PyWavelets - Wavelet Transforms in Python
We are very pleased to announce the release of PyWavelets 1.3. This release is functionally the same as 1.2.0, but we have updated the set of binary wheels provided. Specifically we have added aarch64 wheels for Python 3.7 that were missing in 1.2.0 and have updated the versions of manylinux used for the linux wheels in a manner consistent with NumPy and SciPy. We have also dropped musllinux wheels which were present for some architectures in 1.2.0. We may add them again in the future if they are adopted by NumPy and there is a demand for this architecture.
We are very pleased to announce the release of PyWavelets 1.2.
This release has new discrete wavelet transforms features incleading a series of multiresolution analysis functions (details below).
PyWavelets has dropped support for Python 3.5 and 3.6 and now supports Python 3.7-3.10.
We also now provide aarch64 linux wheels as well as universal2 and arm64 wheels that are compatible with Apple's M1 processors.
There is a new series of multilevel stationary wavelet transforms (mra
, mra2
and mran
) suited for multiresolution analysis of 1D, 2D or nD signals, respectively. This MRA analysis is also known as the additive wavelet decomposition because the corresponding inverse functions (imra
, imra2
or imran
) reconstruct the original signal by simple addition of the components. These are a good alternative to the use of the existing SWT functions when it is important to have features aligned across wavelet scales (see the new demo in demo/mra_vs_swt.py
).
There is now an n-dimensional implementation available for the wavelet packet transforms (see class WaveletPacketND
).
pywt.data.camera
has been replaced by a similar, CC0-licensed image because the original image was determined to only be licensed for non-commercial use. Any users who still need the prior camera image for non-commercial use can find it many places online by performing a web search for "cameraman test image".dwt_single
for reflect modes.imp
.A total of 6 people contributed to this release. People with a "+" by their names contributed a patch for the first time. This list of names is automatically generated, and may not be fully complete.
This release is functionally identical to 1.1.0.
This release modified setup.py
to mark the package as Python 3.5+ only so pip
will not try to install 1.1.1 on older Python versions. To prevent pip from trying to install 1.1.0 on older Python, the source tarball for 1.1.0 was removed from PyPI.
.. contents::
We are very pleased to announce the release of PyWavelets 1.1.
This release includes enhanced functionality for both the stationary wavelet
transforms (swt
, swt2
, swtn
) as well as the continuous wavelet
transform (cwt
). In addition, there are a handful of bug fixes as
described in more detail below.
This release has dropped Python 2.7 support and now requires Python >= 3.5.
In addition to these changes to the software itself, a paper describing PyWavelets was recently published in The Journal of Open Source Software: https://joss.theoj.org/papers/10.21105/joss.01237
All swt
functions now have a new trim_approx
option that can be used
to exclude the approximation coefficients from all but the final level of
decomposition. This mode makes the output of these functions consistent with
the format of the output from the corresponding wavedec
functions.
All swt
functions also now have a new norm
option that, when set to
True
and used in combination with trim_approx=True
, gives a partition
of variance across the transform coefficients. In other words, the sum of
the variances of all coefficients is equal to the variance of the original
data. This partitioning of variance makes the swt
transform more similar
to the multiple-overlap DWT (MODWT) described in Percival and Walden's book,
"Wavelet Methods for Time Series Analysis". (#476)
A demo of this new swt
functionality is available at
https://github.com/PyWavelets/pywt/blob/master/demo/swt_variance.py
The continuous wavelet transform (cwt
) now offers an FFT-based
implementation in addition to the previous convolution based one. The new
method
argument can be set to either 'conv'
or 'fft'
to select
between these two implementations. (#490).
The cwt
now also has axis
support so that CWTs can be applied in
batch along any axis of an n-dimensional array. This enables faster batch
transformation of signals. (#509)
When the input to cwt
is single precision, the computations are now
performed in single precision. This was done both for efficiency and to make
cwt
handle dtypes consistently with the discrete transforms in
PyWavelets. This is a change from the prior behaviour of always performing
the cwt
in double precision. (#507)
When using complex-valued wavelets with the cwt
, the output will now be
the complex conjugate of the result that was produced by PyWavelets 1.0.x.
This was done to account for a bug described below. The magnitude of the
cwt
coefficients will still match those from previous releases. (#439)
For a cwt
with complex wavelets, the results in PyWavelets 1.0.x releases
matched the output of Matlab R2012a's cwt
. Howveer, older Matlab releases
like R2012a had a phase that was of opposite sign to that given in textbook
definitions of the CWT (Eq. 2 of Torrence and Compo's review article, "A
Practical Guide to Wavelet Analysis"). Consequently, the wavelet coefficients
were the complex conjugates of the expected result. This was validated by
comparing the results of a transform using cmor1.0-1.0
as compared to the
cwt
implementation available in Matlab R2017b as well as the function
wt.m
from the Lancaster University Physics department's
MODA toolbox <https://github.com/luphysics/MODA>
_. (#439)
For some boundary modes and data sizes, round-trip dwt
/idwt
can
result in an output that has one additional coefficient. Prior to this
relese, this could cause a failure during WaveletPacket
or
WaveletPacket2D
reconstruction. These wavelet packet transforms have now
been fixed and round-trip wavelet packet transforms always preserve the
original data shape. (#448)
All inverse transforms now handle mixed precision coefficients consistently. Prior to this release some inverse transform raised an error upon encountering mixed precision dtypes in the wavelet subbands. In release 1.1, when the user-provided coefficients are a mixture of single and double precision, all coefficients will be promoted to double precision. (#450)
A bug that caused a failure for iswtn
when using user-provided axes
with non-uniform shape along the transformed axes has been fixed. (#462)
The PyWavelet test suite now uses pytest
rather than nose
. (#477)
Cython code has been updated to use language_level=3
. (#435)
PyWavelets has adopted the SciPy Code of Conduct. (#521)
PyWavelets 1.0.3 is functionally equivalent to the 1.0.2 release. It was made to archive the JOSS paper about PyWavelets to the 1.0.x branch and serve as a reference corresponding to the version that was peer reviewed.