PyWavelets 0.5.0 Release Notes¶
Contents
PyWavelets is a Python toolbox implementing both discrete and continuous wavelet transforms (mathematical time-frequency transforms) with a wide range of built-in wavelets. C/Cython are used for the low-level routines, enabling high performance. Key Features of PyWavelets are:
1D, 2D and nD Forward and Inverse Discrete Wavelet Transform (DWT and IDWT)
1D, 2D and nD Multilevel DWT and IDWT
1D and 2D Forward and Inverse Stationary Wavelet Transform
1D and 2D Wavelet Packet decomposition and reconstruction
1D Continuous Wavelet Transform
When multiple valid implementations are available, we have chosen to maintain consistency with MATLAB™’s Wavelet Toolbox.
PyWavelets 0.5.0 is the culmination of 1 year of work. In addition to several new features, substantial refactoring of the underlying C and Cython code have been made.
This release requires Python 2.6, 2.7 or 3.3-3.5 and NumPy 1.9.1 or greater. This will be the final release supporting Python 2.6 and 3.3.
Highlights of this release include:
1D continuous wavelet transforms
new discrete wavelets added (additional Debauchies and Coiflet wavelets)
new ‘reflect’ extension mode for discrete wavelet transforms
faster performance for multilevel forward stationary wavelet transforms (SWT)
n-dimensional support added to forward SWT
routines to convert multilevel DWT coefficients to and from a single array
axis support for multilevel DWT
substantial refactoring/reorganization of the underlying C and Cython code
New features¶
1D Continous Wavelet Transforms¶
A wide range of continous wavelets are now available. These include the following:
Gaussian wavelets (
gaus1
…``gaus8``)Mexican hat wavelet (
mexh
)Morlet wavelet (
morl
)Complex Gaussian wavelets (
cgau1
…``cgau8``)Shannon wavelet (
shan
)Frequency B-Spline wavelet (
fbsp
)Complex Morlet wavelet (
cmor
)
Also, see the new CWT-related demo: demo/cwt_analysis.py
New discrete wavelets¶
Additional Debauchies wavelets (db20
…``db38``) and Coiflets
(coif6
…``coif17``) have been added.
New extension mode: reflect¶
Discrete wavelet transforms support a new extension mode, reflect
. This
mode pads an array symmetrically, but without repeating the edge value. As an
example:
pad array pad
4 3 2 | 1 2 3 4 5 | 4 3 2
This differs from symmetric
, which repeats the values at the boundaries:
pad array pad
3 2 1 | 1 2 3 4 5 | 5 4 3
Multilevel DWT Coefficient Handling¶
New routines to convert the coefficients returned by multilevel DWT routines
to and from a single n-dimensional array have been added.
pywt.coeffs_to_array
concatenates the output of wavedec
, wavedec2
or wavedecn
into a single numpy array. pywt.array_to_coeffs
can be
used to transform back from a single coefficient array to a format appropriate
for waverec
, waverec2
or waverecn
.
More C function calls release the GIL¶
Cython code calling the wavelet filtering routines (DWT and SWT) now releases the global interpreter lock (GIL) where possible. A potential use case is in speeding up the batch computation of several large DWTs using multi-threading (e.g. via concurrent.futures).
Multilevel wavelet transforms along specific axes¶
The axis specific transform support introduced in the prior release was
extended to the multilevel DWT transforms. All wavedec*
and waverec*
routines have a new axis (1D) or axes (2D, nD) keyword argument. If
unspecified the default behaviour is to transform all axes of the input.
Faster multilevel stationary wavelet transforms¶
Stationary wavelet transforms are now faster when the number of levels is greater than one. The improvement can be very large (multiple orders of magnitude) for transforms with a large number of levels.
Deprecated features¶
Backwards incompatible changes¶
A FutureWarning was added to swt2
and iswt2
to warn about a pending
backwards incompatible change to the order of the coefficients in the list
returned by these routines. The actual change will not occur until the next
release. Transform coefficients will be returned in descending rather than
ascending order. This change is being made for consistency with all other
existing multi-level transforms in PyWavelets.
Bugs Fixed¶
demo/image_blender.py
was updated to support the new api of Pillow 3.x
A bug related to size of assumed size_t
on some platforms/compilers
(e.g. Windows with mingw64) was fixed.
Fix to memory leak in (i)dwt_axis
Fix to a performance regression in idwt
and iswt
that was introduced
in v0.4.0.
Fixed a bug in dwtn
and idwtn
for data with complex dtype when
axes != None
.
Other changes¶
The minimum supported numpy version has been increased to 1.9.1.
Test coverage (including for the Cython and C code) via Codecov was added and the overall test coverage has been improved.
A substantial overhaul of the C extension code has been performed. Custom templating is no longer used. The intention is to make this code easier to maintain and expand in the future.
The Cython code has been split out into a multiple files to hopefully make relevant portions of the wrappers easier to find for future developers.
setup.py
now relies on setuptools in all cases (rather than distutils).