Baillehache Pascal's personal website

LibCapy - nnClassifier

Class implementing a classifier based on a fully connected neural network.

Macros:

#define CAPY_NNCLASSIFIER_H

Enumerations:

None.

Typedefs:

None.

Struct CapyNNClassifier :

Struct CapyNNClassifier's properties:

  CapyClassifierDef;

Inherits CapyClassifier

  CapyNNModel const* nnModel;

Reference to the definition of the neural network model

  CapyNeuralNetwork* nn;

Neural network

  bool verbose;

Verbose mode (default: false)

  CapyPad(bool, 1);

  uint64_t timeTraining;

Time available for training (in second, default: 60, no time limit if 0)

  uint64_t nbIterTrainMax;

Number of iteration available for training (default: 0, no limit if 0)

  CapyDatasetNbRows_t batchSize;

Batch size for training (default: 100)

  uint64_t nbIterTrain;

Counter for iteration during training

  double learnRate;

Step size for the gradient descent (default: 0.1)

  double momentum;

Momentum for the gradient descent (default: 0.1)

  uint64_t seed;

Seed for the pseudo random generator (default: 0)

  double bestLoss;

Best loss during training

  double gradientNorm;

Current gradient norm during training

  double gradientNormEpsilon;

Threshold to stop the training if the gradient is null (default: 1e-6)

  double lossEpsilon;

Threshold to stop the training if the loss is null (default: 1e-6)

  double coeffDampenLearnRate;

Dampening coefficient for the adaptative learning rate (default: 0.5)

  double coeffBoostLearnRate;

Boosting coefficient for the adaptative learning rate (default: 1.1)

Struct CapyNNClassifier's methods:

  void (*destructCapyClassifier)(void);

Destructor

  void (*initParams)(double* const params);

Initialise the parameters value

Input argument(s):

params: the array of parameters
Ouput: The array of parameters is updated

  void (*exportBodyToHtml)(
                 FILE* const stream,
           char const* const title,
    CapyDataset const* const dataset,
                double const expectedAccuracy);

Function herited from the parent to export the body to HTML.

Input argument(s):

stream: the stream where to export
title: the title of the web app
dataset: the training dataset
expectedAccuracy: the expected accuracy of the classifier (in [0,1])

Output and side effect(s):

The and part of a ready to use web app implementing the classifier is written on the stream. The web app can be completed by calling exportToHtml on the classifier to write the