class LazyLearningMixin(SharedDocs):
"""
Mixin for lazy learning Neural Network algorithms.
Notes
-----
- Network uses lazy learning which mean that network doesn't
need iterative training. It just stores parameters
and use them to make a predictions.
Methods
-------
train(input_train, target_train, copy=True)
Network just stores all the information about the data and use
it for the prediction. Parameter ``copy`` copies input data
before saving it inside the network.
"""
step = WithdrawProperty()
show_epoch = WithdrawProperty()
shuffle_data = WithdrawProperty()
train_end_signal = WithdrawProperty()
epoch_end_signal = WithdrawProperty()
def __init__(self, *args, **kwargs):
self.input_train = None
self.target_train = None
super(LazyLearningMixin, self).__init__(*args, **kwargs)
def train(self, input_train, target_train):
self.input_train = input_train
self.target_train = target_train
if input_train.shape[0] != target_train.shape[0]:
raise ValueError("Number of samples in the input and target "
"datasets are different")
class ConjugateGradient(WolfeLineSearchForStep, BaseOptimizer):
"""
Conjugate Gradient algorithm.
Parameters
----------
update_function : ``fletcher_reeves``, ``polak_ribiere``,\
``hentenes_stiefel``, ``dai_yuan``, ``liu_storey``
Update function. Defaults to ``fletcher_reeves``.
epsilon : float
Ensures computational stability during the division in
``update_function`` when denominator is very small number.
Defaults to ``1e-7``.
{WolfeLineSearchForStep.Parameters}
{BaseOptimizer.network}
{BaseOptimizer.loss}
{BaseOptimizer.show_epoch}
{BaseOptimizer.shuffle_data}
{BaseOptimizer.signals}
{BaseOptimizer.verbose}
{BaseOptimizer.regularizer}
Attributes
----------
{BaseOptimizer.Attributes}
Methods
-------
{BaseOptimizer.Methods}
Examples
--------
>>> from sklearn import datasets, preprocessing
>>> from sklearn.model_selection import train_test_split
>>> from neupy import algorithms, layers
>>>
>>> dataset = datasets.load_boston()
>>> data, target = dataset.data, dataset.target
>>>
>>> data_scaler = preprocessing.MinMaxScaler()
>>> target_scaler = preprocessing.MinMaxScaler()
>>>
>>> x_train, x_test, y_train, y_test = train_test_split(
... data_scaler.fit_transform(data),
... target_scaler.fit_transform(target),
... test_size=0.15
... )
>>>
>>> cgnet = algorithms.ConjugateGradient(
... network=[
... layers.Input(13),
... layers.Sigmoid(50),
... layers.Sigmoid(1),
... ],
... update_function='fletcher_reeves',
... verbose=False
... )
>>>
>>> cgnet.train(x_train, y_train, epochs=100)
>>> y_predict = cgnet.predict(x_test).round(1)
>>>
>>> real = target_scaler.inverse_transform(y_test)
>>> predicted = target_scaler.inverse_transform(y_predict)
References
----------
[1] Jorge Nocedal, Stephen J. Wright, Numerical Optimization.
Chapter 5, Conjugate Gradient Methods, p. 101-133
"""
epsilon = NumberProperty(default=1e-7, minval=0)
update_function = ChoiceProperty(
default='fletcher_reeves',
choices={
'fletcher_reeves': fletcher_reeves,
'polak_ribiere': polak_ribiere,
'hentenes_stiefel': hentenes_stiefel,
'liu_storey': liu_storey,
'dai_yuan': dai_yuan,
}
)
step = WithdrawProperty()
def init_functions(self):
n_parameters = self.network.n_parameters
self.variables.update(
prev_delta=tf.Variable(
tf.zeros([n_parameters]),
name="conj-grad/prev-delta",
dtype=tf.float32,
),
prev_gradient=tf.Variable(
tf.zeros([n_parameters]),
name="conj-grad/prev-gradient",
dtype=tf.float32,
),
iteration=tf.Variable(
asfloat(self.last_epoch),
name='conj-grad/current-iteration',
dtype=tf.float32
),
)
super(ConjugateGradient, self).init_functions()
def init_train_updates(self):
iteration = self.variables.iteration
previous_delta = self.variables.prev_delta
previous_gradient = self.variables.prev_gradient
n_parameters = self.network.n_parameters
variables = self.network.variables
parameters = [var for var in variables.values() if var.trainable]
param_vector = make_single_vector(parameters)
gradients = tf.gradients(self.variables.loss, parameters)
full_gradient = make_single_vector(gradients)
beta = self.update_function(
previous_gradient, full_gradient, previous_delta, self.epsilon)
parameter_delta = tf.where(
tf.equal(tf.mod(iteration, n_parameters), 0),
-full_gradient,
-full_gradient + beta * previous_delta
)
step = self.find_optimal_step(param_vector, parameter_delta)
updated_parameters = param_vector + step * parameter_delta
updates = setup_parameter_updates(parameters, updated_parameters)
