A Restricted Boltzmann Machine.
An RBM is an energy based probabilistic model. It consists of two groups of variables: the visible variables \( v \) and the hidden variables \( h \). The key assumption that RBMs make is that the hidden units are conditionally independent given the visible units, and vice versa.
The energy function for RBMs with binary visible units is defined as:
\[ E(v,h) = - b^T v - c^T h - h^T Wv \]
and for RBMs with gaussian (linear) visible units:
\[ E(v,h) = v^T v - b^T v - c^T h - h^T Wv \]
where \( b \) is the bias vector for the visible units, \( c \) is the bias vector for the hidden units, and \( W \) is the weight matrix.
The probability distribution is defined through the energy fucntion as:
\[ P(v,h) = \frac{exp(-E(v,h))}{\sum_{v,h} exp(-E(v,h))} \]
The above definitions along with the independence assumptions result in the following conditionals:
\[ P(h=1|v) = \frac{1}{1+exp(-Wv-c)} \quad \text{for binary hidden units} \]
\[ P(v=1|h) = \frac{1}{1+exp(-W^T h-b)} \quad \text{for binary visible units} \]
\[ P(v|h) \sim \mathcal{N} (W^T h + b,1) \quad \text{for gaussian visible units} \]
Note that when using gaussian visible units, the inputs should be normalized to have zero mean and unity standard deviation.
This class supports having multiple types of visible units in the same RBM. The visible units are divided into groups where each group can have its own type. The hidden units however are just one group of binary units.
Samples can be drawn from the model using Gibbs sampling.
Training is done using contrastive divergence [Hinton, 2002] or persistent contrastive divergence [Tieleman, 2008] (default).
Training progress can be monitored using the reconstruction error (default), which is the average squared difference between a training batch and the RBM's reconstruction of it. The reconstruction is generated using one step of gibbs sampling. Progress can also be monitored using the pseudo-log-likelihood which is an approximation to the log-likelihood. However, this is currently only supported for binary visible units.
The rows of the visible_state matrix are divided into groups, one for each group of visible units. For example, if we have 3 groups of visible units: group 0 with 10 units, group 1 with 5 units, and group 2 with 6 units, the states of group 0 will be stored in visible_state[0:10,:], the states of group 1 will stored in visible_state[10:15,:], and the states of group 2 will be stored in visible_state[15:21,:]. Note that the groups are numbered by the order in which they where added to the RBM using add_visible_group()
Public 成员函数 | |
CRBM () | |
CRBM (int32_t num_hidden) | |
CRBM (int32_t num_hidden, int32_t num_visible, ERBMVisibleUnitType visible_unit_type=RBMVUT_BINARY) | |
virtual | ~CRBM () |
virtual void | add_visible_group (int32_t num_units, ERBMVisibleUnitType unit_type) |
virtual void | initialize (float64_t sigma=0.01) |
virtual void | set_batch_size (int32_t batch_size) |
virtual void | train (CDenseFeatures< float64_t > *features) |
virtual void | sample (int32_t num_gibbs_steps=1, int32_t batch_size=1) |
virtual CDenseFeatures < float64_t > * | sample_group (int32_t V, int32_t num_gibbs_steps=1, int32_t batch_size=1) |
virtual void | sample_with_evidence (int32_t E, CDenseFeatures< float64_t > *evidence, int32_t num_gibbs_steps=1) |
virtual CDenseFeatures < float64_t > * | sample_group_with_evidence (int32_t V, int32_t E, CDenseFeatures< float64_t > *evidence, int32_t num_gibbs_steps=1) |
virtual void | reset_chain () |
virtual float64_t | free_energy (SGMatrix< float64_t > visible, SGMatrix< float64_t > buffer=SGMatrix< float64_t >()) |
virtual void | free_energy_gradients (SGMatrix< float64_t > visible, SGVector< float64_t > gradients, bool positive_phase=true, SGMatrix< float64_t > hidden_mean_given_visible=SGMatrix< float64_t >()) |
virtual void | contrastive_divergence (SGMatrix< float64_t > visible_batch, SGVector< float64_t > gradients) |
virtual float64_t | reconstruction_error (SGMatrix< float64_t > visible, SGMatrix< float64_t > buffer=SGMatrix< float64_t >()) |
virtual float64_t | pseudo_likelihood (SGMatrix< float64_t > visible, SGMatrix< float64_t > buffer=SGMatrix< float64_t >()) |
virtual CDenseFeatures < float64_t > * | visible_state_features () |
virtual SGVector< float64_t > | get_parameters () |
virtual SGMatrix< float64_t > | get_weights (SGVector< float64_t > p=SGVector< float64_t >()) |
virtual SGVector< float64_t > | get_hidden_bias (SGVector< float64_t > p=SGVector< float64_t >()) |
virtual SGVector< float64_t > | get_visible_bias (SGVector< float64_t > p=SGVector< float64_t >()) |
virtual int32_t | get_num_parameters () |
virtual const char * | get_name () const |
virtual CSGObject * | shallow_copy () const |
virtual CSGObject * | deep_copy () const |
virtual bool | is_generic (EPrimitiveType *generic) const |
template<class T > | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
template<> | |
void | set_generic () |
void | unset_generic () |
virtual void | print_serializable (const char *prefix="") |
virtual bool | save_serializable (CSerializableFile *file, const char *prefix="", int32_t param_version=Version::get_version_parameter()) |
virtual bool | load_serializable (CSerializableFile *file, const char *prefix="", int32_t param_version=Version::get_version_parameter()) |
DynArray< TParameter * > * | load_file_parameters (const SGParamInfo *param_info, int32_t file_version, CSerializableFile *file, const char *prefix="") |
DynArray< TParameter * > * | load_all_file_parameters (int32_t file_version, int32_t current_version, CSerializableFile *file, const char *prefix="") |
void | map_parameters (DynArray< TParameter * > *param_base, int32_t &base_version, DynArray< const SGParamInfo * > *target_param_infos) |
void | set_global_io (SGIO *io) |
SGIO * | get_global_io () |
void | set_global_parallel (Parallel *parallel) |
Parallel * | get_global_parallel () |
void | set_global_version (Version *version) |
Version * | get_global_version () |
SGStringList< char > | get_modelsel_names () |
void | print_modsel_params () |
char * | get_modsel_param_descr (const char *param_name) |
index_t | get_modsel_param_index (const char *param_name) |
void | build_gradient_parameter_dictionary (CMap< TParameter *, CSGObject * > *dict) |
virtual void | update_parameter_hash () |
virtual bool | parameter_hash_changed () |
virtual bool | equals (CSGObject *other, float64_t accuracy=0.0, bool tolerant=false) |
virtual CSGObject * | clone () |
Public 属性 | |
int32_t | cd_num_steps |
bool | cd_persistent |
bool | cd_sample_visible |
float64_t | l2_coefficient |
float64_t | l1_coefficient |
int32_t | monitoring_interval |
ERBMMonitoringMethod | monitoring_method |
int32_t | max_num_epochs |
int32_t | gd_mini_batch_size |
float64_t | gd_learning_rate |
float64_t | gd_learning_rate_decay |
float64_t | gd_momentum |
SGMatrix< float64_t > | hidden_state |
SGMatrix< float64_t > | visible_state |
SGIO * | io |
Parallel * | parallel |
Version * | version |
Parameter * | m_parameters |
Parameter * | m_model_selection_parameters |
Parameter * | m_gradient_parameters |
ParameterMap * | m_parameter_map |
uint32_t | m_hash |
Protected 成员函数 | |
virtual void | mean_hidden (SGMatrix< float64_t > visible, SGMatrix< float64_t > result) |
virtual void | mean_visible (SGMatrix< float64_t > hidden, SGMatrix< float64_t > result) |
virtual void | sample_hidden (SGMatrix< float64_t > mean, SGMatrix< float64_t > result) |
virtual void | sample_visible (SGMatrix< float64_t > mean, SGMatrix< float64_t > result) |
virtual void | sample_visible (int32_t index, SGMatrix< float64_t > mean, SGMatrix< float64_t > result) |
