencoder_graphs,
set_params_func=set_params_func, train_set=train_set,
validation_set=validation_set)
def pretrain(self, train_set, validation_set=None):
self.do_pretrain = True
def set_params_func(rbmmachine, rbmgraph):
params = rbmmachine.get_model_parameters(graph=rbmgraph)
self.encoding_w_.append(params['W'])
self.encoding_b_.append(params['bh_'])
return SupervisedModel.pretrain_procedure(
self,
self.rbms,
self.rbm_graphs,
set_params_func=set_params_func,
train_set=train_set,
validation_set=validation_set)
def pretrain(self, train_set, validation_set=None):
"""Perform Unsupervised pretraining of the autoencoder."""
self.do_pretrain = True
def set_params_func(autoenc, autoencgraph):
params = autoenc.get_model_parameters(graph=autoencgraph)
self.encoding_w_.append(params["enc_w"])
self.encoding_b_.append(params["enc_b"])
return SupervisedModel.pretrain_procedure(
self,
self.autoencoders,
self.autoencoder_graphs,
set_params_func=set_params_func,
train_set=train_set,
validation_set=validation_set,
)
def __init__(self, layers, model_name='sdae', main_dir='sdae/', models_dir='models/',
data_dir='data/', summary_dir='logs/',
enc_act_func=[tf.nn.tanh], dec_act_func=[None], loss_func=['cross_entropy'], num_epochs=[10],
batch_size=[10], dataset='mnist', opt=['gradient_descent'], l2reg=[5e-4],
learning_rate=[0.01], momentum=0.5, finetune_dropout=1, corr_type=['none'],
corr_frac=[0.], verbose=1, finetune_loss_func='softmax_cross_entropy', finetune_act_func=tf.nn.relu,
finetune_opt='gradient_descent', finetune_learning_rate=0.001, finetune_num_epochs=10,
finetune_batch_size=20, do_pretrain=False):
"""
:param layers: list containing the hidden units for each layer
:param enc_act_func: Activation function for the encoder. [tf.nn.tanh, tf.nn.sigmoid]
:param dec_act_func: Activation function for the decoder. [[tf.nn.tanh, tf.nn.sigmoid, None]
:param finetune_loss_func: Loss function for the softmax layer. string, default ['softmax_cross_entropy', 'mean_squared']
:param finetune_dropout: dropout parameter
:param finetune_learning_rate: learning rate for the finetuning. float, default 0.001
:param finetune_act_func: activation function for the finetuning phase
:param finetune_opt: optimizer for the finetuning phase
:param finetune_num_epochs: Number of epochs for the finetuning. int, default 20
:param finetune_batch_size: Size of each mini-batch for the finetuning. int, default 20
:param corr_type: Type of input corruption. string, default 'none'. ["none", "masking", "salt_and_pepper"]
:param corr_frac: Fraction of the input to corrupt. float, default 0.0
:param verbose: Level of verbosity. 0 - silent, 1 - print accuracy. int, default 0
:param do_pretrain: True: uses variables from pretraining, False: initialize new variables.
"""
# WARNING! This must be the first expression in the function or else it will send other variables to expanded_args()
# This function takes all the passed parameters that are lists and expands them to the number of layers, if the number
# of layers is more than the list of the parameter.
expanded_args = utilities.expand_args(**locals())
SupervisedModel.__init__(self, model_name, main_dir, models_dir, data_dir, summary_dir)
self._initialize_training_parameters(loss_func=finetune_loss_func, learning_rate=finetune_learning_rate,
dropout=finetune_dropout, num_epochs=finetune_num_epochs,
batch_size=finetune_batch_size, dataset=dataset, opt=finetune_opt,
momentum=momentum)
self.do_pretrain = do_pretrain
self.layers = layers
self.finetune_act_func = finetune_act_func
self.verbose = verbose
# Model parameters
self.encoding_w_ = [] # list of matrices of encoding weights (one per layer)
self.encoding_b_ = [] # list of arrays of encoding biases (one per layer)
self.last_W = None
self.last_b = None
self.autoencoders = []
self.autoencoder_graphs = []
for l, layer in enumerate(layers):
dae_str = 'dae-' + str(l + 1)
self.autoencoders.append(denoising_autoencoder.DenoisingAutoencoder(
n_components=layer, main_dir=self.main_dir, model_name=self.model_name + '-' + dae_str,
models_dir=os.path.join(self.models_dir, dae_str), data_dir=os.path.join(self.data_dir, dae_str),
summary_dir=os.path.join(self.tf_summary_dir, dae_str),
enc_act_func=expanded_args['enc_act_func'][l], dec_act_func=expanded_args['dec_act_func'][l],
loss_func=expanded_args['loss_func'][l],
opt=expanded_args['opt'][l], learning_rate=expanded_args['learning_rate'][l], l2reg=expanded_args['l2reg'],
momentum=self.momentum, corr_type=expanded_args['corr_type'][l], corr_frac=expanded_args['corr_frac'][l],
verbose=self.verbose, num_epochs=expanded_args['num_epochs'][l], batch_size=expanded_args['batch_size'][l],
dataset=self.dataset))
self.autoencoder_graphs.append(tf.Graph())
def __init__(self, rbm_layers, model_name='srbm', do_pretrain=False, main_dir='srbm/', models_dir='models/', data_dir='data/',
summary_dir='logs/', rbm_num_epochs=[10], rbm_gibbs_k=[1], rbm_gauss_visible=False, rbm_stddev=0.1,
rbm_batch_size=[10], dataset='mnist', rbm_learning_rate=[0.01], momentum=0.5,
finetune_dropout=1, verbose=1, finetune_loss_func='softmax_cross_entropy', finetune_act_func=tf.nn.relu,
finetune_opt='gradient_descent', finetune_learning_rate=0.001, finetune_num_epochs=10,
finetune_batch_size=20):
"""
:param rbm_layers: list containing the hidden units for each layer
:param finetune_loss_func: Loss function for the softmax layer. string, default ['softmax_cross_entropy', 'mean_squared']
:param finetune_dropout: dropout parameter
:param finetune_learning_rate: learning rate for the finetuning. float, default 0.001
:param finetune_act_func: activation function for the finetuning phase
:param finetune_opt: optimizer for the finetuning phase
:param finetune_num_epochs: Number of epochs for the finetuning. int, default 20
:param finetune_batch_size: Size of each mini-batch for the finetuning. int, default 20
:param verbose: Level of verbosity. 0 - silent, 1 - print accuracy. int, default 0
:param do_pretrain: True: uses variables from pretraining, False: initialize new variables.
"""
SupervisedModel.__init__(self, model_name, main_dir, models_dir, data_dir, summary_dir)
self._initialize_training_parameters(loss_func=finetune_loss_func, learning_rate=finetune_learning_rate,
dropout=finetune_dropout, num_epochs=finetune_num_epochs,
batch_size=finetune_batch_size, dataset=dataset, opt=finetune_opt,
momentum=momentum)
self.do_pretrain = do_pretrain
self.layers = rbm_layers
self.finetune_act_func = finetune_act_func
self.verbose = verbose
# Model parameters
self.encoding_w_ = [] # list of matrices of encoding weights (one per layer)
self.encoding_b_ = [] # list of arrays of encoding biases (one per layer)
self.softmax_W = None
self.softmax_b = None
rbm_params = {'num_epochs': rbm_num_epochs, 'gibbs_k': rbm_gibbs_k, 'batch_size': rbm_batch_size,
'learning_rate': rbm_learning_rate}
for p in rbm_params:
if len(rbm_params[p]) != len(rbm_layers):
# The current parameter is not specified by the user, should default it for all the layers
rbm_params[p] = [rbm_params[p][0] for _ in rbm_layers]
self.rbms = []
self.rbm_graphs = []
for l, layer in enumerate(rbm_layers):
rbm_str = 'rbm-' + str(l+1)
if l == 0 and rbm_gauss_visible:
self.rbms.append(rbm.RBM(model_name=self.model_name + '-' + rbm_str,
models_dir=os.path.join(self.models_dir, rbm_str), data_dir=os.path.join(self.data_dir, rbm_str), summary_dir=os.path.join(self.tf_summary_dir, rbm_str),
num_hidden=layer, main_dir=self.main_dir, learning_rate=rbm_params['learning_rate'][l], dataset=self.dataset,
verbose=self.verbose, num_epochs=rbm_params['num_epochs'][l], batch_size=rbm_params['batch_size'][l],