# We have to compute these values first, otherwise
# parallelization, in tensorflow, can mix update order
# and, for example, previous gradient can be equal to
# current gradient value. It happens because tensorflow
# try to execute operations in parallel.
with tf.control_dependencies([full_gradient, parameter_delta]):
updates.extend([
previous_gradient.assign(full_gradient),
previous_delta.assign(parameter_delta),
iteration.assign(iteration + 1),
])
return updates
class QuasiNewton(StepSelectionBuiltIn, GradientDescent):
"""
Quasi-Newton algorithm optimization.
Parameters
----------
update_function : {{'bfgs', 'dfp', 'psb', 'sr1'}}
Update function. Defaults to ``bfgs``.
h0_scale : float
Default Hessian matrix is an identity matrix. The
``h0_scale`` parameter scales identity matrix.
Defaults to ``1``.
{GradientDescent.connection}
{GradientDescent.error}
{GradientDescent.show_epoch}
{GradientDescent.shuffle_data}
{GradientDescent.epoch_end_signal}
{GradientDescent.train_end_signal}
{GradientDescent.verbose}
{GradientDescent.addons}
Attributes
----------
{GradientDescent.Attributes}
Methods
-------
{GradientDescent.Methods}
Examples
--------
>>> import numpy as np
>>> from neupy import algorithms
>>>
>>> x_train = np.array([[1, 2], [3, 4]])
>>> y_train = np.array([[1], [0]])
>>>
>>> qnnet = algorithms.QuasiNewton(
... (2, 3, 1),
... update_function='bfgs'
... )
>>> qnnet.train(x_train, y_train, epochs=10)
See Also
--------
:network:`GradientDescent` : GradientDescent algorithm.
"""
update_function = ChoiceProperty(default='bfgs',
choices={
'bfgs': bfgs,
'dfp': dfp,
'psb': psb,
'sr1': sr1,
})
h0_scale = NumberProperty(default=1, minval=0)
step = WithdrawProperty()
def init_variables(self):
super(QuasiNewton, self).init_variables()
n_params = count_parameters(self.connection)
self.variables.update(
inv_hessian=theano.shared(
name='algo:quasi-newton/matrix:inv-hessian',
value=asfloat(self.h0_scale * np.eye(int(n_params))),
),
prev_params=theano.shared(
name='algo:quasi-newton/vector:prev-params',
value=asfloat(np.zeros(n_params)),
),
prev_full_gradient=theano.shared(
name='algo:quasi-newton/vector:prev-full-gradient',
value=asfloat(np.zeros(n_params)),
),
)
def init_train_updates(self):
network_inputs = self.variables.network_inputs
network_output = self.variables.network_output
inv_hessian = self.variables.inv_hessian
prev_params = self.variables.prev_params
prev_full_gradient = self.variables.prev_full_gradient
params = parameter_values(self.connection)
param_vector = T.concatenate([param.flatten() for param in params])
gradients = T.grad(self.variables.error_func, wrt=params)
full_gradient = T.concatenate([grad.flatten() for grad in gradients])
new_inv_hessian = ifelse(
T.eq(self.variables.epoch, 1), inv_hessian,
self.update_function(inv_hessian, param_vector - prev_params,
full_gradient - prev_full_gradient))
param_delta = -new_inv_hessian.dot(full_gradient)
layers_and_parameters = list(iter_parameters(self.layers))
def prediction(step):
updated_params = param_vector + step * param_delta
# This trick allow us to replace shared variables
# with theano variables and get output from the network
start_pos = 0
for layer, attrname, param in layers_and_parameters:
end_pos = start_pos + param.size
updated_param_value = T.reshape(
updated_params[start_pos:end_pos], param.shape)
setattr(layer, attrname, updated_param_value)
start_pos = end_pos
output = self.connection.output(*network_inputs)
# Restore previous parameters
for layer, attrname, param in layers_and_parameters:
setattr(layer, attrname, param)
return output
def phi(step):
return self.error(network_output, prediction(step))
def derphi(step):
error_func = self.error(network_output, prediction(step))
return T.grad(error_func, wrt=step)
step = asfloat(line_search(phi, derphi))
updated_params = param_vector + step * param_delta
updates = setup_parameter_updates(params, updated_params)
updates.extend([
(inv_hessian, new_inv_hessian),
(prev_params, param_vector),
(prev_full_gradient, full_gradient),
])
return updates
class LevenbergMarquardt(StepSelectionBuiltIn, BaseGradientDescent):
"""
Levenberg-Marquardt algorithm is a variation of the Newton's method.
