NEWS | R Documentation |
The operations that involve fitting many PARAFAC models
(feemsplithalf
and feemjackknife
) now handle the
arguments parallel
and cl
on their own. If parallel
processing is enabled, they schedule the calls to
feemparafac(...)
on the cluster. This makes it possible to
run more than nstart
(10 by default) models at the same time
and avoid the wait for one remaining random start before moving on
to the next model. Due to implementation differences, this also
avoids the need to call clusterEvalQ(., library(multiway))
before performing parallel processing.
The default stopping criterion in feemparafac
has been
changed from 10^{-4}
to 10^{-6}
. The
previous value caused misleadingly pessimistic split-half results
for real world datasets, failing the comparisons due to the models
being underfitted.
The feemscale
function now raises an error when the
returned scale factor turns out to be NA
.
The plot.feemparafac
method used to sort the components
lexicographically, making it inconvenient to read plots for models
with more than 9 factors. The problem is now fixed, with
coef(type = 'surfaces')
returning the data.frame
used
to construct the plot.
The feem(..., format = 'table')
method now treats the
dec
argument specially, making it possible to read FEEM
tables with commas as decimal separators.
The coef.feemsplithalf(kind = 'factors')
method now
avoids repeatedly calling feemcube.feemparafac()
, which is
expensive for large feemcube
s due to having to subset them.
This speeds up all related plots.
New split-half plot and coefficient output,
kind = 'aggtcc'
, taking the minimal TCC values over
individual components and plotting their distributions as a boxplot
combined with a violin plot.
New split-half plot and coefficient output,
kind = 'bandfactors'
, summarising the distributions of the factor
value over the different subsets as estimates and bands.
The operations that involve fitting many PARAFAC models
(feemsplithalf
and feemjackknife
) now adjust the signs
of the loadings to be as positive as possible (using
resign(absorb = 'C')
) before matching the components. This
makes it possible to validate models with loadings not constrained
to nonnegativity.
The feemparafac
function now returns the time it took
to compute the decomposition in the time
attribute of the
object. Depending on whether a parallel cluster or a parallel
BLAS has been used, most of the time may be spent in elapsed
(waiting for the cluster), user.self
(computing in more than
one thread), or even sys.self
(BLAS threads spinning in
sched_yield()
).
New function feemcorcondia
to perform the core
consistency diagnostic (Bro & Kiers, 2003) on PARAFAC models, with
added support for missing data.
The feemflame
function gains arguments
control.parafac
and control.cmf
, making it possible to
adjust the termination criteria of the underlying methods.
The previous version of the package was accidentally broken on
versions of R where apply()
did not yet have a
simplify
argument. More strict release testing procedures
will be applied in the future.
New argument like
to reorder.feemparafac
and
rescale.feemparafac
makes it possible to order and scale the
components to match another model. The package now re-exports the
rescale
generic from the multiway package.
Fixed an error in feemscatter(method='whittaker')
where
the nonneg
parameter could be ignored (turning the
non-negativity protection off) if the wavelength grid was not sorted.
This mistake has been present since version 0.3-0, where Whittaker
smoothing had been introduced.
Inverted the luminance of the diverging.colours
palette,
setting it to brighter for lower-magnitude values. The signs of the
residuals are much easier to understand now.
The feemcube.feemflame
method is now actually exported.
The plot()
method for feemflame
objects now
correctly forwards its arguments to lattice functions for
type = 'both'
.
New function absindex
computes absorption coefficients,
absorbance values, spectral slopes and their ratios.
When importing files with a non-default encoding,
feem.character
now avoids leaking a connection per file. (They
were cleaned up anyway by the garbage collector, but with a warning.)
The custom progress bar in feemflame
is now re-drawn
correctly on R GUIs.
On R ≥ 4.2, equations in HTML help are typeset from LaTeX source, not manually translated HTML.
t()
on feem
objects now results in an unclassed
matrix, not a broken feem
object.
feemscatter
now accepts scattering bands with asymmetric
widths. On some fluorometers, this is needed in order to remove
unwanted signal in anti-Stokes zone.
New palette, diverging.colours
, is used to plot residual
objects.
The const
argument of feemparafac
(and other
functions that call it) now defaults to non-negativity. The previous
default of an unconstrained model can be achieved by setting the
argument to NULL
.
Fixed the bugs in feemjackknife
and feemsplithalf
which prevented them from working with 1-component models.
Fixed the bugs preventing feem
and feemcube
objects from being properly indexed by their dimnames
.
The data(feems)
dataset has been replaced with a
synthetic one, with perfectly matching absorbance spectra.
New function feemflame
to model fluorescence and
scattering signal at the same time.
New function feemindex
to calculate fluorescence indices
and named peak values used in dissolved organic matter analysis.
When the samples and the absorbance information don't match,
feemife
now provides an explanation of what went wrong.
When an error happens during the processing of a feem
inside a collection (a feemcube
or a list of feem
s) or
during feemlist
file import, the offending sample (or file) is
now reported, and the error can be trapped without losing the
traceback.
The feemcube
function can now return the original cube
used to fit feemparafac
, feemjackknife
,
feemsplithalf
objects.
feem
now supports text files from Edinburgh Instruments
F900 software.
feemlist
can now take an importer function as an argument.
feemsplithalf
can limit the number of split-combine
comparisons (the splits
argument) or take a fixed list of
halves to fit and compare (the fixed
argument).
HTML version of the manual now passes Tidy validation without warnings.
Fixed a bug in numerical differentiation code for
feemscatter(method = 'whittaker')
that resulted in d
-th
order derivatives being 2^d
times smaller than they should have
been. Existing optimal lambda
values will have to be updated by
dividing them by 2^{2d}
.
When interpolating using “loess” or “kriging” methods, predicted negative values are replaced with zeroes. Previously, this happened for the whole FEEM, but only with “loess”.
feemcube
now can also extract the original cube from
feemparafac
objects.
Simplified the feemlist(..., simplify.names = TRUE)
logic,
hopefully preventing the test failure on some macOS machines.
Made the feem(..., 'table', fileEncoding = '...')
logic
more robust, especially on older versions of R where it could still
fail when fileEncoding
could not be represented in the locale
encoding.
example(write.openfluor)
no longer leaves files it creates
in the per-session temporary directory.
Added a new ‘whittaker’ interpolation method for
feemscatter
and feemgrid
based on Whittaker smoothing
with configurable difference order(s) and optional iterative penalty
against resulting negative values. This brings in a dependency on the
recommended Matrix package.
The ‘pchip’ interpolation method now offers a choice whether to interpolate the EEM by row, by column, or to do both and average the results (thanks staRdom for the trick). It also doesn't crash anymore when trying to 1-D interpolate less than 3 defined points, falling back to linear interpolation instead.
Added feem
methods for strings and connections, making
it possible to read EEMs from files in simple table formats, including
those produced by Horiba and Panorama software.
Added the feemlist
generic to construct lists of
feem
objects from eemR or EEM objects or vectors
of file and directory paths.
Added the groups
argument to feemsplithalf
,
making it possible to stratify the resulting halves according to a
given factor value.
Added the feemgrid
generic to interpolate FEEM(s) on a
given wavelength grid.
Added coef
methods for feemparafac
,
feemsplithalf
, feemjackknife
returning the coefficients
and summary values (e.g. TCC for split-half) in “long”
data.frame format.
Added function write.openfluor
to export a fitted
feemparafac
object for the OpenFluor service.
Equations in the HTML help pages are now typeset using HTML features instead of plain ASCII.
Dependency on R >= 3.3
is now acknowledged. The
package doesn't actually pass R CMD check
on prior versions
of R. Patches to lower the required R version are welcome if they
don't result in extensive rewrites.
Other documentation improvements.
Fixed a serious bug in feemsplithalf
that would result
in partially overlapping splits being compared. It has been present
since v0.1-1. I apologise to any users of albatross who found
out they have been bitten by this. Steps have been taken to make sure
that other bugs like this don't still lurk in the code and to prevent
similar bugs from occurring in the future.
Potentially breaking change: feemife
now accepts either
named absorbance / optical path length data exactly matching unique
names of samples, or a collection of exactly same length with either
names missing. Named collections of the same length as number of
samples which don't contain all unique sample names are going to be
rejected from now on.
Saved feemsplithalf
and feemjackknife
results now
should consume much less space and much less memory when loaded back.
This is achieved by saving the FEEM cube object in an environment and
only referencing it using vectors of indices instead of keeping copies
of its subsets.
New interpolation method: “kriging”. Very
resource-intensive, but results can be nicer-looking than either
loess
or pchip
.
More consistent progress bars for operations on collections of
feem
s. Scatter correction can now be run in parallel.
Minor documentation improvements.
Fixed a bug in feemjackknife
and feemsplithalf
,
which could manifest when rescale
argument is not 3
and
loading scales are wildly different, resulting in poor approximation
for some of the modes.
Made the documentation for PARAFAC/split-half/jack-knife plots more prominent.
Made tests and examples much faster.
Made the gamma
argument in marine.colours
much
more useful.
Minor improvements in documentation.