gibbs_sampling_steps=rbm_params['gibbs_k'][l], visible_unit_type='gauss', stddev=rbm_stddev))
else:
self.rbms.append(rbm.RBM(model_name=self.model_name + '-' + rbm_str,
models_dir=os.path.join(self.models_dir, rbm_str), data_dir=os.path.join(self.data_dir, rbm_str), summary_dir=os.path.join(self.tf_summary_dir, rbm_str),
num_hidden=layer, main_dir=self.main_dir, learning_rate=rbm_params['learning_rate'][l], dataset=self.dataset,
verbose=self.verbose, num_epochs=rbm_params['num_epochs'][l], batch_size=rbm_params['batch_size'][l],
gibbs_sampling_steps=rbm_params['gibbs_k'][l]))
self.rbm_graphs.append(tf.Graph())
def __init__(self,
layers,
model_name='sdae',
main_dir='sdae/',
models_dir='models/',
data_dir='data/',
summary_dir='logs/',
enc_act_func=[tf.nn.tanh],
dec_act_func=[None],
loss_func=['cross_entropy'],
num_epochs=[10],
batch_size=[10],
dataset='mnist',
opt=['gradient_descent'],
l2reg=[5e-4],
learning_rate=[0.01],
momentum=0.5,
finetune_dropout=1,
corr_type=['none'],
corr_frac=[0.],
verbose=1,
finetune_loss_func='softmax_cross_entropy',
finetune_act_func=tf.nn.relu,
finetune_opt='gradient_descent',
finetune_learning_rate=0.001,
finetune_num_epochs=10,
finetune_batch_size=20,
do_pretrain=False):
"""
:param layers: list containing the hidden units for each layer
:param enc_act_func: Activation function for the encoder. [tf.nn.tanh, tf.nn.sigmoid]
:param dec_act_func: Activation function for the decoder. [[tf.nn.tanh, tf.nn.sigmoid, None]
:param finetune_loss_func: Loss function for the softmax layer. string, default ['softmax_cross_entropy', 'mean_squared']
:param finetune_dropout: dropout parameter
:param finetune_learning_rate: learning rate for the finetuning. float, default 0.001
:param finetune_act_func: activation function for the finetuning phase
:param finetune_opt: optimizer for the finetuning phase
:param finetune_num_epochs: Number of epochs for the finetuning. int, default 20
:param finetune_batch_size: Size of each mini-batch for the finetuning. int, default 20
:param corr_type: Type of input corruption. string, default 'none'. ["none", "masking", "salt_and_pepper"]
:param corr_frac: Fraction of the input to corrupt. float, default 0.0
:param verbose: Level of verbosity. 0 - silent, 1 - print accuracy. int, default 0
:param do_pretrain: True: uses variables from pretraining, False: initialize new variables.
"""
# WARNING! This must be the first expression in the function or else it will send other variables to expanded_args()
# This function takes all the passed parameters that are lists and expands them to the number of layers, if the number
# of layers is more than the list of the parameter.
expanded_args = utilities.expand_args(**locals())
SupervisedModel.__init__(self, model_name, main_dir, models_dir,
data_dir, summary_dir)
self._initialize_training_parameters(
loss_func=finetune_loss_func,
learning_rate=finetune_learning_rate,
dropout=finetune_dropout,
num_epochs=finetune_num_epochs,
batch_size=finetune_batch_size,
dataset=dataset,
opt=finetune_opt,
momentum=momentum)
self.do_pretrain = do_pretrain
self.layers = layers
self.finetune_act_func = finetune_act_func
self.verbose = verbose
# Model parameters
self.encoding_w_ = [
] # list of matrices of encoding weights (one per layer)
self.encoding_b_ = [
] # list of arrays of encoding biases (one per layer)
self.last_W = None
self.last_b = None
self.autoencoders = []
self.autoencoder_graphs = []
for l, layer in enumerate(layers):
dae_str = 'dae-' + str(l + 1)
self.autoencoders.append(
denoising_autoencoder.DenoisingAutoencoder(
n_components=layer,
main_dir=self.main_dir,
model_name=self.model_name + '-' + dae_str,
models_dir=os.path.join(self.models_dir, dae_str),
data_dir=os.path.join(self.data_dir, dae_str),
summary_dir=os.path.join(self.tf_summary_dir, dae_str),
enc_act_func=expanded_args['enc_act_func'][l],
dec_act_func=expanded_args['dec_act_func'][l],
loss_func=expanded_args['loss_func'][l],
opt=expanded_args['opt'][l],
learning_rate=expanded_args['learning_rate'][l],
l2reg=expanded_args['l2reg'],
momentum=self.momentum,
corr_type=expanded_args['corr_type'][l],
corr_frac=expanded_args['corr_frac'][l],
verbose=self.verbose,
num_epochs=expanded_args['num_epochs'][l],
batch_size=expanded_args['batch_size'][l],
dataset=self.dataset))
self.autoencoder_graphs.append(tf.Graph())
def __init__(
self,
rbm_layers,
model_name='dbn',
do_pretrain=False,
main_dir='dbn/',
models_dir='models/',
data_dir='data/',
summary_dir='logs/',
rbm_num_epochs=[10],
rbm_gibbs_k=[1],
rbm_gauss_visible=False,
rbm_stddev=0.1,
rbm_batch_size=[10],
dataset='mnist',
rbm_learning_rate=[0.01],
finetune_dropout=1,
finetune_loss_func='softmax_cross_entropy',
finetune_act_func=tf.nn.sigmoid,
finetune_opt='gradient_descent',
finetune_learning_rate=0.001,
finetune_num_epochs=10,
finetune_batch_size=20,
verbose=1,
momentum=0.5,
):
"""Constructor.