It minimizes MSE error. The algorithm approximates Hessian matrix using
dot product between two jacobian matrices.
Notes
-----
- Method requires all training data during propagation, which means
it's not allowed to use mini-batches.
- Network minimizes only Mean Squared Error (MSE) loss function.
- Efficient for small training datasets, because it
computes gradient per each sample separately.
- Efficient for small-sized networks.
Parameters
----------
{BaseGradientDescent.connection}
mu : float
Control invertion for J.T * J matrix, defaults to ``0.1``.
mu_update_factor : float
Factor to decrease the mu if update decrese the error, otherwise
increse mu by the same factor. Defaults to ``1.2``
error : {{``mse``}}
Levenberg-Marquardt works only for quadratic functions.
Defaults to ``mse``.
{BaseGradientDescent.show_epoch}
{BaseGradientDescent.shuffle_data}
{BaseGradientDescent.epoch_end_signal}
{BaseGradientDescent.train_end_signal}
{BaseGradientDescent.verbose}
{BaseGradientDescent.addons}
Attributes
----------
{BaseGradientDescent.Attributes}
Methods
-------
{BaseGradientDescent.Methods}
Examples
--------
>>> import numpy as np
>>> from neupy import algorithms
>>>
>>> x_train = np.array([[1, 2], [3, 4]])
>>> y_train = np.array([[1], [0]])
>>>
>>> lmnet = algorithms.LevenbergMarquardt((2, 3, 1))
>>> lmnet.train(x_train, y_train)
See Also
--------
:network:`BaseGradientDescent` : BaseGradientDescent algorithm.
"""
mu = BoundedProperty(default=0.01, minval=0)
mu_update_factor = BoundedProperty(default=1.2, minval=1)
error = ChoiceProperty(default='mse', choices={'mse': errors.mse})
step = WithdrawProperty()
def init_variables(self):
super(LevenbergMarquardt, self).init_variables()
self.variables.update(
mu=tf.Variable(self.mu, name='lev-marq/mu'),
last_error=tf.Variable(np.nan, name='lev-marq/last-error'),
)
def init_train_updates(self):
network_output = self.variables.network_output
prediction_func = self.variables.train_prediction_func
last_error = self.variables.last_error
error_func = self.variables.error_func
mu = self.variables.mu
new_mu = tf.where(
tf.less(last_error, error_func),
mu * self.mu_update_factor,
mu / self.mu_update_factor,
)
err_for_each_sample = flatten((network_output - prediction_func) ** 2)
params = parameter_values(self.connection)
param_vector = make_single_vector(params)
J = compute_jacobian(err_for_each_sample, params)
J_T = tf.transpose(J)
n_params = J.shape[1]
parameter_update = tf.matrix_solve(
tf.matmul(J_T, J) + new_mu * tf.eye(n_params.value),
tf.matmul(J_T, tf.expand_dims(err_for_each_sample, 1))
)
updated_params = param_vector - flatten(parameter_update)
updates = [(mu, new_mu)]
parameter_updates = setup_parameter_updates(params, updated_params)
updates.extend(parameter_updates)
return updates
def on_epoch_start_update(self, epoch):
super(LevenbergMarquardt, self).on_epoch_start_update(epoch)
last_error = self.errors.last()
if last_error is not None:
self.variables.last_error.load(last_error, tensorflow_session())
class B(A):
prop = WithdrawProperty()
class QuasiNewton(WolfeLineSearchForStep, BaseGradientDescent):
"""
Quasi-Newton algorithm. Every iteration quasi-Network method approximates
inverse Hessian matrix with iterative updates. It doesn't have ``step``
parameter. Instead, algorithm applies line search for the step parameter
that satisfies strong Wolfe condition. Parameters that control wolfe
search start with the ``wolfe_`` prefix.
Parameters
----------
update_function : ``bfgs``, ``dfp``, ``sr1``
Update function for the iterative inverse hessian matrix
approximation. Defaults to ``bfgs``.
- ``bfgs`` - It's rank 2 formula update. It can suffer from
round-off error and inaccurate line searches.
- ``dfp`` - DFP is a method very similar to BFGS. It's rank 2 formula
update. It can suffer from round-off error and inaccurate line
searches.