virtual TParameter * | migrate (DynArray< TParameter * > *param_base, const SGParamInfo *target) |
virtual void | one_to_one_migration_prepare (DynArray< TParameter * > *param_base, const SGParamInfo *target, TParameter *&replacement, TParameter *&to_migrate, char *old_name=NULL) |
virtual void | load_serializable_pre () throw (ShogunException) |
virtual void | load_serializable_post () throw (ShogunException) |
virtual void | save_serializable_pre () throw (ShogunException) |
virtual void | save_serializable_post () throw (ShogunException) |
Protected 属性 | |
int32_t | m_num_hidden |
int32_t | m_num_visible |
int32_t | m_batch_size |
int32_t | m_num_visible_groups |
CDynamicArray< int32_t > * | m_visible_group_types |
CDynamicArray< int32_t > * | m_visible_group_sizes |
CDynamicArray< int32_t > * | m_visible_state_offsets |
int32_t | m_num_params |
SGVector< float64_t > | m_params |
友元 | |
class | CDeepBeliefNetwork |
CRBM | ( | int32_t | num_hidden | ) |
Constructs an RBM with no visible units. The visible units can be added later using add_visible_group()
num_hidden | Number of hidden units |
CRBM | ( | int32_t | num_hidden, |
int32_t | num_visible, | ||
ERBMVisibleUnitType | visible_unit_type = RBMVUT_BINARY |
||
) |
|
virtual |
|
inherited |
Builds a dictionary of all parameters in SGObject as well of those of SGObjects that are parameters of this object. Dictionary maps parameters to the objects that own them.
dict | dictionary of parameters to be built. |
在文件 SGObject.cpp 第 1243 行定义.
|
virtualinherited |
Creates a clone of the current object. This is done via recursively traversing all parameters, which corresponds to a deep copy. Calling equals on the cloned object always returns true although none of the memory of both objects overlaps.
在文件 SGObject.cpp 第 1360 行定义.
|
virtual |
Computes the gradients using contrastive divergence
visible_batch | States of the visible units |
gradients | Array in which the results are stored. Length get_num_parameters() |
|
virtualinherited |
A deep copy. All the instance variables will also be copied.
在文件 SGObject.cpp 第 200 行定义.
Recursively compares the current SGObject to another one. Compares all registered numerical parameters, recursion upon complex (SGObject) parameters. Does not compare pointers!
May be overwritten but please do with care! Should not be necessary in most cases.
other | object to compare with |
accuracy | accuracy to use for comparison (optional) |
tolerant | allows linient check on float equality (within accuracy) |
在文件 SGObject.cpp 第 1264 行定义.
|
virtual |
Computes the average free energy on a given batch of visible unit states.
The free energy for a vector \( v \) is defined as:
\[ F(v) = - log(\sum_h exp(-E(v,h)) \]
which yields the following (in vectorized form):
\[ F(v) = -b^T v - \sum log(1+exp(Wv+c)) \quad \text{for binary visible units}\]
\[ F(v) = \frac{1}{2} v^T v - b^T v - \sum log(1+exp(Wv+c)) \quad \text{for gaussian visible units}\]
visible | States of the visible units |
buffer | A matrix of size num_hidden*batch_size. used as a buffer during computation. If not given, a new matrix is allocated and used as a buffer. |
|
virtual |
Computes the gradients of the free energy function with respect to the RBM's parameters
visible | States of the visible units |
gradients | Array in which the results are stored. Length get_num_parameters() |
positive_phase | If true, the result vector is reset to zero and the gradients are added to it with a positive sign. If false, the result vector is not reset and the gradients are added to it with a negative sign. This is useful during contrastive divergence. |
hidden_mean_given_visible | Means of the hidden states given the visible states. If not given, means will be computed by calling mean_hidden() |
|
inherited |
|
inherited |
|
inherited |
|
inherited |
在文件 SGObject.cpp 第 1135 行定义.
