class DiagonalGMM

This class can be used to model Diagonal Gaussian Mixture Models.

Inheritance:

Public Fields

int n_gaussians

number of Gaussians in the mixture

real prior_weights

prior weights of the Gaussians, used in EM to give a small prior on each Gaussian

EMTrainer* initial_kmeans_trainer

optional initializations if nothing is given, then random calling by reset(), at your own risks or model loaded by the user.

MeasurerList* initial_kmeans_trainer_measurers

as well as a measurer of this trainer

real* log_weights

the pointers to the parameters

real* dlog_weights

the pointers to the derivative of the parameters

real* var_threshold

this contains the minimal value of each variance

Sequence* log_probabilities_g

for each frame, for each gaussian, keep its log probability

int best_gauss

gaussian that maximize the observed frame

Sequence* best_gauss_per_frame

gaussian that maximize the observed frame

real* sum_log_var_plus_n_obs_log_2_pi

in order to faster the computation, we can do some "pre-computation" pre-computed sum_log_var + n_obs * log_2_pi

real** minus_half_over_var

pre-computed -05 / var

real** means_acc

accumulators for EM

Public Methods

DiagonalGMM(int n_inputs_, int n_gaussians_, EMTrainer* initial_kmeans_trainer_ = NULL)

virtual void reset()

This methods have to call by the user to initialized the random parameters

virtual void setDataSet(DataSet* data_)

If the KmeanTrainer has been given to the constructor, Kmeans algorithm is perform

virtual void setVarThreshold(real* var_threshold_)

Set the minimum value for the variances

virtual void eMIterInitialize()

Methods used to initialize the model at the beginning of each EM iteration

virtual void iterInitialize()

Methods used to initialize the model at the beginning of each gradient descent iteration

virtual real frameLogProbability(int t, real* inputs)

Returns the log probability of a frame of a sequence

virtual real viterbiFrameLogProbability(int t, real* inputs)

Returns the log probability of a frame of a sequence on viterbi mode

virtual real frameLogProbabilityOneGaussian(int g, real* inputs)

this method returns the log probability of the "g" Gaussian

virtual void sequenceInitialize(Sequence* inputs)

Methods used to initialize the model at the beginning of each example during gradient descent training

virtual void eMSequenceInitialize(Sequence* inputs)

Methods used to initialize the model at the beginning of each example during EM training

virtual void frameEMAccPosteriors(int t, real* inputs, real log_posterior)

The backward step of EM for a frame

virtual void frameViterbiAccPosteriors(int t, real* inputs, real log_posterior)

The backward step of Viterbi for a frame

virtual void eMUpdate()

The update after each iteration for EM

virtual void update()

The update after each iteration for gradient

virtual void frameBackward(int t, real* f_inputs, real* beta_, real* f_outputs, real* alpha_)

Same as backward, but for one frame only

virtual void frameDecision(int t, real* decision)

Returns the expected value in decision

Inherited from Distribution:

Public Fields

real log_probability

Sequence* log_probabilities

Public Methods

virtual real logProbability(Sequence* inputs)

virtual real viterbiLogProbability(Sequence* inputs)

virtual void eMAccPosteriors(Sequence* inputs, real log_posterior)

virtual void viterbiAccPosteriors(Sequence* inputs, real log_posterior)

virtual void decode(Sequence* inputs)

virtual void eMForward(Sequence* inputs)

virtual void viterbiForward(Sequence* inputs)

virtual void viterbiBackward(Sequence* inputs, Sequence* alpha)

Public Members

Returns the decision of the distribution

Inherited from GradientMachine:

Public Fields

int n_inputs

int n_outputs

Parameters* params

Parameters* der_params

Sequence* beta

Public Methods

virtual void forward(Sequence* inputs)

virtual void backward(Sequence* inputs, Sequence* alpha)

virtual void setPartialBackprop(bool flag=true)

virtual void frameForward(int t, real* f_inputs, real* f_outputs)

virtual void loadXFile(XFile* file)

virtual void saveXFile(XFile* file)

Inherited from Machine:

Public Fields

Sequence* outputs

Inherited from Object:

Public Fields

Allocator* allocator

Public Methods

void addOption(const char* name, int size, void* ptr, const char* help="")

void addIOption(const char* name, int* ptr, int init_value, const char* help="")

void addROption(const char* name, real* ptr, real init_value, const char* help="")

void addBOption(const char* name, bool* ptr, bool init_value, const char* help="")

void addOOption(const char* name, Object** ptr, Object* init_value, const char* help="")

void setOption(const char* name, void* ptr)

void setIOption(const char* name, int option)

void setROption(const char* name, real option)

void setBOption(const char* name, bool option)

void setOOption(const char* name, Object* option)

void load(const char* filename)

void save(const char* filename)

void* operator new(size_t size, Allocator* allocator_=NULL)

void* operator new(size_t size, Allocator* allocator_, void* ptr_)

void operator delete(void* ptr)

Documentation

This class can be used to model Diagonal Gaussian Mixture Models. They can be trained using either EM (with EMTrainer) or gradient descent (with GMTrainer).

int n_gaussians

number of Gaussians in the mixture

real prior_weights

prior weights of the Gaussians, used in EM to give a small prior on each Gaussian

EMTrainer* initial_kmeans_trainer

optional initializations if nothing is given, then random calling by reset(), at your own risks or model loaded by the user. one can give a initial trainer containing a kmeans

MeasurerList* initial_kmeans_trainer_measurers

as well as a measurer of this trainer

real* log_weights

the pointers to the parameters

real* dlog_weights

the pointers to the derivative of the parameters

real* var_threshold

this contains the minimal value of each variance

Sequence* log_probabilities_g

for each frame, for each gaussian, keep its log probability

int best_gauss

gaussian that maximize the observed frame

Sequence* best_gauss_per_frame

gaussian that maximize the observed frame

real* sum_log_var_plus_n_obs_log_2_pi

in order to faster the computation, we can do some "pre-computation" pre-computed sum_log_var + n_obs * log_2_pi

real** minus_half_over_var

pre-computed -05 / var

real** means_acc

accumulators for EM

DiagonalGMM(int n_inputs_, int n_gaussians_, EMTrainer* initial_kmeans_trainer_ = NULL)

virtual void reset()

This methods have to call by the user to initialized the random parameters

virtual void setDataSet(DataSet* data_)

If the KmeanTrainer has been given to the constructor, Kmeans algorithm is perform

virtual void setVarThreshold(real* var_threshold_)

Set the minimum value for the variances

virtual void eMIterInitialize()

Methods used to initialize the model at the beginning of each EM iteration

virtual void iterInitialize()

Methods used to initialize the model at the beginning of each gradient descent iteration

virtual real frameLogProbability(int t, real* inputs)

Returns the log probability of a frame of a sequence

virtual real viterbiFrameLogProbability(int t, real* inputs)

Returns the log probability of a frame of a sequence on viterbi mode

virtual real frameLogProbabilityOneGaussian(int g, real* inputs)

this method returns the log probability of the "g" Gaussian

virtual void sequenceInitialize(Sequence* inputs)

Methods used to initialize the model at the beginning of each example during gradient descent training

virtual void eMSequenceInitialize(Sequence* inputs)

Methods used to initialize the model at the beginning of each example during EM training

virtual void frameEMAccPosteriors(int t, real* inputs, real log_posterior)

The backward step of EM for a frame

virtual void frameViterbiAccPosteriors(int t, real* inputs, real log_posterior)

The backward step of Viterbi for a frame

virtual void eMUpdate()

The update after each iteration for EM

virtual void update()

The update after each iteration for gradient

virtual void frameBackward(int t, real* f_inputs, real* beta_, real* f_outputs, real* alpha_)

Same as backward, but for one frame only

virtual void frameDecision(int t, real* decision)

Returns the expected value in decision

Direct child classes:

MAPDiagonalGMM
KMeans

Author:

Samy Bengio (bengio@idiap.ch) Johnny Mariethoz (Johnny.Mariethoz@idiap.ch)

Alphabetic index HTML hierarchy of classes or Java