- ``sr1`` - Symmetric rank 1 (SR1). Generates update for the
inverse hessian matrix adding symmetric rank-1 matrix. It's
possible that there is no rank 1 updates for the matrix and in
this case update won't be applied and original inverse hessian
will be returned.
h0_scale : float
Default Hessian matrix is an identity matrix. The
``h0_scale`` parameter scales identity matrix.
Defaults to ``1``.
epsilon : float
Controls numerical stability for the ``update_function`` parameter.
Defaults to ``1e-7``.
{WolfeLineSearchForStep.Parameters}
{BaseGradientDescent.connection}
{BaseGradientDescent.error}
{BaseGradientDescent.show_epoch}
{BaseGradientDescent.shuffle_data}
{BaseGradientDescent.epoch_end_signal}
{BaseGradientDescent.train_end_signal}
{BaseGradientDescent.verbose}
{BaseGradientDescent.addons}
Notes
-----
- Method requires all training data during propagation, which means
it's not allowed to use mini-batches.
Attributes
----------
{BaseGradientDescent.Attributes}
Methods
-------
{BaseGradientDescent.Methods}
Examples
--------
>>> import numpy as np
>>> from neupy import algorithms
>>>
>>> x_train = np.array([[1, 2], [3, 4]])
>>> y_train = np.array([[1], [0]])
>>>
>>> qnnet = algorithms.QuasiNewton(
... (2, 3, 1),
... update_function='bfgs'
... )
>>> qnnet.train(x_train, y_train, epochs=10)
References
----------
[1] Yang Ding, Enkeleida Lushi, Qingguo Li,
Investigation of quasi-Newton methods for unconstrained optimization.
http://people.math.sfu.ca/~elushi/project_833.pdf
[2] Jorge Nocedal, Stephen J. Wright, Numerical Optimization.
Chapter 6, Quasi-Newton Methods, p. 135-163
"""
update_function = ChoiceProperty(default='bfgs',
choices={
'bfgs': bfgs,
'dfp': dfp,
'sr1': sr1,
})
epsilon = NumberProperty(default=1e-7, minval=0)
h0_scale = NumberProperty(default=1, minval=0)
step = WithdrawProperty()
def init_variables(self):
super(QuasiNewton, self).init_variables()
n_parameters = count_parameters(self.connection)
self.variables.update(
inv_hessian=tf.Variable(
asfloat(self.h0_scale) * tf.eye(n_parameters),
name="quasi-newton/inv-hessian",
dtype=tf.float32,
),
prev_params=tf.Variable(
tf.zeros([n_parameters]),
name="quasi-newton/prev-params",
dtype=tf.float32,
),
prev_full_gradient=tf.Variable(
tf.zeros([n_parameters]),
name="quasi-newton/prev-full-gradient",
dtype=tf.float32,
),
)
def init_train_updates(self):
inv_hessian = self.variables.inv_hessian
prev_params = self.variables.prev_params
prev_full_gradient = self.variables.prev_full_gradient
params = parameter_values(self.connection)
param_vector = make_single_vector(params)
gradients = tf.gradients(self.variables.error_func, params)
full_gradient = make_single_vector(gradients)
new_inv_hessian = tf.where(
tf.equal(self.variables.epoch, 1), inv_hessian,
self.update_function(inv_H=inv_hessian,
delta_w=param_vector - prev_params,
delta_grad=full_gradient - prev_full_gradient,
epsilon=self.epsilon))
param_delta = -dot(new_inv_hessian, full_gradient)
step = self.find_optimal_step(param_vector, param_delta)
updated_params = param_vector + step * param_delta
updates = setup_parameter_updates(params, updated_params)
# We have to compute these values first, otherwise
# parallelization in tensorflow can mix update order
# and, for example, previous gradient can be equal to
# current gradient value. It happens because tensorflow
# try to execute operations in parallel.
required_variables = [new_inv_hessian, param_vector, full_gradient]
with tf.control_dependencies(required_variables):
updates.extend([
inv_hessian.assign(new_inv_hessian),
prev_params.assign(param_vector),
prev_full_gradient.assign(full_gradient),
])
return updates
class Hessian(BaseOptimizer):
"""
Hessian gradient decent optimization, also known as Newton's method. This
algorithm uses second-order derivative (hessian matrix) in order to
choose correct step during the training iteration. Because of this,
method doesn't have ``step`` parameter.
Parameters
----------
penalty_const : float
Inverse hessian could be singular matrix. For this reason
algorithm include penalty that add to hessian matrix identity
multiplied by defined constant. Defaults to ``1``.
