def bernoulli_csl(switch=0):
mnist = MNIST()
train_x, _ = mnist.getSubset(TRAIN)
valid_x, _ = mnist.getSubset(VALID)
test_x, _ = mnist.getSubset(TEST)
mnist_b = MNIST(binary=True)
train_x_b, _ = mnist_b.getSubset(TRAIN)
valid_x_b, _ = mnist_b.getSubset(VALID)
test_x_b, _ = mnist_b.getSubset(TEST)
means = as_floatX(test_x).eval()
means = numpy.clip(a=means, a_min=1e-10, a_max=(1 - (1e-5)))
# means = as_floatX(numpy.random.uniform(size=(10000,784))) * 0 + 0.5
minibatches = as_floatX(test_x_b.reshape((1000, 10, 784))).eval()
if switch:
# when means is a matrix of (N,D), representing only 1 chain
csl_fn = _compile_csl_fn_v2(means)
compute_CSL_with_minibatches_one_chain(csl_fn, minibatches)
else:
# when means is a 3D tensor (N, K, D)
# When there are N chains, each chain having K samples of dimension D
chains = means.reshape(10, 100, 10, 784)
csl_fn = _compile_csl_fn()
compute_CSL_with_minibatches(csl_fn, minibatches, chains)
del mnist
del mnist_b
def bernoulli_csl(switch=0):
mnist = MNIST()
train_x, _ = mnist.getSubset(TRAIN)
valid_x, _ = mnist.getSubset(VALID)
test_x, _ = mnist.getSubset(TEST)
mnist_b = MNIST(binary=True)
train_x_b, _ = mnist_b.getSubset(TRAIN)
valid_x_b, _ = mnist_b.getSubset(VALID)
test_x_b, _ = mnist_b.getSubset(TEST)
means = as_floatX(test_x).eval()
means = numpy.clip(a=means, a_min=1e-10, a_max=(1 - (1e-5)))
#means = as_floatX(numpy.random.uniform(size=(10000,784))) * 0 + 0.5
minibatches = as_floatX(test_x_b.reshape((1000, 10, 784))).eval()
if switch:
# when means is a matrix of (N,D), representing only 1 chain
csl_fn = _compile_csl_fn_v2(means)
compute_CSL_with_minibatches_one_chain(csl_fn, minibatches)
else:
# when means is a 3D tensor (N, K, D)
# When there are N chains, each chain having K samples of dimension D
chains = means.reshape(10, 100, 10, 784)
csl_fn = _compile_csl_fn()
compute_CSL_with_minibatches(csl_fn, minibatches, chains)
del mnist
del mnist_b
def _compile_csl_fn():
"""
BUG HERE, not doing properly by chains (still has the bug, I don't see it)
This is taking too much GPU mem
mean: N(# of chains)*K(samples per chain)*D(data dim)
minibatch: M(# of examples)*D (data dim)
M * N matrix where each element is LL of one example against one chain.
This function is for computing CSL over parallel chains of minibatches.
Returns
-------
theano function
Function computing M * N matrix where each element is LL of one example against one chain.
"""
# when means is a 3D tensor (N, K, D)
# When there are N chains, each chain having K samples of dimension D
log.debug('building theano fn for Bernoulli CSL')
means = T.tensor3('chains')
minibatch = T.matrix('inputs')
# how many chains CSL average over
N = 5
# minibatch size
M = 10
# data dim
D = 784
minibatch.tag.test_value = as_floatX(numpy.random.binomial(1, 0.5, size=(M, D)))
# chain length
K = 100
means.tag.test_value = as_floatX(numpy.random.uniform(size=(N, K, D)))
# computing LL
# the length of each chain
sample_size = means.shape[1]
_minibatch = minibatch.dimshuffle(0, 'x', 'x', 1)
_means = means.dimshuffle('x', 0, 1, 2)
A = T.log(sample_size)
B = _minibatch * T.log(_means) + (1. - _minibatch) * T.log(1. - _means)
C = B.sum(axis=3)
D = log_sum_exp_theano(C, axis=2)
E = D - A
# G = E.mean(axis=1)
f = function(
inputs=[minibatch, means],
outputs=E,
name='CSL_independent_bernoulli_fn'
)
return f
def _compile_csl_fn():
"""
BUG HERE, not doing properly by chains (still has the bug, I don't see it)
This is taking too much GPU mem
mean: N(# of chains)*K(samples per chain)*D(data dim)
minibatch: M(# of examples)*D (data dim)
M * N matrix where each element is LL of one example against one chain.
This function is for computing CSL over parallel chains of minibatches.
Returns
-------
theano function
Function computing M * N matrix where each element is LL of one example against one chain.
