grlee77
commented
3 years ago Hi @codesteller. It sounds like your use case is primarily for 1D signals?
I am a maintainer at PyWavelets and had worked on a new package for discrete wavelet transforms (and more general, overcomplete framelet transforms such as shearlets, dual-tree wavelets, etc) that have both NumPy and CuPy backends. This was done as a research project and is currently still in a private repository. A few years back, I had unsuccessfully applied for an academic grant to support its development and transition into a more mature open source package. I have since moved to a software development role at Quansight and no longer have an immediate research need for these transforms, but would like to see the past work I did be of benefit to others.
Although there are still some things in the API that were a bit experimental, I am interested in making it available. It is mainly a matter of finding time to do so.
One possibility would be extracting a minimal subset of features for inclusion here (maybe just traditional covering PyWavelets dwt, idwt, wavedec and waverec). This repository already has an upfirdn function that can be used for the core convolution operations in these transforms. If we want to reproduce all of the boundary modes from PyWavelets, though, those would need to adapt that code from another upfirdn implementation I previously published at https://github.com/mritools/fast_upfirdn.
A second possibility is publishing in a new PyFramelets repository under the PyWavelets organization. The main drawback of that approach is that we don't have CI infrastructure there for testing on GPUs and I suspect we will find it challenging to find resources to adequately maintain the package.
I am interested in hearing from cuSignal devs on the relative level of interest in wavelet transforms for cuSignal. In the past, I was most interested in use cases applying to 3D (or 3D + channels) medical imaging datasets, so the implementations I have are n-dimensional and have the general ability to select which axes the transform would be applied over.
cc @jakirkham , @quasiben , @rgommers for awareness
awthomp
codesteller86
github-actions[bot]
Is your feature request related to a problem? Please describe. I wish I could use cuSignal to work on Discrete Wavelet Transforms in a way like I use PyWavelets
(cA, cD) = pywt.dwt(arr, 'db1',axis=-1)Describe the solution you'd like I would like to have a single api to take an array of data and the mother wavelets, which is GPU accelerated and produces output as in PyWavelets
Describe alternatives you've considered cuSignal has CWT, but this does not work for our usecase. Also tried https://github.com/pierrepaleo/PDWT, but it does not work with variations in wavelets as input.
Additional context A similar to PyWavelets
(cA, cD) = pywt.dwt(arr, 'db1',axis=-1)