{BaseOptimizer.network}
{BaseOptimizer.loss}
{BaseOptimizer.regularizer}
{BaseOptimizer.show_epoch}
{BaseOptimizer.shuffle_data}
{BaseOptimizer.signals}
{BaseOptimizer.verbose}
Attributes
----------
{BaseOptimizer.Attributes}
Methods
-------
{BaseOptimizer.Methods}
Notes
-----
- Method requires all training data during propagation, which means
it cannot be trained with mini-batches.
- This method calculates full hessian matrix which means it will compute
matrix with NxN parameters, where N = number of parameters in the
network.
Examples
--------
>>> import numpy as np
>>> from neupy import algorithms
>>> from neupy.layers import *
>>>
>>> x_train = np.array([[1, 2], [3, 4]])
>>> y_train = np.array([[1], [0]])
>>>
>>> network = Input(2) >> Sigmoid(3) >> Sigmoid(1)
>>> optimizer = algorithms.Hessian(network)
>>> optimizer.train(x_train, y_train)
See Also
--------
:network:`HessianDiagonal` : Hessian diagonal approximation.
"""
penalty_const = BoundedProperty(default=1, minval=0)
step = WithdrawProperty()
def init_train_updates(self):
penalty_const = asfloat(self.penalty_const)
n_parameters = self.network.n_parameters
variables = self.network.variables
parameters = [var for var in variables.values() if var.trainable]
param_vector = make_single_vector(parameters)
hessian_matrix, full_gradient = find_hessian_and_gradient(
self.variables.loss, parameters
)
parameter_update = tf.matrix_solve(
hessian_matrix + penalty_const * tf.eye(n_parameters),
tf.reshape(full_gradient, [-1, 1])
)
updated_parameters = param_vector - flatten(parameter_update)
updates = setup_parameter_updates(parameters, updated_parameters)
return updates
class Deconvolution(Convolution):
"""
Deconvolution layer. It's commonly called like this in the literature,
but it's just gradient of the convolution and not actual deconvolution.
Parameters
----------
{Convolution.size}
{Convolution.padding}
{Convolution.stride}
weight : array-like, Tensorfow variable, scalar or Initializer
Defines layer's weights. Shape of the weight will be equal to
``(filter rows, filter columns, output channels, input channels)``.
Default initialization methods you can find
:ref:`here <init-methods>`. Defaults to
:class:`HeNormal(gain=2) <neupy.init.HeNormal>`.
{ParameterBasedLayer.bias}
{BaseLayer.Parameters}
Methods
-------
{ParameterBasedLayer.Methods}
Examples
--------
>>> from neupy import layers
>>>
>>> layers.join(
... layers.Input((28, 28, 3)),
... layers.Convolution((3, 3, 16)),
... layers.Deconvolution((3, 3, 1)),
... )
Attributes
----------
{ParameterBasedLayer.Attributes}
"""
dilation = WithdrawProperty()
def output_shape_per_dim(self, *args, **kwargs):
return deconv_output_shape(*args, **kwargs)
@property
def weight_shape(self):
# Compare to the regular convolution weights
# have switched input and output channels.
return as_tuple(self.size, self.input_shape[-1])
def output(self, input_value):
input_shape = tf.shape(input_value)
# We need to get information about output shape from the input
# tensor's shape, because for some inputs we might have
# height and width specified as None and shape value won't be
# computed for these dimensions.
output_shape = self.find_output_from_input_shape(
tf.unstack(input_shape[1:]))
batch_size = input_shape[0]
padding = self.padding
if isinstance(self.padding, (list, tuple)):
height_pad, width_pad = self.padding
# VALID option will make sure that
# deconvolution won't use any padding.
padding = 'VALID'
# conv2d_transpose doesn't know about extra paddings that we added
# in the convolution. For this reason we have to expand our
# expected output shape and later we will remove these paddings
# manually after transpose convolution.
output_shape = (
output_shape[0] + 2 * height_pad,
output_shape[1] + 2 * width_pad,
output_shape[2],
)
output = tf.nn.conv2d_transpose(
input_value,
self.weight,
as_tuple(batch_size, output_shape),
as_tuple(1, self.stride, 1),
padding,
data_format="NHWC"
)
if isinstance(self.padding, (list, tuple)):
h_pad, w_pad = self.padding
if h_pad > 0:
output = output[:, h_pad:-h_pad, :, :]
if w_pad > 0:
output = output[:, :, w_pad:-w_pad, :]
if self.bias is not None:
bias = tf.reshape(self.bias, (1, 1, 1, -1))
output += bias
return output
class LevenbergMarquardt(BaseOptimizer):
"""
Levenberg-Marquardt algorithm is a variation of the Newton's method.
