pycudwt-multitarget 1.0.4

pycudwt

pycudwt is a python module for parallel Discrete Wavelet Transform. This is a wrapper of PDWT.

Note: this project was formerly named pypwt. It has been renamed pycudwt to have a spot on pypi.

Installation

Requirements

You need cython and nvcc (the Nvidia CUDA compiler, available in the NVIDIA CUDA Toolkit).

For the tests, you need pywavelets. python-pywt is packaged for Debian-like distributions, more recent changes are available on the new repository.

Stable version (from pypi)

pip install pycudwt

From conda recipe

Conda build for a specific cudatoolkit version that matches one in your conda environment, e.g.:

export CUDA_VERSION="10.1.243"
conda build conda-recipe/

Development version (from github)

git clone https://github.com/pierrepaleo/pypwt
cd pypwt
pip install .

You can specify the compute capability when building the library:

PYCUDWT_CC=86 pip install .

# or to target multiple specific GPUs
PYCUDWT_CC=80,90

# or to let nvcc target your current GPU(s)
PYCUDWT_CC=native
PYCUDWT_CC=all

Learn more here.

Testing

If pywt is available, you can check if pycudwt gives consistent results :

cd test
python test_all.py

the results are stored in results.log.

Getting started

Computing a Wavelet Transform wity pycudwt is simple. In ipython:

from pycudwt import Wavelets
from scipy.misc import lena
l = lena()
W = Wavelets(l, "db2", 3)
W
------------- Wavelet transform infos ------------
Wavelet name : db2
Number of levels : 3
Stationary WT : no
Cycle spinning : no
Separable transform : yes
Estimated memory footprint : 5.2 MB
Running on device : GeForce GTX TITAN X
--------------------------------------------------
W.forward()
W.soft_threshold(10)
W.inverse()
imshow(W.image)