"""
# when means is a 3D tensor (N, K, D)
# When there are N chains, each chain having K samples of dimension D
log.debug('building theano fn for Bernoulli CSL')
means = T.tensor3('chains')
minibatch = T.matrix('inputs')
# how many chains CSL average over
N = 5
# minibatch size
M = 10
# data dim
D = 784
minibatch.tag.test_value = as_floatX(
numpy.random.binomial(1, 0.5, size=(M, D)))
# chain length
K = 100
means.tag.test_value = as_floatX(numpy.random.uniform(size=(N, K, D)))
# computing LL
# the length of each chain
sample_size = means.shape[1]
_minibatch = minibatch.dimshuffle(0, 'x', 'x', 1)
_means = means.dimshuffle('x', 0, 1, 2)
A = T.log(sample_size)
B = _minibatch * T.log(_means) + (1. - _minibatch) * T.log(1. - _means)
C = B.sum(axis=3)
D = log_sum_exp_theano(C, axis=2)
E = D - A
# G = E.mean(axis=1)
f = function(inputs=[minibatch, means],
outputs=E,
name='CSL_independent_bernoulli_fn')
return f
def get_bias(shape, name="b", init_values=None):
"""
This creates a theano shared variable for the bias parameter - normally initialized to zeros,
but you can specify other values
Parameters
----------
shape : tuple
The shape to use for the bias vector/matrix.
name : str
The name to give the shared variable.
offset : float or array_like
Values to add to the zeros, if you want a nonzero bias initially.
Returns
-------
shared variable
The theano shared variable with given shape.
"""
default_init = 0
init_values = init_values or default_init
log.debug("Initializing bias variable with shape %s" % str(shape))
# init to zeros plus the offset
val = as_floatX(
numpy.ones(shape=shape, dtype=theano.config.floatX) * init_values)
return theano.shared(value=val, name=name)
def __init__(self, param, initial, reduction_factor):
"""
A generic class for decaying a theano variable.
Parameters
----------
param : shared variable
The theano variable you want to decay. This must already be a shared variable.
initial : float
The initial value the variable should have.
reduction_factor : float
The amount of reduction (depending on subclass's algorithm) each epoch.
"""
# make sure the parameter is a Theano shared variable
if not hasattr(param, 'get_value'):
log.error('Parameter doesn\'t have a get_value() function! It is supposed to be a shared variable...')
if not hasattr(param, 'set_value'):
log.error('Parameter doesn\'t have a set_value() function! It is supposed to be a shared variable...')
assert hasattr(param, 'get_value')
assert hasattr(param, 'set_value')
self.param = param
self.initial = initial
self.param.set_value(as_floatX(self.initial))
self.reduction_factor = reduction_factor
def __init__(self, param, initial, reduction_factor):
"""
A generic class for decaying a theano variable.
Parameters
----------
param : shared variable
The theano variable you want to decay. This must already be a shared variable.
initial : float
The initial value the variable should have.
reduction_factor : float
The amount of reduction (depending on subclass's algorithm) each epoch.
"""
# make sure the parameter is a Theano shared variable
if not hasattr(param, 'get_value'):
log.error(
'Parameter doesn\'t have a get_value() function! It is supposed to be a shared variable...'
)
if not hasattr(param, 'set_value'):
log.error(
'Parameter doesn\'t have a set_value() function! It is supposed to be a shared variable...'
)
assert hasattr(param, 'get_value')
assert hasattr(param, 'set_value')
self.param = param
self.initial = initial
self.param.set_value(as_floatX(self.initial))
self.reduction_factor = reduction_factor
def get_bias(shape, name="b", init_values=None):
"""
This creates a theano shared variable for the bias parameter - normally initialized to zeros,
but you can specify other values
Parameters
----------
shape : tuple
The shape to use for the bias vector/matrix.
name : str
The name to give the shared variable.
offset : float or array_like
Values to add to the zeros, if you want a nonzero bias initially.
Returns
-------
shared variable
The theano shared variable with given shape.
"""
default_init = 0
init_values = init_values or default_init
log.debug("Initializing bias variable with shape %s" % str(shape))
# init to zeros plus the offset
val = as_floatX(numpy.ones(shape=shape, dtype=theano.config.floatX) * init_values)
return theano.shared(value=val, name=name)
def rectifier(x):
"""
Returns the element-wise rectifier (ReLU) applied to x.
Parameters
----------
x : tensor
Symbolic Tensor (or compatible).
Returns
-------
tensor
Element-wise rectifier: rectifier(x) = max(0,x) applied to `x`.
.. note::
This implementation uses rectifier(x) = (x + abs(x)) / 2
which is faster than max(0,x)
See https://github.com/SnippyHolloW/abnet/blob/807aeb9/layers.py#L15
"""
# return T.maximum(as_floatX(0), x)
# below fix is taken from Lasagne framework:
# https://github.com/benanne/Lasagne/blob/master/lasagne/nonlinearities.py
# The following is faster than lambda x: T.maximum(0, x)
# Thanks to @SnippyHolloW for pointing this out.
# See: https://github.com/SnippyHolloW/abnet/blob/807aeb9/layers.py#L15
return (x + abs(x)) / as_floatX(2.0)
def rectifier(x):
"""
Returns the element-wise rectifier (ReLU) applied to x.
Parameters
----------
x : tensor
Symbolic Tensor (or compatible).
Returns
-------
tensor
Element-wise rectifier: rectifier(x) = max(0,x) applied to `x`.
.. note::
This implementation uses rectifier(x) = (x + abs(x)) / 2
which is faster than max(0,x)
See https://github.com/SnippyHolloW/abnet/blob/807aeb9/layers.py#L15
"""
# return T.maximum(as_floatX(0), x)
# below fix is taken from Lasagne framework:
# https://github.com/benanne/Lasagne/blob/master/lasagne/nonlinearities.py
# The following is faster than lambda x: T.maximum(0, x)
# Thanks to @SnippyHolloW for pointing this out.
# See: https://github.com/SnippyHolloW/abnet/blob/807aeb9/layers.py#L15
return (x + abs(x)) / as_floatX(2.0)
def get_weights_uniform(shape, interval='montreal', name="W", rng=None):
"""
This initializes a shared variable with a given shape for weights drawn from a Uniform distribution with
low = -interval and high = interval.
Interval can either be a number to use, or a string key to one of the predefined formulas in the
_uniform_interval dictionary.
Parameters
----------
shape : tuple
A tuple giving the shape information for this weight matrix.
interval : float or str
Either a number for your own custom interval, or a string key to one of the predefined formulas.
name : str
The name to give the shared variable.
rng : random
The random number generator to use with a .uniform method.
Returns
-------
shared variable
The theano shared variable with given shape and name drawn from a uniform distribution.
Raises
------
NotImplementedError
If the string name for the interval couldn't be found in the dictionary.
"""
if rng is None:
rng = numpy.random
# If the interval parameter is a string, grab the appropriate formula from the function dictionary,
# and apply the appropriate shape numbers to it.
if isinstance(interval, six.string_types):
interval_func = _uniform_interval.get(interval)
if interval_func is None:
log.error(
'Could not find uniform interval formula %s, try one of %s instead.'
% str(interval), str(_uniform_interval.keys()))
raise NotImplementedError(
'Could not find uniform interval formula %s, try one of %s instead.'
% str(interval), str(_uniform_interval.keys()))
else:
log.debug(
"Creating weights with shape %s from Uniform distribution with formula name: %s",
str(shape), str(interval))
interval = interval_func(shape)
else:
log.debug(
"Creating weights with shape %s from Uniform distribution with given interval +- %s",
str(shape), str(interval))
# build the uniform weights tensor
val = as_floatX(rng.uniform(low=-interval, high=interval, size=shape))
# check if a theano rng was used
if isinstance(val, T.TensorVariable):
val = val.eval()
# make it into a shared variable
return theano.shared(value=val, name=name)
def get_weights_uniform(shape, interval='montreal', name="W", rng=None):
"""
This initializes a shared variable with a given shape for weights drawn from a Uniform distribution with
low = -interval and high = interval.
Interval can either be a number to use, or a string key to one of the predefined formulas in the
_uniform_interval dictionary.
Parameters
----------
shape : tuple
A tuple giving the shape information for this weight matrix.
interval : float or str
Either a number for your own custom interval, or a string key to one of the predefined formulas.
name : str
The name to give the shared variable.
rng : random
The random number generator to use with a .uniform method.
Returns
-------
shared variable
The theano shared variable with given shape and name drawn from a uniform distribution.
Raises
------
NotImplementedError
If the string name for the interval couldn't be found in the dictionary.
"""
if rng is None:
rng = numpy.random
# If the interval parameter is a string, grab the appropriate formula from the function dictionary,
# and apply the appropriate shape numbers to it.
if isinstance(interval, six.string_types):
interval_func = _uniform_interval.get(interval)
if interval_func is None:
log.error('Could not find uniform interval formula %s, try one of %s instead.' %
str(interval), str(_uniform_interval.keys()))
raise NotImplementedError('Could not find uniform interval formula %s, try one of %s instead.' %
str(interval), str(_uniform_interval.keys()))
else:
log.debug("Creating weights with shape %s from Uniform distribution with formula name: %s",
str(shape), str(interval))
interval = interval_func(shape)
else:
log.debug("Creating weights with shape %s from Uniform distribution with given interval +- %s",
str(shape), str(interval))
# build the uniform weights tensor
val = as_floatX(rng.uniform(low=-interval, high=interval, size=shape))
# check if a theano rng was used
if isinstance(val, T.TensorVariable):
val = val.eval()
# make it into a shared variable
return theano.shared(value=val, name=name)
def get_weights_gaussian(shape, mean=None, std=None, name="W", rng=None):
"""
This initializes a shared variable with the given shape for weights drawn from a
Gaussian distribution with mean and std.
Parameters
----------
shape : tuple
A tuple giving the shape information for this weight matrix.
mean : float
The mean to use for the Gaussian distribution.
std : float
The standard deviation to use dor the Gaussian distribution.
name : str
The name to give the shared variable.
rng : random
A given random number generator to use with .normal method.