It minimizes MSE error. The algorithm approximates Hessian matrix using
dot product between two jacobian matrices.
Notes
-----
- Method requires all training data during propagation, which means
it's not allowed to use mini-batches.
- Network minimizes only Mean Squared Error (MSE) loss function.
- Efficient for small training datasets, because it
computes gradient per each sample separately.
- Efficient for small-sized networks.
Parameters
----------
{BaseOptimizer.network}
mu : float
Control invertion for J.T * J matrix, defaults to ``0.1``.
mu_update_factor : float
Factor to decrease the mu if update decrese the error, otherwise
increse mu by the same factor. Defaults to ``1.2``
error : {{``mse``}}
Levenberg-Marquardt works only for quadratic functions.
Defaults to ``mse``.
{BaseOptimizer.show_epoch}
{BaseOptimizer.shuffle_data}
{BaseOptimizer.signals}
{BaseOptimizer.verbose}
Attributes
----------
{BaseOptimizer.Attributes}
Methods
-------
{BaseOptimizer.Methods}
Examples
--------
>>> import numpy as np
>>> from neupy import algorithms
>>> from neupy.layers import *
>>>
>>> x_train = np.array([[1, 2], [3, 4]])
>>> y_train = np.array([[1], [0]])
>>>
>>> network = Input(2) >> Sigmoid(3) >> Sigmoid(1)
>>> optimizer = algorithms.LevenbergMarquardt(network)
>>> optimizer.train(x_train, y_train)
See Also
--------
:network:`BaseOptimizer` : BaseOptimizer algorithm.
"""
mu = BoundedProperty(default=0.01, minval=0)
mu_update_factor = BoundedProperty(default=1.2, minval=1)
loss = ChoiceProperty(default='mse', choices={'mse': objectives.mse})
step = WithdrawProperty()
regularizer = WithdrawProperty()
def init_functions(self):
self.variables.update(
mu=tf.Variable(self.mu, name='lev-marq/mu'),
last_error=tf.Variable(np.nan, name='lev-marq/last-error'),
)
super(LevenbergMarquardt, self).init_functions()
def init_train_updates(self):
training_outputs = self.network.training_outputs
last_error = self.variables.last_error
error_func = self.variables.loss
mu = self.variables.mu
new_mu = tf.where(
tf.less(last_error, error_func),
mu * self.mu_update_factor,
mu / self.mu_update_factor,
)
err_for_each_sample = flatten((self.target - training_outputs)**2)
variables = self.network.variables
params = [var for var in variables.values() if var.trainable]
param_vector = make_single_vector(params)
J = compute_jacobian(err_for_each_sample, params)
J_T = tf.transpose(J)
n_params = J.shape[1]
parameter_update = tf.matrix_solve(
tf.matmul(J_T, J) + new_mu * tf.eye(n_params.value),
tf.matmul(J_T, tf.expand_dims(err_for_each_sample, 1)))
updated_params = param_vector - flatten(parameter_update)
updates = [(mu, new_mu)]
parameter_updates = setup_parameter_updates(params, updated_params)
updates.extend(parameter_updates)
return updates
def one_training_update(self, X_train, y_train):
if self.errors.train:
last_error = self.errors.train[-1]
self.variables.last_error.load(last_error, tensorflow_session())
return super(LevenbergMarquardt,
self).one_training_update(X_train, y_train)
class RBFKMeans(StepSelectionBuiltIn, BaseNetwork):
"""
Radial basis function K-means for clustering.
Parameters
----------
n_clusters : int
Number of clusters.
{BaseNetwork.show_epoch}
{BaseNetwork.shuffle_data}
{BaseNetwork.epoch_end_signal}
{BaseNetwork.train_end_signal}
{Verbose.verbose}
Attributes
----------
centers : array-like with shape (n_clusters, n_futures)
Cluster centers.
Methods
-------
train(input_train, epsilon=1e-5, epochs=100)
Trains network.
{BaseSkeleton.predict}
{BaseSkeleton.fit}
Examples
--------
>>> import numpy as np
>>> from neupy.algorithms import RBFKMeans
>>>
>>> data = np.array([
... [0.11, 0.20],
... [0.25, 0.32],
... [0.64, 0.60],
... [0.12, 0.42],
... [0.70, 0.73],
... [0.30, 0.27],
... [0.43, 0.81],
... [0.44, 0.87],
... [0.12, 0.92],
... [0.56, 0.67],
... [0.36, 0.35],
... ])
>>> rbfk_net = RBFKMeans(n_clusters=2, verbose=False)
>>> rbfk_net.train(data, epsilon=1e-5)
>>> rbfk_net.centers
array([[ 0.228 , 0.312 ],
[ 0.48166667, 0.76666667]])
>>>
>>> new_data = np.array([[0.1, 0.1], [0.9, 0.9]])
>>> rbfk_net.predict(new_data)
array([[ 0.],
[ 1.]])
