libeblearn
Public Member Functions | Protected Member Functions | Protected Attributes
ebl::module_tester< T > Class Template Reference

#include <ebl_tester.h>

List of all members.

Public Member Functions

 module_tester (double thres=1e-6, T rrange_min=-1.0, T rrange_max=1.0, FILE *out=stdout)
idx< double > test_jacobian (module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out)
idx< double > test_jacobian_param (parameter< bbstate_idx< T > > &p, module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out)
idx< double > test_hessian (module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out)
idx< double > test_hessian_param (parameter< bbstate_idx< T > > &p, module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out)
double get_acc_thres () const
 get accuracy threshold
void set_acc_thres (double acc_thres)
 set accuracy threshold
T get_rrange () const
 get random range
void set_rrange (T rrange)
 set random range
FILE * get_out () const
 get ostream used
void set_out (FILE *out)
 set ostream

Protected Member Functions

void get_jacobian_fprop (module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out, idx< T > &jac)
void get_jacobian_fprop_param (parameter< bbstate_idx< T > > &p, module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out, idx< T > &jac)
void get_jacobian_bprop (module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out, idx< T > &jac)
 Get jacobian using bprop for derivative of 'out' w.r.t. 'in'.
void get_jacobian_bprop_param (parameter< bbstate_idx< T > > &p, module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out, idx< T > &jac)
void get_hessian_fprop (module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out, idx< T > &jac)
void get_hessian_fprop_param (parameter< bbstate_idx< T > > &p, module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out, idx< T > &jac)
void get_hessian_bprop (module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out, idx< T > &jac)
 Get hessian using bprop for derivative of 'out' w.r.t. 'in'.
void get_hessian_bprop_param (parameter< bbstate_idx< T > > &p, module_1_1< T > &m, bbstate_idx< T > &in, bbstate_idx< T > &out, idx< T > &jac)
void report_err (idx< T > &a, idx< T > &b, const char *msg)
 Print error between 'a' and 'b' with message 'msg'.
idx< double > get_errs (idx< T > &a, idx< T > &b)
 Returns a 2-elements idx with 0: max error 1: total error.

Protected Attributes

double acc_thres
 accuracy threshold
T rrange_min
T rrange_max
 random initialization range
FILE * out
 output stream
idx< T > jac_fprop
idx< T > jac_bprop
idx< T > jac_bbprop
 Jacobians from *prop.
idx< T > jac_pfprop
idx< T > jac_pbprop
idx< T > jac_pbbprop
 Jacobians wrt params.
idx< T > hes_fprop
idx< T > hes_bprop
idx< T > hes_bbprop
 Hessians from *prop.
idx< T > hes_pfprop
idx< T > hes_pbprop
idx< T > hes_pbbprop
 Hessians wrt params.

Detailed Description

template<class T>
class ebl::module_tester< T >

Tests any module that is descendant of module_1_1. (includes layers) bprop to input and to parameters are validated with finite difference approximations.

This class is not meant to replace the testing framework for eblearn developers but it should be useful for eblearn users that create new modules using this library as a binary reference

Constructor & Destructor Documentation

template<class T >
ebl::module_tester< T >::module_tester ( double  thres = 1e-6,
T  rrange_min = -1.0,
T  rrange_max = 1.0,
FILE *  out = stdout 
)

out : reference to ostream to push results thres : threshold to decide bprop and finite diff jacobian are same or different this might be tricky because if the modules accumulates derivatives, then the difference will inevitably grow, so user has to have a sense of what is valid default : 1e-8 rrange : range that is used to randomize the input state and parameter objects. if this value is very small, then the jacobians will always be very similar

Member Function Documentation

template<class T >
void ebl::module_tester< T >::get_hessian_bprop_param ( parameter< bbstate_idx< T > > &  p,
module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out,
idx< T > &  jac 
) [protected]

Get hessian using bprop for derivative of 'out' w.r.t. 'p'.

Parameters:
pThe parameter object that was used to allocate the parameters of module. This should contain the parameter of this module only.
template<class T >
void ebl::module_tester< T >::get_hessian_fprop ( module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out,
idx< T > &  jac 
) [protected]

Get hessian using 1st order central finite differnce approximation for derivative of 'out' w.r.t. 'in' using module 'm'.

template<class T >
void ebl::module_tester< T >::get_hessian_fprop_param ( parameter< bbstate_idx< T > > &  p,
module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out,
idx< T > &  jac 
) [protected]

Get hessian using 1st order central finite differnce approximation for derivative of 'out' w.r.t. 'p' using module 'm'.

Parameters:
pThe parameter object that was used to allocate the parameters of module. This should contain the parameter of this module only.
template<class T >
void ebl::module_tester< T >::get_jacobian_bprop_param ( parameter< bbstate_idx< T > > &  p,
module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out,
idx< T > &  jac 
) [protected]

Get jacobian using bprop for derivative of 'out' w.r.t. 'p'.

Parameters:
pThe parameter object that was used to allocate the parameters of module. This should contain the parameter of this module only.
template<class T >
void ebl::module_tester< T >::get_jacobian_fprop ( module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out,
idx< T > &  jac 
) [protected]

Get jacobian using 1st order central finite differnce approximation for derivative of 'out' w.r.t. 'in' using module 'm'.

template<class T >
void ebl::module_tester< T >::get_jacobian_fprop_param ( parameter< bbstate_idx< T > > &  p,
module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out,
idx< T > &  jac 
) [protected]

Get jacobian using 1st order central finite differnce approximation for derivative of 'out' w.r.t. 'p' using module 'm'.

Parameters:
pThe parameter object that was used to allocate the parameters of module. This should contain the parameter of this module only.
template<class T >
idx< double > ebl::module_tester< T >::test_hessian ( module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out 
)

Test the 2nd derivatives of 'in' w.r.t. 'out' with module 'm'. This returns an idx with 2 elements, 0: maximum error 1: total error.

template<class T >
idx< double > ebl::module_tester< T >::test_hessian_param ( parameter< bbstate_idx< T > > &  p,
module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out 
)

Test the 2nd derivatives of 'out' w.r.t. parameters 'p' of module 'm'.

Parameters:
pThe parameter object that was used to allocate the parameters of module. this parameter object should contain the parameter of only this module. This returns an idx with 2 elements, 0: maximum error 1: total error.
template<class T >
idx< double > ebl::module_tester< T >::test_jacobian ( module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out 
)

Test the 1st derivatives of 'in' w.r.t. 'out' with module 'm'. This returns an idx with 2 elements, 0: maximum error 1: total error.

template<class T >
idx< double > ebl::module_tester< T >::test_jacobian_param ( parameter< bbstate_idx< T > > &  p,
module_1_1< T > &  m,
bbstate_idx< T > &  in,
bbstate_idx< T > &  out 
)

Test the 1st derivatives of 'out' w.r.t. parameters 'p' of module 'm'.

Parameters:
pThe parameter object that was used to allocate the parameters of module. this parameter object should contain the parameter of only this module. This returns an idx with 2 elements, 0: maximum error 1: total error.
The documentation for this class was generated from the following files: