The hardware and bandwidth for this mirror is donated by METANET, the Webhosting and Full Service-Cloud Provider.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]metanet.ch.
Due to the bug fixes below, the results of fcwt()
and fcwt_batch() change numerically, in some cases
considerably (in particular when averaging over time windows, for odd
signal lengths, and for freq_scale = "linear").
fcwt_batch(): the defaults of sigma and
n_freqs are now the same as for fcwt()
(previously sigma = 1, and n_freqs was a third
lower).
[ (scalogram subsetting): indices that do not select
a contiguous range now raise an error, since the result cannot be
represented by evenly spaced time and frequency axes. Use
as.matrix() for arbitrary subsetting.
fcwt(): n_threads must be at least
1.
R >= 4.2 is required.
fcwt() (and hence
fcwt_batch())Values of a frequency band no longer leak into the following (lower-frequency) bands.
For odd signal lengths, every other frequency band was shifted by one sample in time.
Averaging over time windows (y_sample_freq lower
than x_sample_freq) under-estimated the mean.
With freq_scale = "linear", the frequencies actually
computed did not match the reported ones (only the lowest one did): they
now range from freq_begin to freq_end (both
inclusive), as for freq_scale = "log".
n_freqs = 1 with freq_scale = "log"
returned garbage (and could crash on CPUs with AVX support).
The sampling frequency x_sample_freq was truncated
to an integer (in Hz), so non-integer sampling frequencies lead to wrong
frequencies, and sampling frequencies below 1 Hz did not work at
all.
The time axis of the result was wrong if
x_sample_freq / y_sample_freq is not an integer: the result
now records the actual output sampling rate (which can be slightly
higher than y_sample_freq).
Fix undefined behavior (memory alignment) that could lead to crashes on some platforms.
Very large inputs (more than 2^31 output values) lead to an integer overflow, failing memory allocations to undefined behavior. The former are now supported, the latter raise an error.
fcwt_batch()n_freqs = 1 did not work.[: selecting frequencies (columns) now adjusts the
frequency range (previously, the frequency labels were wrong), and
negative and logical indices now adjust the time offset
correctly.
plot() / autoplot(): the axis labels
show the units again with ggplot2 >= 4.0.
fcwt()Considerably less memory is needed: the complex transform is no longer stored as a whole, but processed frequency by frequency (e.g. about 4x less peak memory when averaging over time windows).
With rm_coi = TRUE (the default), the boundary
values are now removed with much less memory overhead (peak memory about
2.5x instead of 6x the size of the result).
The global OpenMP thread setting of the R session is no longer changed.
fcwt_batch()max_batch_size is now chosen such that
processing a batch needs about 500 MB of memory, taking into account the
reduced memory usage of fcwt() and the averaging over time
windows (y_sample_freq). Previously, the batches were often
unnecessarily small, which lead to large losses from boundary
effects.time column of as.data.frame() /
as_tibble(), the time offset and time resolutions in
print(), and the time axis of plot() /
autoplot() (whose time_unit argument now
defaults to NULL, i.e. automatic).The installation now explicitly checks for the single-precision
fftw library (falling back to a direct link test if pkg-config does not
find it), and gives clear instructions on which packages to install if
it is missing. SystemRequirements describes the
dependencies in more detail (fixes CRAN issues).
The installation no longer fails if the compiler supports OpenMP,
but the fftw library comes without OpenMP support
(libfftw3f_omp): the Fourier transforms then run
single-threaded.
?fCWTr).fcwt():
fix a bug where fcwt(..., n_freqs = 1, ...) leads to
a faulty data representation.
add freq_scale argument to choose between a linear
or logarithmic frequency scale.
allow (and prefer) sample_freq,
freq_begin, freq_end arguments to be passed as
a dimensionful quantities, via u().
minor argument name changes.
add new argument y_sample_freq which allows to
specify the sampling rate of the output signal.
add meaningful default values for most of the arguments.
fcwt_batch():
allow (and prefer) sample_freq,
freq_begin, freq_end and
time_resolution arguments to be passed as a dimensionful
quantities, via u().
in some cases, the averaging procedure together with the batching could lead to accidentally dropped or duplicated time slices: the mechanism has been overhauled, every slice should be perfectly aligned now, even when averaged.
shared argument with fcwt(): same changes as in
fcwt().
new unit helper functions: u() and
du(): include improved physical unit treatment with the units package.
Frequency and time parameters can now be “units” objects, created with
u(). Allow user to adjust plot() scales to use
arbitrary time/frequency units.
new helper functions to determine the correct value of sigma for
a given use case: sigma_res(),
sigma_freq_res_rel(), sigma_from_time_res(),
sigma_from_freq_res(),
sigma_from_freq_res_rel().
add more convenience S3 methods: print(),
[], as.matrix(), rbind(),
tibble::as_tibble().
openmp_enabled()/avx_enabled() new
functions to check for system OpenMP/AVX support.
new vignette about controlling the time-frequency resolution
vignette("sigma").
add a package logo
These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