"""
n_clusters = IntProperty(minval=2)
step = WithdrawProperty()
def __init__(self, **options):
self.centers = None
super(RBFKMeans, self).__init__(**options)
def predict(self, input_data):
input_data = format_data(input_data)
centers = self.centers
classes = np.zeros((input_data.shape[0], 1))
for i, value in enumerate(input_data):
classes[i] = np.argmin(norm(centers - value, axis=1))
return classes
def train_epoch(self, input_train, target_train):
centers = self.centers
old_centers = centers.copy()
output_train = self.predict(input_train)
for i, center in enumerate(centers):
positions = np.argwhere(output_train[:, 0] == i)
if not np.any(positions):
continue
class_data = np.take(input_train, positions, axis=0)
centers[i, :] = (1 / len(class_data)) * np.sum(class_data, axis=0)
return np.abs(old_centers - centers)
def train(self, input_train, epsilon=1e-5, epochs=100):
n_clusters = self.n_clusters
input_train = format_data(input_train)
n_samples = input_train.shape[0]
if n_samples <= n_clusters:
raise ValueError("Number of samples in the dataset is less than "
"spcified number of clusters. Got {} samples, "
"expected at least {} (for {} clusters)"
"".format(n_samples, n_clusters + 1, n_clusters))
self.centers = input_train[:n_clusters, :].copy()
super(RBFKMeans, self).train(input_train,
epsilon=epsilon,
epochs=epochs)
class Hessian(StepSelectionBuiltIn, GradientDescent):
"""
Hessian gradient decent optimization. This GD algorithm
variation using second derivative information helps choose better
gradient direction and as a consequence better weight update
parameter after each epoch.
Parameters
----------
penalty_const : float
Inverse hessian could be singular matrix. For this reason
algorithm include penalty that add to hessian matrix identity
multiplied by defined constant. Defaults to ``1``.
{GradientDescent.connection}
{GradientDescent.error}
{GradientDescent.show_epoch}
{GradientDescent.shuffle_data}
{GradientDescent.epoch_end_signal}
{GradientDescent.train_end_signal}
{GradientDescent.verbose}
{GradientDescent.addons}
Attributes
----------
{GradientDescent.Attributes}
Methods
-------
{GradientDescent.Methods}
Examples
--------
>>> import numpy as np
>>> from neupy import algorithms
>>>
>>> x_train = np.array([[1, 2], [3, 4]])
>>> y_train = np.array([[1], [0]])
>>>
>>> mnet = algorithms.Hessian((2, 3, 1))
>>> mnet.train(x_train, y_train)
See Also
--------
:network:`HessianDiagonal` : Hessian diagonal approximation.
"""
penalty_const = BoundedProperty(default=1, minval=0)
step = WithdrawProperty()
def init_train_updates(self):
n_parameters = count_parameters(self.connection)
parameters = parameter_values(self.connection)
param_vector = T.concatenate([param.flatten() for param in parameters])
penalty_const = asfloat(self.penalty_const)
print n_parameters
self.variables.hessian = theano.shared(value=asfloat(
np.zeros((n_parameters, n_parameters))),
name='hessian_inverse')
hessian_matrix, full_gradient = find_hessian_and_gradient(
self.variables.error_func, parameters)
updated_parameters = hessian_matrix
updates = setup_parameter_updates([self.variables.hessian],
updated_parameters)
return updates
class LevenbergMarquardt(StepSelectionBuiltIn, GradientDescent):
"""
Levenberg-Marquardt algorithm.
Notes
-----
- Network minimizes only Mean Squared Error function.
- Efficient for small training datasets, because it
computes gradient per each sample separately.
- Efficient for small-sized networks.
Parameters
----------
{GradientDescent.connection}
mu : float
Control invertion for J.T * J matrix, defaults to `0.1`.
mu_update_factor : float
Factor to decrease the mu if update decrese the error, otherwise
increse mu by the same factor. Defaults to ``1.2``
error : {{``mse``}}
Levenberg-Marquardt works only for quadratic functions.
Defaults to ``mse``.