|
inherited |
Returns description of a given parameter string, if it exists. SG_ERROR otherwise
param_name | name of the parameter |
在文件 SGObject.cpp 第 1159 行定义.
|
inherited |
Returns index of model selection parameter with provided index
param_name | name of model selection parameter |
在文件 SGObject.cpp 第 1172 行定义.
|
virtual |
|
virtual |
|
virtual |
|
virtualinherited |
If the SGSerializable is a class template then TRUE will be returned and GENERIC is set to the type of the generic.
generic | set to the type of the generic if returning TRUE |
在文件 SGObject.cpp 第 297 行定义.
|
inherited |
maps all parameters of this instance to the provided file version and loads all parameter data from the file into an array, which is sorted (basically calls load_file_parameter(...) for all parameters and puts all results into a sorted array)
file_version | parameter version of the file |
current_version | version from which mapping begins (you want to use Version::get_version_parameter() for this in most cases) |
file | file to load from |
prefix | prefix for members |
在文件 SGObject.cpp 第 704 行定义.
|
inherited |
loads some specified parameters from a file with a specified version The provided parameter info has a version which is recursively mapped until the file parameter version is reached. Note that there may be possibly multiple parameters in the mapping, therefore, a set of TParameter instances is returned
param_info | information of parameter |
file_version | parameter version of the file, must be <= provided parameter version |
file | file to load from |
prefix | prefix for members |
在文件 SGObject.cpp 第 545 行定义.
|
virtualinherited |
Load this object from file. If it will fail (returning FALSE) then this object will contain inconsistent data and should not be used!
file | where to load from |
prefix | prefix for members |
param_version | (optional) a parameter version different to (this is mainly for testing, better do not use) |
在文件 SGObject.cpp 第 374 行定义.
|
protectedvirtualinherited |
Can (optionally) be overridden to post-initialize some member variables which are not PARAMETER::ADD'ed. Make sure that at first the overridden method BASE_CLASS::LOAD_SERIALIZABLE_POST is called.
ShogunException | will be thrown if an error occurs. |
被 CKernel, CWeightedDegreePositionStringKernel, CList, CAlphabet, CLinearHMM, CGaussianKernel, CInverseMultiQuadricKernel, CCircularKernel , 以及 CExponentialKernel 重载.
在文件 SGObject.cpp 第 1062 行定义.
|
protectedvirtualinherited |
Can (optionally) be overridden to pre-initialize some member variables which are not PARAMETER::ADD'ed. Make sure that at first the overridden method BASE_CLASS::LOAD_SERIALIZABLE_PRE is called.
ShogunException | will be thrown if an error occurs. |
被 CDynamicArray< T >, CDynamicArray< float64_t >, CDynamicArray< float32_t >, CDynamicArray< int32_t >, CDynamicArray< char >, CDynamicArray< bool > , 以及 CDynamicObjectArray 重载.
在文件 SGObject.cpp 第 1057 行定义.
|
inherited |
Takes a set of TParameter instances (base) with a certain version and a set of target parameter infos and recursively maps the base level wise to the current version using CSGObject::migrate(...). The base is replaced. After this call, the base version containing parameters should be of same version/type as the initial target parameter infos. Note for this to work, the migrate methods and all the internal parameter mappings have to match
param_base | set of TParameter instances that are mapped to the provided target parameter infos |
base_version | version of the parameter base |
target_param_infos | set of SGParamInfo instances that specify the target parameter base |
在文件 SGObject.cpp 第 742 行定义.
|
protectedvirtualinherited |
creates a new TParameter instance, which contains migrated data from the version that is provided. The provided parameter data base is used for migration, this base is a collection of all parameter data of the previous version. Migration is done FROM the data in param_base TO the provided param info Migration is always one version step. Method has to be implemented in subclasses, if no match is found, base method has to be called.
If there is an element in the param_base which equals the target, a copy of the element is returned. This represents the case when nothing has changed and therefore, the migrate method is not overloaded in a subclass
param_base | set of TParameter instances to use for migration |
target | parameter info for the resulting TParameter |
在文件 SGObject.cpp 第 949 行定义.
|
protectedvirtualinherited |
This method prepares everything for a one-to-one parameter migration. One to one here means that only ONE element of the parameter base is needed for the migration (the one with the same name as the target). Data is allocated for the target (in the type as provided in the target SGParamInfo), and a corresponding new TParameter instance is written to replacement. The to_migrate pointer points to the single needed TParameter instance needed for migration. If a name change happened, the old name may be specified by old_name. In addition, the m_delete_data flag of to_migrate is set to true. So if you want to migrate data, the only thing to do after this call is converting the data in the m_parameter fields. If unsure how to use - have a look into an example for this. (base_migration_type_conversion.cpp for example)
param_base | set of TParameter instances to use for migration |
target | parameter info for the resulting TParameter |
replacement | (used as output) here the TParameter instance which is returned by migration is created into |
to_migrate | the only source that is used for migration |
old_name | with this parameter, a name change may be specified |
在文件 SGObject.cpp 第 889 行定义.
|
virtualinherited |
在文件 SGObject.cpp 第 263 行定义.
|
inherited |
prints all parameter registered for model selection and their type
在文件 SGObject.cpp 第 1111 行定义.
|
virtualinherited |
|
virtual |
Computes an approximation to the pseudo-likelihood. See this tutorial for more details. Only works with binary visible units
visible | States of the visible units |
buffer | A matrix of size num_visible*batch_size. used as a buffer during computation. If not given, a new matrix is allocated and used as a buffer. |
return | Approximation to the average pseudo-likelihood over the given batch |
|
virtual |
|
virtual |
Draws samples from the marginal distribution of the visible units using Gibbs sampling. The sampling starts from the values in the RBM's visible_state matrix and result of the sampling is stored there too.
num_gibbs_steps | Number of Gibbs sampling steps |
batch_size | Number of samples to be drawn. A seperate chain is used for each sample |
|
virtual |
Draws Samples from \( P(V) \) where \( V \) is one of the visible unit groups. The sampling starts from the values in the RBM's visible_state matrix and result of the sampling is stored there too.
V | Index of the visible unit group to be sampled |
num_gibbs_steps | Number of Gibbs sampling steps |
batch_size | Number of samples to be drawn. A seperate chain is used for each sample |
|
virtual |
Draws Samples from \( P(V|E=evidence) \) where \( E \) is one of the visible unit groups and \( V \) is another visible unit group. The sampling starts from the values in the RBM's visible_state matrix and result of the sampling is stored there too.
V | Index of the visible unit group to be sampled |
E | Index of the evidence visible unit group |
evidence | States of the evidence visible unit group |
num_gibbs_steps | Number of Gibbs sampling steps |
|
virtual |
Draws Samples from \( P(V|E=evidence) \) where \( E \) is one of the visible unit groups and \( V \) is all the visible unit excluding the ones in group \( E \). The sampling starts from the values in the RBM's visible_state matrix and result of the sampling is stored there too.
