`funcalls` output from scipy.optimize.fmin_l_bfgs_b unreliable?

Question

My function:

count = 0
def fake(x):
    global count
    print count
    count += 1
    return x ** 4 + 10 * x ** 3 + 4 * x ** 2 + 7 * x + 1

‘Nelder-Mead’ method, gives me correct number of function calls.

scipy.optimize.fmin(fake, [1])

0
1
...

45
Optimization terminated successfully.
         Current function value: -887.470826
         Iterations: 23
         Function evaluations: 46
Out[377]:
array([-7.25761719])

BFGS method, gives me correct number of function calls.

scipy.optimize.fmin_bfgs(fake, [1])
0
1
...
61
62
Optimization terminated successfully.
         Current function value: -887.470826
         Iterations: 6
         Function evaluations: 63
         Gradient evaluations: 21
Out[380]:
array([-7.25765231])

However, L-BFGS-B, gives me strange number of function calls. What happened?

scipy.optimize.fmin_l_bfgs_b(fake, [1], approx_grad=True)
0
1
...
43
44
Out[374]:
(array([-7.25765246]),
 array([-887.47082639]),
 {'funcalls': 15,
  'grad': array([ -3.41060513e-05]),
  'nit': 6,
  'task': 'CONVERGENCE: REL_REDUCTION_OF_F_<=_FACTR*EPSMCH',
  'warnflag': 0})

Answer 1

From the code, it appears that the number of function evaluations is counted when the function and gradient is evaluated, but does not count the number of function calls needed to approximate the gradient.

Inside my version of lbfgsb.py:

197             n_function_evals += 1
198             # Overwrite f and g:
199             f, g = func_and_grad(x)

I recommend that you report this as a bug.

Edit: As per the comment below, this was actually a bug and was fixed.