{GradientDescent.show_epoch}
{GradientDescent.shuffle_data}
{GradientDescent.epoch_end_signal}
{GradientDescent.train_end_signal}
{GradientDescent.verbose}
{GradientDescent.addons}
Attributes
----------
{GradientDescent.Attributes}
Methods
-------
{GradientDescent.Methods}
Examples
--------
>>> import numpy as np
>>> from neupy import algorithms
>>>
>>> x_train = np.array([[1, 2], [3, 4]])
>>> y_train = np.array([[1], [0]])
>>>
>>> lmnet = algorithms.LevenbergMarquardt((2, 3, 1))
>>> lmnet.train(x_train, y_train)
See Also
--------
:network:`GradientDescent` : GradientDescent algorithm.
"""
mu = BoundedProperty(default=0.01, minval=0)
mu_update_factor = BoundedProperty(default=1.2, minval=1)
error = ChoiceProperty(default='mse', choices={'mse': errors.mse})
step = WithdrawProperty()
def init_variables(self):
super(LevenbergMarquardt, self).init_variables()
self.variables.update(
mu=theano.shared(name='lev-marq/mu', value=asfloat(self.mu)),
last_error=theano.shared(name='lev-marq/last-error', value=np.nan),
)
def init_train_updates(self):
network_output = self.variables.network_output
prediction_func = self.variables.train_prediction_func
last_error = self.variables.last_error
error_func = self.variables.error_func
mu = self.variables.mu
new_mu = ifelse(
T.lt(last_error, error_func),
mu * self.mu_update_factor,
mu / self.mu_update_factor,
)
se_for_each_sample = ((network_output - prediction_func)**2).ravel()
params = parameter_values(self.connection)
param_vector = T.concatenate([param.flatten() for param in params])
J = compute_jacobian(se_for_each_sample, params)
n_params = J.shape[1]
updated_params = param_vector - slinalg.solve(
J.T.dot(J) + new_mu * T.eye(n_params), J.T.dot(se_for_each_sample))
updates = [(mu, new_mu)]
parameter_updates = setup_parameter_updates(params, updated_params)
updates.extend(parameter_updates)
return updates
def on_epoch_start_update(self, epoch):
super(LevenbergMarquardt, self).on_epoch_start_update(epoch)
last_error = self.errors.last()
if last_error is not None:
self.variables.last_error.set_value(last_error)
class Hessian(StepSelectionBuiltIn, BaseGradientDescent):
"""
Hessian gradient decent optimization, also known as Newton's method. This
algorithm uses second-order derivative (hessian matrix) in order to
choose correct step during the training itration. Because of this,
method doesn't have ``step`` parameter.
Parameters
----------
penalty_const : float
Inverse hessian could be singular matrix. For this reason
algorithm include penalty that add to hessian matrix identity
multiplied by defined constant. Defaults to ``1``.
{BaseGradientDescent.connection}
{BaseGradientDescent.error}
{BaseGradientDescent.show_epoch}
{BaseGradientDescent.shuffle_data}
{BaseGradientDescent.epoch_end_signal}
{BaseGradientDescent.train_end_signal}
{BaseGradientDescent.verbose}
{BaseGradientDescent.addons}
Attributes
----------
{BaseGradientDescent.Attributes}
Methods
-------
{BaseGradientDescent.Methods}
Notes
-----
- Method requires all training data during propagation, which means
it's not allowed to use mini-batches.
- This method calculates full hessian matrix which means it will compute
matrix with NxN parameters, where N = number of parameters in the
network.
Examples
--------
>>> import numpy as np
>>> from neupy import algorithms
>>>
>>> x_train = np.array([[1, 2], [3, 4]])
>>> y_train = np.array([[1], [0]])
>>>
>>> mnet = algorithms.Hessian((2, 3, 1))
>>> mnet.train(x_train, y_train)
See Also
--------
:network:`HessianDiagonal` : Hessian diagonal approximation.
"""
penalty_const = BoundedProperty(default=1, minval=0)
step = WithdrawProperty()
def init_train_updates(self):
penalty_const = asfloat(self.penalty_const)
n_parameters = count_parameters(self.connection)
parameters = parameter_values(self.connection)
param_vector = make_single_vector(parameters)
hessian_matrix, full_gradient = find_hessian_and_gradient(
self.variables.error_func, parameters)
parameter_update = tf.matrix_solve(
hessian_matrix + penalty_const * tf.eye(n_parameters),
tf.reshape(full_gradient, [-1, 1]))
updated_parameters = param_vector - flatten(parameter_update)
updates = setup_parameter_updates(parameters, updated_parameters)
return updates