Erroneous Hessian on second evaluation

[dpo]

I'm having an issue where the first (sparse) Hessian evaluation is correct, but the second is not. I'm not sure if it's an Adol-C issue or a pyadolc issue. I'm using ColPack 1.0.9 and Adol-C 2.6.0 on OSX. Here's a minimum working example:

import adolc
import numpy as np

def rosenbrock(x):
    return np.sum(100 * (x[1:] - x[:-1]**2)**2 + (1 - x[:-1])**2)

nvar = 5
x = -np.ones(nvar)

# trace function value
obj_trace_id = 1
adolc.trace_on(obj_trace_id)
ax = adolc.adouble(x)
adolc.independent(ax)
y = rosenbrock(ax)
adolc.dependent(y)
adolc.trace_off()

# evaluate gradient
g = adolc.gradient(obj_trace_id, x)

# evaluate sparse hessian
options = np.zeros(2, dtype=int)
nnz, rind0, cind0, values0 =  \
    adolc.colpack.sparse_hess_no_repeat(obj_trace_id,
                                        x, options=options)
print "rind: ", rind0
print "cind: ", cind0
print "vals: ", values0

# evaluate sparse hessian again
nnz, rind, cind, values =  \
    adolc.colpack.sparse_hess_repeat(obj_trace_id,
                                     x,
                                     rind0,
                                     cind0,
                                     values0)

print "rind: ", rind
print "cind: ", cind
print "vals: ", values

Here is the output I receive:

rind:  [0 0 1 1 2 2 3 3 4]
cind:  [0 1 1 2 2 3 3 4 4]
vals:  [ 1602.   400.  1802.   400.  1802.   400.  1802.   400.   200.]
rind:  [0 0 1 1 2 2 3 3 4]
cind:  [0 1 1 2 2 3 3 4 4]
vals:  [ 1602.   400.  1802.   800.  1802.   800.  1802.   400.   200.]

On the second sparse Hessian evaluation, certain off-diagonals changed. Can you reproduce this issue?

[dpo]

I get the same results with ColPack 1.0.10 and Adol-C 2.6.1. I noticed that some pyadolc tests are failing; not sure if it's related: https://gist.github.com/e01e5d6348c84f816b71fdea7161153f

[b45ch1]

Hi Dominique,

I can reproduce the error with ADOL-C-2.6.0 and ColPack-1.0.9.

adolc/colpack/wrapped_functions.py:119: FutureWarning: comparison to `None` will result in an elementwise object comparison in the future.
  if options == None:
rind:  [0 0 1 1 2 2 3 3 4]
cind:  [0 1 1 2 2 3 3 4 4]
vals:  [ 1602.   400.  1802.   400.  1802.   400.  1802.   400.   200.]
rind:  [0 0 1 1 2 2 3 3 4]
cind:  [0 1 1 2 2 3 3 4 4]
vals:  [ 1602.   400.  1802.   800.  1802.   800.  1802.   400.   200.]

The error does not appear if you use options[:] = 0,1.

/home/swalter/projects/pyadolc/build/lib.linux-x86_64-2.7/adolc/colpack/wrapped_functions.py:119: FutureWarning: comparison to `None` will result in an elementwise object comparison in the future.
  if options == None:
rind:  [0 0 1 1 2 2 3 3 4]
cind:  [0 1 1 2 2 3 3 4 4]
vals:  [ 1602.   400.  1802.   400.  1802.   400.  1802.   400.   200.]
rind:  [0 0 1 1 2 2 3 3 4]
cind:  [0 1 1 2 2 3 3 4 4]
vals:  [ 1602.   400.  1802.   400.  1802.   400.  1802.   400.   200.]

#15 seems to be related. Back then, the error resided in ADOL-C/Colpack, not in pyadolc. Unfortunately, I don't have time to look into it right now. You may want to involve the ADOL-C maintainers.

Regarding the errors in the tests:
I cannot reproduce the errors in the tests. Are the errors in adolc.test() reproducible on your machine?

[dpo]

Many thanks Sebastian! The options you give work for me. I'll be in touch with the Adol-C folks.