Activity

1. horta commented on Apr 17, 2019

   

   horta

   on Apr 17, 2019

   Author

   More actions

   The problem lies in numpy.dual.svd:

   $ docker run -it continuumio/anaconda3:latest /bin/bash
   $ conda create -n test_mkl python=3.7 numpy --yes
   $ conda activate test_mkl
   $ wget https://srv-file1.gofile.io/download/bRLqH7/41e47ac59aa2fea4dd9645dce24021ca/script-save.py
   $ python script-save.py
   $ conda create -n test_nomkl python=3.7 --yes
   $ conda activate test_nomkl
   $ pip install numpy
   $ wget https://srv-file1.gofile.io/download/JlJCVt/41e47ac59aa2fea4dd9645dce24021ca/script-load.py
   $ python script-load.py
   2.3534794732541338
   2.3068937965514184e-15
   2.3534794732541626

   The scripts are:

   # script-save.py
   import urllib.request
   import numpy as np
   import numpy.dual
   
   url = "https://srv-file1.gofile.io/download/0CZT37/41e47ac59aa2fea4dd9645dce24021ca/K.npy"
   urllib.request.urlretrieve(url, "K.npy")
   K = np.load("K.npy")
   S, V, D = np.dual.svd(K)
   np.save("S", S)
   np.save("V", V)
   np.save("D", D)

   # script-load.py
   import numpy as np
   import numpy.dual
   
   S = np.load("S.npy")
   V = np.load("V.npy")
   D = np.load("D.npy")
   
   S2, V2, D2 = np.dual.svd(np.load("K.npy"))
   
   print(np.linalg.norm(S - S2))
   print(np.linalg.norm(V - V2))
   print(np.linalg.norm(D - D2))
2. eric-wieser

   added

   29 - Intel/Anaconda

   component: numpy.random

   on Apr 18, 2019
3. bashtage commented on Apr 20, 2019

   

   bashtage

   on Apr 20, 2019

   Contributor

   More actions

   This is inevitable since it depends on a SVD, and so the result conditional on the linear algebra library.
4. horta commented on Apr 22, 2019

   

   horta

   on Apr 22, 2019

   Author

   More actions

   We could replace SVD decomposition with pivoted Cholesky decomposition. Please, refer to the above issue I just created.
5. bashtage commented on Sep 13, 2019

   

   bashtage

   on Sep 13, 2019

   Contributor

   More actions

   @horta Since does pivoted Cholesky ultimately use BLAS routines? If it does, then this type of non-determinacy might not be easy to resolve.
6. horta commented on Sep 13, 2019

   

   horta

   on Sep 13, 2019

   Author

   More actions

   There are multiple (infinite) SVD decompositions for semi-definite positive matrices, whereas pivoted Cholesky decomposition has a single one, as far as I know.

   (rounding error is not being tackled by the above proposition...)
7. rossbar commented on May 20, 2020

   

   rossbar

   on May 20, 2020

   Contributor

   More actions

   It's been a few months since this was reported - given the deprecation of dual in #16156 and the improvements in #14197, is this still an issue?
8. horta commented on May 21, 2020

   

   horta

   on May 21, 2020

   Author

   More actions

   I'm fine with closing it.
9. seberg

   closed this as completedon May 21, 2020