E | Index of the evidence visible unit group |
evidence | States of the evidence visible unit group |
num_gibbs_steps | Number of Gibbs sampling steps |
|
virtualinherited |
Save this object to file.
file | where to save the object; will be closed during returning if PREFIX is an empty string. |
prefix | prefix for members |
param_version | (optional) a parameter version different to (this is mainly for testing, better do not use) |
在文件 SGObject.cpp 第 315 行定义.
|
protectedvirtualinherited |
Can (optionally) be overridden to post-initialize some member variables which are not PARAMETER::ADD'ed. Make sure that at first the overridden method BASE_CLASS::SAVE_SERIALIZABLE_POST is called.
ShogunException | will be thrown if an error occurs. |
被 CKernel 重载.
在文件 SGObject.cpp 第 1072 行定义.
|
protectedvirtualinherited |
Can (optionally) be overridden to pre-initialize some member variables which are not PARAMETER::ADD'ed. Make sure that at first the overridden method BASE_CLASS::SAVE_SERIALIZABLE_PRE is called.
ShogunException | will be thrown if an error occurs. |
被 CKernel, CDynamicArray< T >, CDynamicArray< float64_t >, CDynamicArray< float32_t >, CDynamicArray< int32_t >, CDynamicArray< char >, CDynamicArray< bool > , 以及 CDynamicObjectArray 重载.
在文件 SGObject.cpp 第 1067 行定义.
|
virtual |
|
inherited |
在文件 SGObject.cpp 第 42 行定义.
|
inherited |
在文件 SGObject.cpp 第 47 行定义.
|
inherited |
在文件 SGObject.cpp 第 52 行定义.
|
inherited |
在文件 SGObject.cpp 第 57 行定义.
|
inherited |
在文件 SGObject.cpp 第 62 行定义.
|
inherited |
在文件 SGObject.cpp 第 67 行定义.
|
inherited |
在文件 SGObject.cpp 第 72 行定义.
|
inherited |
在文件 SGObject.cpp 第 77 行定义.
|
inherited |
在文件 SGObject.cpp 第 82 行定义.
|
inherited |
在文件 SGObject.cpp 第 87 行定义.
|
inherited |
在文件 SGObject.cpp 第 92 行定义.
|
inherited |
在文件 SGObject.cpp 第 97 行定义.
|
inherited |
在文件 SGObject.cpp 第 102 行定义.
|
inherited |
在文件 SGObject.cpp 第 107 行定义.
|
inherited |
在文件 SGObject.cpp 第 112 行定义.
|
inherited |
set generic type to T
|
inherited |
|
inherited |
|
inherited |
|
virtualinherited |
A shallow copy. All the SGObject instance variables will be simply assigned and SG_REF-ed.
被 CGaussianKernel 重载.
在文件 SGObject.cpp 第 194 行定义.
|
virtual |
|
inherited |
unset generic type
this has to be called in classes specializing a template class
在文件 SGObject.cpp 第 304 行定义.
|
virtualinherited |
Updates the hash of current parameter combination
在文件 SGObject.cpp 第 250 行定义.
|
virtual |
Returns the states of the visible unit as CDenseFeatures<float64_t>
|
friend |
int32_t cd_num_steps |
bool cd_persistent |
bool cd_sample_visible |
float64_t gd_learning_rate_decay |
int32_t gd_mini_batch_size |
float64_t gd_momentum |
|
inherited |
io
在文件 SGObject.h 第 496 行定义.
|
inherited |
parameters wrt which we can compute gradients
在文件 SGObject.h 第 511 行定义.
|
inherited |
Hash of parameter values
在文件 SGObject.h 第 517 行定义.
|
inherited |
model selection parameters
在文件 SGObject.h 第 508 行定义.
|
inherited |
map for different parameter versions
在文件 SGObject.h 第 514 行定义.
|
inherited |
parameters
在文件 SGObject.h 第 505 行定义.
|
protected |
|
protected |
|
protected |
int32_t max_num_epochs |
int32_t monitoring_interval |
ERBMMonitoringMethod monitoring_method |
|
inherited |
parallel
在文件 SGObject.h 第 499 行定义.
|
inherited |
version
在文件 SGObject.h 第 502 行定义.