Returns
-------
shared variable
The theano shared variable with given shape and drawn from a Gaussian distribution.
"""
default_mean = 0
default_std = 0.05
mean = mean or default_mean
std = std or default_std
log.debug("Creating weights with shape %s from Gaussian mean=%s, std=%s",
str(shape), str(mean), str(std))
if rng is None:
rng = numpy.random
if std != 0:
if isinstance(rng, type(numpy.random)):
val = numpy.asarray(rng.normal(loc=mean, scale=std, size=shape),
dtype=theano.config.floatX)
else:
val = numpy.asarray(rng.normal(avg=mean, std=std,
size=shape).eval(),
dtype=theano.config.floatX)
else:
val = as_floatX(mean * numpy.ones(shape, dtype=theano.config.floatX))
# check if a theano rng was used
if isinstance(val, T.TensorVariable):
val = val.eval()
# make it into a shared variable
return theano.shared(value=val, name=name)
def get_weights_gaussian(shape, mean=None, std=None, name="W", rng=None):
"""
This initializes a shared variable with the given shape for weights drawn from a
Gaussian distribution with mean and std.
Parameters
----------
shape : tuple
A tuple giving the shape information for this weight matrix.
mean : float
The mean to use for the Gaussian distribution.
std : float
The standard deviation to use dor the Gaussian distribution.
name : str
The name to give the shared variable.
rng : random
A given random number generator to use with .normal method.
Returns
-------
shared variable
The theano shared variable with given shape and drawn from a Gaussian distribution.
"""
default_mean = 0
default_std = 0.05
mean = mean or default_mean
std = std or default_std
log.debug("Creating weights with shape %s from Gaussian mean=%s, std=%s", str(shape), str(mean), str(std))
if rng is None:
rng = numpy.random
if std != 0:
if isinstance(rng, type(numpy.random)):
val = numpy.asarray(rng.normal(loc=mean, scale=std, size=shape), dtype=theano.config.floatX)
else:
val = numpy.asarray(rng.normal(avg=mean, std=std, size=shape).eval(), dtype=theano.config.floatX)
else:
val = as_floatX(mean * numpy.ones(shape, dtype=theano.config.floatX))
# check if a theano rng was used
if isinstance(val, T.TensorVariable):
val = val.eval()
# make it into a shared variable
return theano.shared(value=val, name=name)
def _compile_csl_fn_v2(mu):
"""
p(x) = sum_h p(x|h)p(h) where p(x|h) is independent Bernoulli with
a vector mu, mu_i for dim_i
This function is for computing CSL over minibatches (in a single chain).
Parameters
----------
mu : array_like
mu is (N,D) numpy array
Returns
-------
theano function
Function computing the Bernoulli CSL log likelihood.
"""
#
log.debug('building theano fn for Bernoulli CSL')
x = T.fmatrix('inputs')
x.tag.test_value = as_floatX(numpy.random.uniform(size=(10, 784)))
mu = numpy.clip(mu, 1e-10, (1 - (1e-5)))
mu = mu[None, :, :]
inner_1 = numpy.log(mu)
inner_2 = numpy.log(1. - mu)
k = mu.shape[1]
D = mu.shape[2]
# there are two terms in the log(p(x|mu))
term_1 = -T.log(k)
c = T.sum(x.dimshuffle(0, 'x', 1) * inner_1 +
(1. - x.dimshuffle(0, 'x', 1)) * inner_2,
axis=2)
debug = c.sum(axis=1)
term_2 = log_sum_exp_theano(c, axis=1)
log_likelihood = term_1 + term_2
f = function([x], log_likelihood, name='CSL_independent_bernoulli_fn')
return f
def _compile_csl_fn_v2(mu):
"""
p(x) = sum_h p(x|h)p(h) where p(x|h) is independent Bernoulli with
a vector mu, mu_i for dim_i
This function is for computing CSL over minibatches (in a single chain).
Parameters
----------
mu : array_like
mu is (N,D) numpy array
Returns
-------
theano function
Function computing the Bernoulli CSL log likelihood.
"""
#
log.debug('building theano fn for Bernoulli CSL')
x = T.fmatrix('inputs')
x.tag.test_value = as_floatX(numpy.random.uniform(size=(10, 784)))
mu = numpy.clip(mu, 1e-10, (1 - (1e-5)))
mu = mu[None, :, :]
inner_1 = numpy.log(mu)
inner_2 = numpy.log(1. - mu)
k = mu.shape[1]
D = mu.shape[2]
# there are two terms in the log(p(x|mu))
term_1 = -T.log(k)
c = T.sum(x.dimshuffle(0, 'x', 1) * inner_1 +
(1. - x.dimshuffle(0, 'x', 1)) * inner_2,
axis=2)
debug = c.sum(axis=1)
term_2 = log_sum_exp_theano(c, axis=1)
log_likelihood = term_1 + term_2
f = function([x], log_likelihood, name='CSL_independent_bernoulli_fn')
return f
def __init__(self,
inputs_hook=None,
hiddens_hook=None,
params_hook=None,
outdir='outputs/gsn/',
input_size=None,
hidden_size=1000,
layers=2,
walkbacks=4,
visible_activation='sigmoid',
hidden_activation='tanh',
input_sampling=True,
mrg=RNG_MRG.MRG_RandomStreams(1),
tied_weights=True,
weights_init='uniform',
weights_interval='montreal',
weights_mean=0,
weights_std=5e-3,
bias_init=0.0,
cost_function='binary_crossentropy',
cost_args=None,
add_noise=True,
noiseless_h1=True,
hidden_noise='gaussian',
hidden_noise_level=2,
input_noise='salt_and_pepper',
input_noise_level=0.4,
noise_decay='exponential',
noise_annealing=1,
image_width=None,
image_height=None,
**kwargs):
"""
Initialize a GSN.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere.
