Are you still working on this?

[miguelcarcamov]

Hi there,

I am interested on working with this API since it is very useful for my thesis work. Do you know if the functions are compatible with CUDA 10?

Best regards

[pierrepaleo]

Hi @miguelcarcamov

This project has not been updated for a while, but as far as I know things are working.
I just tested the code on a machine with CUDA 10 and all pypwt tests passed.

You will certainly need to modify the -arch compilation flag in PDWT/Makefile or pypwt/setup.py. For example I used -arch=sm_70 for a Tesla V100-SXM2.

The pypwt python wrapper runs numerous tests against the pywavelets module. However the API of pywavelets changed (notably for SWT), so tests fail with recent versions of pywavelets. The version 0.5.0 has to be used.

I tested the following with Python 3.6 in a virtual environment:

git clone https://github.com/pierrepaleo/pypwt
cd pypwt
# Modify the -arch compilation flag in setup.py depending on your device
python setup.py build
python setup.py install
pip install pywavelets==0.5.0 # see above comments 
python test/test_all.py

Tests have to be updated to compare against recent version of pywavelets.

[miguelcarcamov]

what if I want to use it just in C++/CUDA? It should work despite my version of pywavelets right?

[lydiazajiczek]

Hi Pierre, Sorry to hijack your response to the other issue but I have been meaning to email you for ages! First of all thanks for this great library. I unfortunately ran into a very strange behaviour that I was never able to figure out and wondered if you had any insight. I use your library to run the two dimensional wavelet transform on a stack of images, then I find the maximum wavelet coefficient for each pixel along z and do an inverse transform on the composite/maximum wavelet coefficients. The problem is, running the transform/inverse transform is non deterministic and the result looks different every time. This made me think it was a race condition - CUDA has a race condition checker but it didn't find anything. Have you ever seen this behaviour with your library? Thanks a lot! Lydia Neary-Zajiczek Get Outlook for Android<https://aka.ms/ghei36>

…

________________________________ From: Pierre Paleo <notifications@github.com> Sent: Thursday, August 15, 2019 5:46:30 PM To: pierrepaleo/PDWT <PDWT@noreply.github.com> Cc: Subscribed <subscribed@noreply.github.com> Subject: Re: [pierrepaleo/PDWT] Are you still working on this? (#1) Hi @miguelcarcamov<https://github.com/miguelcarcamov> This project has not been updated for a while, but as far as I know things are working. I just tested the code on a machine with CUDA 10 and all pypwt<https://github.com/pierrepaleo/pypwt> tests passed. You will certainly need to modify the -arch compilation flag in PDWT/Makefile or pypwt/setup.py. For example I used -arch=sm_70 for a Tesla V100-SXM2. The pypwt<https://github.com/pierrepaleo/pypwt> python wrapper runs numerous tests against the pywavelets<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpypi.org%2Fproject%2FPyWavelets&data=02%7C01%7C%7C6378701bcaff455f964808d721a01fdf%7C1faf88fea9984c5b93c9210a11d9a5c2%7C0%7C0%7C637014843979558401&sdata=2zK93KKzqmuUnGmig8oYl7ZwspIj%2FIS3TeGctOyXjDo%3D&reserved=0> module. However the API of pywavelets changed (notably for SWT), so tests fail with recent versions of pywavelets. The version 0.5.0 has to be used. I tested the following with Python 3.6 in a virtual environment: git clone https://github.com/pierrepaleo/pypwt cd pypwt # Modify the -arch compilation flag in setup.py depending on your device python setup.py build python setup.py install pip install pywavelets==0.5.0 # see above comments python test/test_all.py Tests have to be updated to compare against recent version of pywavelets. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<#1?email_source=notifications&email_token=AH7GIBS24UCU7VYA56PYRMDQEWB6NA5CNFSM4IL7SBJKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4MLBZA#issuecomment-521711844>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AH7GIBT4PQBQGOT6SLPFYFTQEWB6NANCNFSM4IL7SBJA>.

[pierrepaleo]

@miguelcarcamov Yes pywavelets is just for testing the Python wrapper, so the C++/CUDA API will work without it. What I meant is that if the Python tests pass, then the underlying C++/CUDA code should be valid. I also ran the bare demo of PDWT without issues.

@lydiazajiczek Thanks for the report ; would you be able to provide a minimal code reproducing this behavior ? As far as I remember, the only cases when I got these behaviors is when an illegal memory access is done on the image/coefficients (ex. buggy thresholding function).