PrePostProcessing {SBCK}

PrePostProcessing base class

Description

Base class to pre/post process data before/after a bias correction

Details

This base class can be considered as the identity pre-post processing, and is used to be herited by others pre/post processing class. The key ideas are:
- A PrePostProcessing based class contains a bias correction method, initalized by the 'bc_method' argument, always available for all herited class
- The 'pipe' keyword is a list of pre/post processing class, applied one after the other.

Try with an example, start with a dataset similar to tas/pr:
>>> XY = SBCK::dataset_like_tas_pr(2000)
>>> X0 = XY$X0
>>> X1 = XY$X1
>>> Y0 = XY$Y0

The first column is Gaussian, but the second is an exponential law with a Dirac mass at 0, represented the 0 of precipitations. For a quantile mapping correction in the calibration period, we just apply
>>> qm = SBCK::QM$new()
>>> qm$fit(Y0,X0)
>>> Z0 = qm$predict(X0)

Now, if we want to pre-post process with the SSR method (0 are replaced by random values between 0 (excluded) and the minimal non zero value), we write:
>>> ppp = SBCK::PPPSSR$new( bc_method = QM , cols = 2 )
>>> ppp$fit(Y0,X0)
>>> Z0 = ppp$predict(X0)

The SSR approach is applied only on the second column (the precipitation), and the syntax is the same than for a simple bias correction method.

Imagine now that we want to apply the SSR, and to ensure the positivity of CDFt for precipitation, we also want to use the LogLinLink pre-post processing method. This can be done with the following syntax:
>>> ppp = PPPLogLinLink$new( bc_method = CDFt , cols = 2 ,
>>> pipe = list(PPPSSR) ,
>>> pipe_kwargs = list( list(cols = 2) ) )
>>> ppp$fit(Y0,X0,X1)
>>> Z = ppp$predict(X1,X0)

With this syntax, the pre processing operation is PPPLogLinLink$transform(PPPSSR$transform(data)) and post processing operation PPPSSR$itransform(PPPLogLinLink$itransform(bc_data)). So the formula can read from right to left (as the mathematical composition). Note it is equivalent to define:
>>> ppp = PrePostProcessing$new( bc_method = CDFt,
>>> pipe = list(PPPLogLinLink,PPPSSR),
>>> pipe_kwargs = list( list(cols=2) , list(cols=2) ) )

Methods

Public methods

• PrePostProcessing$new()

• PrePostProcessing$transform()

• PrePostProcessing$itransform()

• PrePostProcessing$fit()

• PrePostProcessing$predict()

• PrePostProcessing$clone()

Method new()

Create a new PrePostProcessing object.

Usage

PrePostProcessing$new(
  bc_method = NULL,
  bc_method_kwargs = list(),
  pipe = list(),
  pipe_kwargs = list()
)

Arguments

bc_method

The bias correction method

bc_method_kwargs

Dict of keyword arguments passed to bc_method

pipe

list of others PrePostProcessing class to pipe

pipe_kwargs

list of list of keyword arguments passed to each elements of pipe

Returns

A new 'PrePostProcessing' object.

Method transform()

Transformation applied to data before the bias correction. Just the identity for this class

Usage

PrePostProcessing$transform(X)

Arguments

X

[matrix: n_samples * n_features]

Returns

Xt [matrix: n_samples * n_features]

Method itransform()

Transformation applied to data after the bias correction. Just the identity for this class

Usage

PrePostProcessing$itransform(Xt)

Arguments

Xt

[matrix: n_samples * n_features]

Returns

X [matrix: n_samples * n_features]

Method fit()

Apply the pre processing and fit the bias correction method. If X1 is NULL, the method is considered as stationary

Usage

PrePostProcessing$fit(Y0, X0, X1 = NULL)

Arguments

Y0

[matrix: n_samples * n_features] Observations in calibration

X0

[matrix: n_samples * n_features] Model in calibration

X1

[matrix: n_samples * n_features] Model in projection

Returns

NULL

Method predict()

Predict the correction, apply pre-processing before, and post-processing after

Usage

PrePostProcessing$predict(X1 = NULL, X0 = NULL)

Arguments

X1

[matrix: n_samples * n_features or NULL] Model in projection

X0

[matrix: n_samples * n_features or NULL] Model in calibration

Returns

[matrix or list] Return the matrix of correction of X1 if X0 is NULL (and vice-versa), else return a list containing Z1 and Z0, the corrections of X1 and X0

Method clone()

The objects of this class are cloneable with this method.

Usage

PrePostProcessing$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

Examples

## Start with data
XY = SBCK::dataset_like_tas_pr(2000)
X0 = XY$X0
X1 = XY$X1
Y0 = XY$Y0

## Define pre/post processing method
ppp = PrePostProcessing$new( bc_method = CDFt,
                             pipe = list(PPPLogLinLink,PPPSSR),
                             pipe_kwargs = list( list(cols=2) , list(cols=2) ) )

## Bias correction
ppp$fit(Y0,X0,X1)
Z = ppp$predict(X1,X0)