This is used for linking different models together (e.g. setting the DAE model's hidden layers to the RNN's
output layer gives a generative recurrent model.) For now, it needs to include the shape
information (normally the dimensionality of the hiddens i.e. n_hidden).
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters - such as a training model with dropout applied
to layers and one without for testing, where the parameters are shared between the two.
outdir : str
The directory you want outputs (parameters, images, etc.) to save to. If None, nothing will
be saved.
input_size : int
The size (dimensionality) of the input to the DAE. If shape is provided in `inputs_hook`, this is optional.
The :class:`Model` requires an `output_size`, which gets set to this value because the DAE is an
unsupervised model. The output is a reconstruction of the input.
hidden_size : int
The size (dimensionality) of the hidden layer for the DAE. Generally, you want it to be larger than
`input_size`, which is known as *overcomplete*.
visible_activation : str or callable
The nonlinear (or linear) visible activation to perform after the dot product from hiddens -> visible layer.
This activation function should be appropriate for the input unit types, i.e. 'sigmoid' for binary inputs.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The nonlinear (or linear) hidden activation to perform after the dot product from visible -> hiddens layer.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
layers : int
The number of hidden layers to use.
walkbacks : int
The number of walkbacks to perform (the variable K in Bengio's paper above). A walkback is a Gibbs sample
from the DAE, which means the model generates inputs in sequence, where each generated input is compared
to the original input to create the reconstruction cost for training. For running the model, the very last
generated input in the Gibbs chain is used as the output.
input_sampling : bool
During walkbacks, whether to sample from the generated input to create a new starting point for the next
walkback (next step in the Gibbs chain). This generally makes walkbacks more effective by making the
process more stochastic - more likely to find spurious modes in the model's representation.
mrg : random
A random number generator that is used when adding noise into the network and for sampling from the input.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
tied_weights : bool
DAE has two weight matrices - W from input -> hiddens and V from hiddens -> input. This boolean
determines if V = W.T, which 'ties' V to W and reduces the number of parameters necessary during training.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the reconstruction cost of the model. This should be appropriate
for the type of input, i.e. use 'binary_crossentropy' for binary inputs, or 'mse' for real-valued inputs.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
add_noise : bool
Whether to add noise (corrupt) the input before passing it through the computation graph during training.
This should most likely be set to the default of True, because this is a *denoising* autoencoder after all.
noiseless_h1 : bool
Whether to not add noise (corrupt) the hidden layer during computation.
hidden_noise : str
What type of noise to use for corrupting the hidden layer (if not `noiseless_h1`). See opendeep.utils.noise
for options. This should be appropriate for the hidden unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
hidden_noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
input_noise : str
What type of noise to use for corrupting the input before computation (if `add_noise`).
See opendeep.utils.noise for options. This should be appropriate for the input units, i.e. salt-and-pepper
for binary units, etc.
input_noise_level : float
The amount of noise used to corrupt the input. This could be the masking probability for salt-and-pepper,
standard deviation for Gaussian, interval for Uniform, etc.
noise_decay : str or False
Whether to use `input_noise` scheduling (decay `input_noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the DAE learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_annealing : float
The amount to reduce the `input_noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
image_width : int
If the input should be represented as an image, the width of the input image. If not specified, it will be
close to the square factor of the `input_size`.
image_height : int
If the input should be represented as an image, the height of the input image. If not specified, it will be
close to the square factor of the `input_size`.
"""
# init Model to combine the defaults and config dictionaries with the initial parameters.
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(GSN, self).__init__(**initial_parameters)
# when the input should be thought of as an image, either use the specified width and height,
# or try to make as square as possible.
if image_height is None and image_width is None:
(_h, _w) = closest_to_square_factors(self.input_size)
self.image_width = _w
self.image_height = _h
else:
self.image_height = image_height
self.image_width = image_width
############################
# Theano variables and RNG #
############################
if self.inputs_hook is None:
self.X = T.matrix('X')
else:
# inputs_hook is a (shape, input) tuple
self.X = self.inputs_hook[1]
##########################
# Network specifications #
##########################
# generally, walkbacks should be at least 2*layers
if layers % 2 == 0:
if walkbacks < 2 * layers:
log.warning(
'Not enough walkbacks for the layers! Layers is %s and walkbacks is %s. '
'Generaly want 2X walkbacks to layers', str(layers),
str(walkbacks))
else:
if walkbacks < 2 * layers - 1:
log.warning(
'Not enough walkbacks for the layers! Layers is %s and walkbacks is %s. '
'Generaly want 2X walkbacks to layers', str(layers),
str(walkbacks))
self.add_noise = add_noise
self.noise_annealing = as_floatX(
noise_annealing) # noise schedule parameter
self.hidden_noise_level = sharedX(hidden_noise_level,
dtype=theano.config.floatX)
self.hidden_noise = get_noise(name=hidden_noise,
noise_level=self.hidden_noise_level,
mrg=mrg)
self.input_noise_level = sharedX(input_noise_level,
dtype=theano.config.floatX)
self.input_noise = get_noise(name=input_noise,
noise_level=self.input_noise_level,
mrg=mrg)
self.walkbacks = walkbacks
self.tied_weights = tied_weights
self.layers = layers
self.noiseless_h1 = noiseless_h1
self.input_sampling = input_sampling
self.noise_decay = noise_decay
# if there was a hiddens_hook, unpack the hidden layers in the tensor
if self.hiddens_hook is not None:
hidden_size = self.hiddens_hook[0]
self.hiddens_flag = True
else:
self.hiddens_flag = False
# determine the sizes of each layer in a list.
# layer sizes, from h0 to hK (h0 is the visible layer)
hidden_size = list(raise_to_list(hidden_size))
if len(hidden_size) == 1:
self.layer_sizes = [self.input_size] + hidden_size * self.layers
else:
assert len(hidden_size) == self.layers, "Hiddens sizes and number of hidden layers mismatch." + \
"Hiddens %d and layers %d" % (len(hidden_size), self.layers)
self.layer_sizes = [self.input_size] + hidden_size
if self.hiddens_hook is not None:
self.hiddens = self.unpack_hiddens(self.hiddens_hook[1])
#########################
# Activation functions! #
#########################
# hidden unit activation
self.hidden_activation = get_activation_function(hidden_activation)
# Visible layer activation
self.visible_activation = get_activation_function(visible_activation)
# make sure the sampling functions are appropriate for the activation functions.
if is_binary(self.visible_activation):
self.visible_sampling = mrg.binomial
else:
# TODO: implement non-binary activation
log.error("Non-binary visible activation not supported yet!")
raise NotImplementedError(
"Non-binary visible activation not supported yet!")
# Cost function
self.cost_function = get_cost_function(cost_function)
self.cost_args = cost_args or dict()
###############
# Parameters! #
###############
# make sure to deal with params_hook!
if self.params_hook is not None:
# if tied weights, expect layers*2 + 1 params
if self.tied_weights:
assert len(self.params_hook) == 2*layers + 1, \
"Tied weights: expected {0!s} params, found {1!s}!".format(2*layers+1, len(self.params_hook))
self.weights_list = self.params_hook[:layers]
self.bias_list = self.params_hook[layers:]
# if untied weights, expect layers*3 + 1 params
else:
assert len(self.params_hook) == 3*layers + 1, \
"Untied weights: expected {0!s} params, found {1!s}!".format(3*layers+1, len(self.params_hook))
self.weights_list = self.params_hook[:2 * layers]
self.bias_list = self.params_hook[2 * layers:]
# otherwise, construct our params
else:
# initialize a list of weights and biases based on layer_sizes for the GSN
self.weights_list = [
get_weights(
weights_init=weights_init,
shape=(self.layer_sizes[i], self.layer_sizes[i + 1]),
name="W_{0!s}_{1!s}".format(i, i + 1),
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval) for i in range(layers)
]
# add more weights if we aren't tying weights between layers (need to add for higher-lower layers now)
if not tied_weights:
self.weights_list.extend([
get_weights(
weights_init=weights_init,
shape=(self.layer_sizes[i + 1], self.layer_sizes[i]),
name="W_{0!s}_{1!s}".format(i + 1, i),
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in reversed(range(layers))
])
