[Insight-users] AdvanceOneStep in ClassGradientDescentOptimizer(itkGradientDescentOptimizer.cpp)

[Karthik Krishnan]

Cause different parameters might have different degrees of influence and the
scales are used to dictate those degrees. So a large scale for a paramter
will restrict influence of that parameter. See Page 268 of the
SoftwareGuide2.0 at http://itk.org/HTML/Documentation.htm

Thanks
kk
  -----Original Message-----
  From: insight-users-bounces at itk.org
[mailto:insight-users-bounces at itk.org]On Behalf Of 庄吓海(Xiahai Zhuang)
  Sent: Sunday, May 01, 2005 5:00 AM
  To: insight-users at itk.org
  Subject: [Insight-users] AdvanceOneStep in
ClassGradientDescentOptimizer(itkGradientDescentOptimizer.cpp)


  /**
   * Advance one Step following the gradient direction
   */
    typedef Array< double>              ParametersType;
   ParametersType m_currentPosition;
  void
  GradientDescentOptimizer
  ::AdvanceOneStep( void )
  {

    itkDebugMacro("AdvanceOneStep");
  .....

    ScalesType scales = this-> GetScales();
  ...

    DerivativeType transformedGradient( spaceDimension );

    for(unsigned int j = 0; j <  spaceDimension; j++)
      {
      transformedGradient[j] = m_Gradient[j] / scales[j];
//?????????????????????? why here takes a (/scales[j])
      }
//as in method m_CostFunction-> GetValueAndDerivative(this->
GetCurrentPosition(), m_Value, m_Gradient )m_Gradient derivative

// m_Gradient is the derivative of MutualInformation with respect to
transform parameters,

// why transformedGradient[j] = m_Gradient[j] / scales[j];
    ParametersType newPosition( spaceDimension );
    for(unsigned int j = 0; j <  spaceDimension; j++)
      {
      newPosition[j] = currentPosition[j] +
        direction * m_LearningRate * transformedGradient[j];
      }

    this-> SetCurrentPosition( newPosition );

    this-> InvokeEvent( IterationEvent() );

  }


  　　　　　　　　致
  礼！


  　　　　　　　　庄吓海(Xiahai Zhuang)
  　　　　　　　　arhye at 163.com
  　　　　　　　　　　2005-05-01
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://public.kitware.com/pipermail/insight-users/attachments/20050501/51db1ba2/attachment.html