[Numpy-tickets] [NumPy] #649: numpy.linalg.lstsq segfaults on i686 32-bit and ia64 linux

[NumPy]

#649: numpy.linalg.lstsq segfaults on i686 32-bit and ia64 linux
------------------------+---------------------------------------------------
 Reporter:  psederberg  |       Owner:  somebody
     Type:  defect      |      Status:  new     
 Priority:  normal      |   Milestone:  1.0.5   
Component:  Other       |     Version:  none    
 Severity:  normal      |    Keywords:          
------------------------+---------------------------------------------------
 Howdy Everybody:

 I am coding a simple ridge regression with numpy and seem to have
 uncovered a bug that exists in numpy.linalg.lstsq, but NOT in
 scipy.linalg,lstsq.

 Running the following code gives rise to a segfault on 32-bit i686 and
 64-bit ia64 linux, but not on a amd64 linux machine:

 ############ START CODE ##################
 import numpy as N

 # set dimensions
 nsamples = 53
 nfeatures = 1000

 # create random data
 data = N.random.rand(nsamples,nfeatures)
 lab = N.random.rand(nsamples)

 # run a ridge regression with constant term
 Lambda = .05*nfeatures*N.eye(nfeatures)
 a = N.concatenate( \
     (N.concatenate((data, N.ones((nsamples, 1))), 1),
      N.concatenate((Lambda, N.zeros((nfeatures, 1))), 1)))
 b = N.concatenate((lab, N.zeros(nfeatures)))
 w = N.linalg.lstsq(a,b)[0]  # here is the segfault

 ################ END CODE ####################

 All machines were tested with Debian testing and numpy version 1.0.4.

 I've attached the results of running valgrind on the above code.

 It should also be added that replacing the numpy lstsq with the
 scipy.linalg.lstsq prevents the segfault and gives rise to the correct
 result.

 Thanks,
 Per

-- 
Ticket URL: <http://scipy.org/scipy/numpy/ticket/649>
NumPy <http://projects.scipy.org/scipy/numpy>
The fundamental package needed for scientific computing with Python.