# initialize each layer bias to 0's.
self.bias_list = [
get_bias(shape=(self.layer_sizes[i], ),
name='b_' + str(i),
init_values=bias_init) for i in range(layers + 1)
]
# build the params of the model into a list
self.params = self.weights_list + self.bias_list
log.debug("gsn params: %s", str(self.params))
# using the properties, build the computational graph
self.cost, self.monitors, self.output, self.hiddens = self.build_computation_graph(
)
def decay(self):
new_value = self.initial / (1 + self.reduction_factor * self.epoch)
self.param.set_value(as_floatX(new_value))
self.epoch += 1
def decay(self):
new_value = self.param.get_value() * self.reduction_factor
self.param.set_value(as_floatX(new_value))
def decay(self):
new_value = self.param.get_value() - self.reduction_factor
self.param.set_value(as_floatX(numpy.max([0, new_value])))
def decay(self):
new_value = self.param.get_value()*self.reduction_factor
self.param.set_value(as_floatX(new_value))
def decay(self):
new_value = self.initial / (1 + self.reduction_factor*self.epoch)
self.param.set_value(as_floatX(new_value))
self.epoch += 1
def __init__(self, inputs_hook=None, hiddens_hook=None, params_hook=None, outdir='outputs/gsn/',
input_size=None, hidden_size=1000,
layers=2, walkbacks=4,
visible_activation='sigmoid', hidden_activation='tanh',
input_sampling=True, mrg=RNG_MRG.MRG_RandomStreams(1),
tied_weights=True,
weights_init='uniform', weights_interval='montreal', weights_mean=0, weights_std=5e-3,
bias_init=0.0,
cost_function='binary_crossentropy', cost_args=None,
add_noise=True, noiseless_h1=True,
hidden_noise='gaussian', hidden_noise_level=2, input_noise='salt_and_pepper', input_noise_level=0.4,
noise_decay='exponential', noise_annealing=1,
image_width=None, image_height=None,
**kwargs):
"""
Initialize a GSN.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere.
This is used for linking different models together (e.g. setting the DAE model's hidden layers to the RNN's
output layer gives a generative recurrent model.) For now, it needs to include the shape
information (normally the dimensionality of the hiddens i.e. n_hidden).
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters - such as a training model with dropout applied
to layers and one without for testing, where the parameters are shared between the two.
outdir : str
The directory you want outputs (parameters, images, etc.) to save to. If None, nothing will
be saved.
input_size : int
The size (dimensionality) of the input to the DAE. If shape is provided in `inputs_hook`, this is optional.
The :class:`Model` requires an `output_size`, which gets set to this value because the DAE is an
unsupervised model. The output is a reconstruction of the input.
hidden_size : int
The size (dimensionality) of the hidden layer for the DAE. Generally, you want it to be larger than
`input_size`, which is known as *overcomplete*.
visible_activation : str or callable
The nonlinear (or linear) visible activation to perform after the dot product from hiddens -> visible layer.
This activation function should be appropriate for the input unit types, i.e. 'sigmoid' for binary inputs.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The nonlinear (or linear) hidden activation to perform after the dot product from visible -> hiddens layer.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
layers : int
The number of hidden layers to use.
walkbacks : int
The number of walkbacks to perform (the variable K in Bengio's paper above). A walkback is a Gibbs sample
from the DAE, which means the model generates inputs in sequence, where each generated input is compared
to the original input to create the reconstruction cost for training. For running the model, the very last
generated input in the Gibbs chain is used as the output.
input_sampling : bool
During walkbacks, whether to sample from the generated input to create a new starting point for the next
walkback (next step in the Gibbs chain). This generally makes walkbacks more effective by making the
process more stochastic - more likely to find spurious modes in the model's representation.
mrg : random
A random number generator that is used when adding noise into the network and for sampling from the input.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
tied_weights : bool
DAE has two weight matrices - W from input -> hiddens and V from hiddens -> input. This boolean
determines if V = W.T, which 'ties' V to W and reduces the number of parameters necessary during training.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the reconstruction cost of the model. This should be appropriate
for the type of input, i.e. use 'binary_crossentropy' for binary inputs, or 'mse' for real-valued inputs.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
add_noise : bool
Whether to add noise (corrupt) the input before passing it through the computation graph during training.
This should most likely be set to the default of True, because this is a *denoising* autoencoder after all.
noiseless_h1 : bool
Whether to not add noise (corrupt) the hidden layer during computation.
hidden_noise : str
What type of noise to use for corrupting the hidden layer (if not `noiseless_h1`). See opendeep.utils.noise
for options. This should be appropriate for the hidden unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
hidden_noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
input_noise : str
What type of noise to use for corrupting the input before computation (if `add_noise`).
See opendeep.utils.noise for options. This should be appropriate for the input units, i.e. salt-and-pepper
for binary units, etc.
input_noise_level : float
The amount of noise used to corrupt the input. This could be the masking probability for salt-and-pepper,
standard deviation for Gaussian, interval for Uniform, etc.
noise_decay : str or False
Whether to use `input_noise` scheduling (decay `input_noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the DAE learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_annealing : float
The amount to reduce the `input_noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
image_width : int
If the input should be represented as an image, the width of the input image. If not specified, it will be
close to the square factor of the `input_size`.
image_height : int
If the input should be represented as an image, the height of the input image. If not specified, it will be
close to the square factor of the `input_size`.
