Print current residual from callback in scipy.sparse.linalg.cg

Question

I am using scipy.sparse.linalg.cg to solve a large, sparse linear system, and it works fine, except that I would like to add a progress report, so that I can monitor the residual as the solver works. I've managed to set up a callback, but I can't figure out how to access the current residual from inside the callback. Calculating the residual myself is possible, of course, but that is a rather heavy operation, which I'd like to avoid. Have I missed something, or is there no efficient way of getting at the residual?

Answer 1

The callback is only sent xk, the current solution vector. So you don't have direct access to the residual. However, the source code shows resid is a local variable in the cg function.

So, with CPython, it is possible to use the inspect module to peek at the local variables in the caller's frame:

import inspect
import numpy as np
import scipy as sp
import scipy.sparse as sparse
import scipy.sparse.linalg as splinalg
import random

def report(xk):
    frame = inspect.currentframe().f_back
    print(frame.f_locals['resid'])

N = 200
A = sparse.lil_matrix( (N, N) )
for _ in xrange(N):
    A[random.randint(0, N-1), random.randint(0, N-1)] = random.randint(1, 100)

b = np.random.randint(0, N-1, size = N)
x, info = splinalg.cg(A, b, callback = report)