:param rbm_layers: list containing the hidden units for each layer
:param finetune_loss_func: Loss function for the softmax layer.
string, default ['softmax_cross_entropy', 'mean_squared']
:param finetune_dropout: dropout parameter
:param finetune_learning_rate: learning rate for the finetuning.
float, default 0.001
:param finetune_act_func: activation function for the finetuning phase
:param finetune_opt: optimizer for the finetuning phase
:param finetune_num_epochs: Number of epochs for the finetuning.
int, default 20
:param finetune_batch_size: Size of each mini-batch for the finetuning.
int, default 20
:param verbose: Level of verbosity. 0 - silent, 1 - print accuracy.
int, default 0
:param do_pretrain: True: uses variables from pretraining,
False: initialize new variables.
"""
SupervisedModel.__init__(self, model_name, main_dir, models_dir,
data_dir, summary_dir)
self._initialize_training_parameters(
loss_func=finetune_loss_func,
learning_rate=finetune_learning_rate,
dropout=finetune_dropout,
num_epochs=finetune_num_epochs,
batch_size=finetune_batch_size,
dataset=dataset,
opt=finetune_opt,
momentum=momentum)
self.do_pretrain = do_pretrain
self.layers = rbm_layers
self.finetune_act_func = finetune_act_func
self.verbose = verbose
# Model parameters
self.encoding_w_ = [] # list of matrices of encoding weights per layer
self.encoding_b_ = [] # list of arrays of encoding biases per layer
self.softmax_W = None
self.softmax_b = None
rbm_params = {
'num_epochs': rbm_num_epochs,
'gibbs_k': rbm_gibbs_k,
'batch_size': rbm_batch_size,
'learning_rate': rbm_learning_rate
}
for p in rbm_params:
if len(rbm_params[p]) != len(rbm_layers):
# The current parameter is not specified by the user,
# should default it for all the layers
rbm_params[p] = [rbm_params[p][0] for _ in rbm_layers]
self.rbms = []
self.rbm_graphs = []
for l, layer in enumerate(rbm_layers):
rbm_str = 'rbm-' + str(l + 1)
if l == 0 and rbm_gauss_visible:
self.rbms.append(
rbm.RBM(model_name=self.model_name + '-' + rbm_str,
models_dir=os.path.join(self.models_dir, rbm_str),
data_dir=os.path.join(self.data_dir, rbm_str),
summary_dir=os.path.join(self.tf_summary_dir,
rbm_str),
num_hidden=layer,
main_dir=self.main_dir,
learning_rate=rbm_params['learning_rate'][l],
dataset=self.dataset,
verbose=self.verbose,
num_epochs=rbm_params['num_epochs'][l],
batch_size=rbm_params['batch_size'][l],
gibbs_sampling_steps=rbm_params['gibbs_k'][l],
visible_unit_type='gauss',
stddev=rbm_stddev))
else:
self.rbms.append(
rbm.RBM(model_name=self.model_name + '-' + rbm_str,
models_dir=os.path.join(self.models_dir, rbm_str),
data_dir=os.path.join(self.data_dir, rbm_str),
summary_dir=os.path.join(self.tf_summary_dir,
rbm_str),
num_hidden=layer,
main_dir=self.main_dir,
learning_rate=rbm_params['learning_rate'][l],
dataset=self.dataset,
verbose=self.verbose,
num_epochs=rbm_params['num_epochs'][l],
batch_size=rbm_params['batch_size'][l],
gibbs_sampling_steps=rbm_params['gibbs_k'][l]))
self.rbm_graphs.append(tf.Graph())