"""
# init Model to combine the defaults and config dictionaries with the initial parameters.
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(GSN, self).__init__(**initial_parameters)
# when the input should be thought of as an image, either use the specified width and height,
# or try to make as square as possible.
if image_height is None and image_width is None:
(_h, _w) = closest_to_square_factors(self.input_size)
self.image_width = _w
self.image_height = _h
else:
self.image_height = image_height
self.image_width = image_width
############################
# Theano variables and RNG #
############################
if self.inputs_hook is None:
self.X = T.matrix('X')
else:
# inputs_hook is a (shape, input) tuple
self.X = self.inputs_hook[1]
##########################
# Network specifications #
##########################
# generally, walkbacks should be at least 2*layers
if layers % 2 == 0:
if walkbacks < 2*layers:
log.warning('Not enough walkbacks for the layers! Layers is %s and walkbacks is %s. '
'Generaly want 2X walkbacks to layers',
str(layers), str(walkbacks))
else:
if walkbacks < 2*layers-1:
log.warning('Not enough walkbacks for the layers! Layers is %s and walkbacks is %s. '
'Generaly want 2X walkbacks to layers',
str(layers), str(walkbacks))
self.add_noise = add_noise
self.noise_annealing = as_floatX(noise_annealing) # noise schedule parameter
self.hidden_noise_level = sharedX(hidden_noise_level, dtype=theano.config.floatX)
self.hidden_noise = get_noise(name=hidden_noise, noise_level=self.hidden_noise_level, mrg=mrg)
self.input_noise_level = sharedX(input_noise_level, dtype=theano.config.floatX)
self.input_noise = get_noise(name=input_noise, noise_level=self.input_noise_level, mrg=mrg)
self.walkbacks = walkbacks
self.tied_weights = tied_weights
self.layers = layers
self.noiseless_h1 = noiseless_h1
self.input_sampling = input_sampling
self.noise_decay = noise_decay
# if there was a hiddens_hook, unpack the hidden layers in the tensor
if self.hiddens_hook is not None:
hidden_size = self.hiddens_hook[0]
self.hiddens_flag = True
else:
self.hiddens_flag = False
# determine the sizes of each layer in a list.
# layer sizes, from h0 to hK (h0 is the visible layer)
hidden_size = list(raise_to_list(hidden_size))
if len(hidden_size) == 1:
self.layer_sizes = [self.input_size] + hidden_size * self.layers
else:
assert len(hidden_size) == self.layers, "Hiddens sizes and number of hidden layers mismatch." + \
"Hiddens %d and layers %d" % (len(hidden_size), self.layers)
self.layer_sizes = [self.input_size] + hidden_size
if self.hiddens_hook is not None:
self.hiddens = self.unpack_hiddens(self.hiddens_hook[1])
#########################
# Activation functions! #
#########################
# hidden unit activation
self.hidden_activation = get_activation_function(hidden_activation)
# Visible layer activation
self.visible_activation = get_activation_function(visible_activation)
# make sure the sampling functions are appropriate for the activation functions.
if is_binary(self.visible_activation):
self.visible_sampling = mrg.binomial
else:
# TODO: implement non-binary activation
log.error("Non-binary visible activation not supported yet!")
raise NotImplementedError("Non-binary visible activation not supported yet!")
# Cost function
self.cost_function = get_cost_function(cost_function)
self.cost_args = cost_args or dict()
###############
# Parameters! #
###############
# make sure to deal with params_hook!
if self.params_hook is not None:
# if tied weights, expect layers*2 + 1 params
if self.tied_weights:
assert len(self.params_hook) == 2*layers + 1, \
"Tied weights: expected {0!s} params, found {1!s}!".format(2*layers+1, len(self.params_hook))
self.weights_list = self.params_hook[:layers]
self.bias_list = self.params_hook[layers:]
# if untied weights, expect layers*3 + 1 params
else:
assert len(self.params_hook) == 3*layers + 1, \
"Untied weights: expected {0!s} params, found {1!s}!".format(3*layers+1, len(self.params_hook))
self.weights_list = self.params_hook[:2*layers]
self.bias_list = self.params_hook[2*layers:]
# otherwise, construct our params
else:
# initialize a list of weights and biases based on layer_sizes for the GSN
self.weights_list = [get_weights(weights_init=weights_init,
shape=(self.layer_sizes[i], self.layer_sizes[i+1]),
name="W_{0!s}_{1!s}".format(i, i+1),
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in range(layers)]
# add more weights if we aren't tying weights between layers (need to add for higher-lower layers now)
if not tied_weights:
self.weights_list.extend(
[get_weights(weights_init=weights_init,
shape=(self.layer_sizes[i+1], self.layer_sizes[i]),
name="W_{0!s}_{1!s}".format(i+1, i),
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in reversed(range(layers))]
)
# initialize each layer bias to 0's.
self.bias_list = [get_bias(shape=(self.layer_sizes[i],),
name='b_' + str(i),
init_values=bias_init)
for i in range(layers+1)]
# build the params of the model into a list
self.params = self.weights_list + self.bias_list
log.debug("gsn params: %s", str(self.params))
# using the properties, build the computational graph
self.cost, self.monitors, self.output, self.hiddens = self.build_computation_graph()
def decay(self):
new_value = self.param.get_value() - self.reduction_factor
self.param.set_value(as_floatX(numpy.max([0, new_value])))
