""" Unit tests for optimization routines from optimize.py Authors: Ed Schofield, Nov 2005 Andrew Straw, April 2008 To run it in its simplest form:: nosetests test_optimize.py """ import itertools import numpy as np from numpy.testing import (assert_allclose, assert_equal, assert_, assert_almost_equal, assert_warns, assert_array_less, suppress_warnings) import pytest from pytest import raises as assert_raises from scipy import optimize from scipy.optimize._minimize import MINIMIZE_METHODS from scipy.optimize._differentiable_functions import ScalarFunction from scipy.optimize.optimize import MemoizeJac def test_check_grad(): # Verify if check_grad is able to estimate the derivative of the # logistic function. def logit(x): return 1 / (1 + np.exp(-x)) def der_logit(x): return np.exp(-x) / (1 + np.exp(-x))**2 x0 = np.array([1.5]) r = optimize.check_grad(logit, der_logit, x0) assert_almost_equal(r, 0) r = optimize.check_grad(logit, der_logit, x0, epsilon=1e-6) assert_almost_equal(r, 0) # Check if the epsilon parameter is being considered. r = abs(optimize.check_grad(logit, der_logit, x0, epsilon=1e-1) - 0) assert_(r > 1e-7) class CheckOptimize(object): """ Base test case for a simple constrained entropy maximization problem (the machine translation example of Berger et al in Computational Linguistics, vol 22, num 1, pp 39--72, 1996.) """ def setup_method(self): self.F = np.array([[1, 1, 1], [1, 1, 0], [1, 0, 1], [1, 0, 0], [1, 0, 0]]) self.K = np.array([1., 0.3, 0.5]) self.startparams = np.zeros(3, np.float64) self.solution = np.array([0., -0.524869316, 0.487525860]) self.maxiter = 1000 self.funccalls = 0 self.gradcalls = 0 self.trace = [] def func(self, x): self.funccalls += 1 if self.funccalls > 6000: raise RuntimeError("too many iterations in optimization routine") log_pdot = np.dot(self.F, x) logZ = np.log(sum(np.exp(log_pdot))) f = logZ - np.dot(self.K, x) self.trace.append(np.copy(x)) return f def grad(self, x): self.gradcalls += 1 log_pdot = np.dot(self.F, x) logZ = np.log(sum(np.exp(log_pdot))) p = np.exp(log_pdot - logZ) return np.dot(self.F.transpose(), p) - self.K def hess(self, x): log_pdot = np.dot(self.F, x) logZ = np.log(sum(np.exp(log_pdot))) p = np.exp(log_pdot - logZ) return np.dot(self.F.T, np.dot(np.diag(p), self.F - np.dot(self.F.T, p))) def hessp(self, x, p): return np.dot(self.hess(x), p) class CheckOptimizeParameterized(CheckOptimize): def test_cg(self): # conjugate gradient optimization routine if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': self.disp, 'return_all': False} res = optimize.minimize(self.func, self.startparams, args=(), method='CG', jac=self.grad, options=opts) params, fopt, func_calls, grad_calls, warnflag = \ res['x'], res['fun'], res['nfev'], res['njev'], res['status'] else: retval = optimize.fmin_cg(self.func, self.startparams, self.grad, (), maxiter=self.maxiter, full_output=True, disp=self.disp, retall=False) (params, fopt, func_calls, grad_calls, warnflag) = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.7.0. Don't allow them to increase. assert_(self.funccalls == 9, self.funccalls) assert_(self.gradcalls == 7, self.gradcalls) # Ensure that the function behaves the same; this is from SciPy 0.7.0 assert_allclose(self.trace[2:4], [[0, -0.5, 0.5], [0, -5.05700028e-01, 4.95985862e-01]], atol=1e-14, rtol=1e-7) def test_cg_cornercase(self): def f(r): return 2.5 * (1 - np.exp(-1.5*(r - 0.5)))**2 # Check several initial guesses. (Too far away from the # minimum, the function ends up in the flat region of exp.) for x0 in np.linspace(-0.75, 3, 71): sol = optimize.minimize(f, [x0], method='CG') assert_(sol.success) assert_allclose(sol.x, [0.5], rtol=1e-5) def test_bfgs(self): # Broyden-Fletcher-Goldfarb-Shanno optimization routine if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': self.disp, 'return_all': False} res = optimize.minimize(self.func, self.startparams, jac=self.grad, method='BFGS', args=(), options=opts) params, fopt, gopt, Hopt, func_calls, grad_calls, warnflag = ( res['x'], res['fun'], res['jac'], res['hess_inv'], res['nfev'], res['njev'], res['status']) else: retval = optimize.fmin_bfgs(self.func, self.startparams, self.grad, args=(), maxiter=self.maxiter, full_output=True, disp=self.disp, retall=False) (params, fopt, gopt, Hopt, func_calls, grad_calls, warnflag) = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.7.0. Don't allow them to increase. assert_(self.funccalls == 10, self.funccalls) assert_(self.gradcalls == 8, self.gradcalls) # Ensure that the function behaves the same; this is from SciPy 0.7.0 assert_allclose(self.trace[6:8], [[0, -5.25060743e-01, 4.87748473e-01], [0, -5.24885582e-01, 4.87530347e-01]], atol=1e-14, rtol=1e-7) def test_bfgs_infinite(self): # Test corner case where -Inf is the minimum. See gh-2019. func = lambda x: -np.e**-x fprime = lambda x: -func(x) x0 = [0] with np.errstate(over='ignore'): if self.use_wrapper: opts = {'disp': self.disp} x = optimize.minimize(func, x0, jac=fprime, method='BFGS', args=(), options=opts)['x'] else: x = optimize.fmin_bfgs(func, x0, fprime, disp=self.disp) assert_(not np.isfinite(func(x))) def test_powell(self): # Powell (direction set) optimization routine if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': self.disp, 'return_all': False} res = optimize.minimize(self.func, self.startparams, args=(), method='Powell', options=opts) params, fopt, direc, numiter, func_calls, warnflag = ( res['x'], res['fun'], res['direc'], res['nit'], res['nfev'], res['status']) else: retval = optimize.fmin_powell(self.func, self.startparams, args=(), maxiter=self.maxiter, full_output=True, disp=self.disp, retall=False) (params, fopt, direc, numiter, func_calls, warnflag) = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.7.0. Don't allow them to increase. # # However, some leeway must be added: the exact evaluation # count is sensitive to numerical error, and floating-point # computations are not bit-for-bit reproducible across # machines, and when using e.g., MKL, data alignment # etc., affect the rounding error. # assert_(self.funccalls <= 116 + 20, self.funccalls) assert_(self.gradcalls == 0, self.gradcalls) # Ensure that the function behaves the same; this is from SciPy 0.7.0 assert_allclose(self.trace[34:39], [[0.72949016, -0.44156936, 0.47100962], [0.72949016, -0.44156936, 0.48052496], [1.45898031, -0.88313872, 0.95153458], [0.72949016, -0.44156936, 0.47576729], [1.72949016, -0.44156936, 0.47576729]], atol=1e-14, rtol=1e-7) def test_powell_bounded(self): # Powell (direction set) optimization routine # same as test_powell above, but with bounds bounds = [(-np.pi, np.pi) for _ in self.startparams] if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': self.disp, 'return_all': False} res = optimize.minimize(self.func, self.startparams, args=(), bounds=bounds, method='Powell', options=opts) params, fopt, direc, numiter, func_calls, warnflag = ( res['x'], res['fun'], res['direc'], res['nit'], res['nfev'], res['status']) assert func_calls == self.funccalls assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'. # Generally, this takes 131 function calls. However, on some CI # checks it finds 138 funccalls. This 20 call leeway was also # included in the test_powell function. # The exact evaluation count is sensitive to numerical error, and # floating-point computations are not bit-for-bit reproducible # across machines, and when using e.g. MKL, data alignment etc. # affect the rounding error. assert self.funccalls <= 131 + 20 assert self.gradcalls == 0 def test_neldermead(self): # Nelder-Mead simplex algorithm if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': self.disp, 'return_all': False} res = optimize.minimize(self.func, self.startparams, args=(), method='Nelder-mead', options=opts) params, fopt, numiter, func_calls, warnflag = ( res['x'], res['fun'], res['nit'], res['nfev'], res['status']) else: retval = optimize.fmin(self.func, self.startparams, args=(), maxiter=self.maxiter, full_output=True, disp=self.disp, retall=False) (params, fopt, numiter, func_calls, warnflag) = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.7.0. Don't allow them to increase. assert_(self.funccalls == 167, self.funccalls) assert_(self.gradcalls == 0, self.gradcalls) # Ensure that the function behaves the same; this is from SciPy 0.7.0 assert_allclose(self.trace[76:78], [[0.1928968, -0.62780447, 0.35166118], [0.19572515, -0.63648426, 0.35838135]], atol=1e-14, rtol=1e-7) def test_neldermead_initial_simplex(self): # Nelder-Mead simplex algorithm simplex = np.zeros((4, 3)) simplex[...] = self.startparams for j in range(3): simplex[j+1, j] += 0.1 if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': False, 'return_all': True, 'initial_simplex': simplex} res = optimize.minimize(self.func, self.startparams, args=(), method='Nelder-mead', options=opts) params, fopt, numiter, func_calls, warnflag = (res['x'], res['fun'], res['nit'], res['nfev'], res['status']) assert_allclose(res['allvecs'][0], simplex[0]) else: retval = optimize.fmin(self.func, self.startparams, args=(), maxiter=self.maxiter, full_output=True, disp=False, retall=False, initial_simplex=simplex) (params, fopt, numiter, func_calls, warnflag) = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.17.0. Don't allow them to increase. assert_(self.funccalls == 100, self.funccalls) assert_(self.gradcalls == 0, self.gradcalls) # Ensure that the function behaves the same; this is from SciPy 0.15.0 assert_allclose(self.trace[50:52], [[0.14687474, -0.5103282, 0.48252111], [0.14474003, -0.5282084, 0.48743951]], atol=1e-14, rtol=1e-7) def test_neldermead_initial_simplex_bad(self): # Check it fails with a bad simplices bad_simplices = [] simplex = np.zeros((3, 2)) simplex[...] = self.startparams[:2] for j in range(2): simplex[j+1, j] += 0.1 bad_simplices.append(simplex) simplex = np.zeros((3, 3)) bad_simplices.append(simplex) for simplex in bad_simplices: if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': False, 'return_all': False, 'initial_simplex': simplex} assert_raises(ValueError, optimize.minimize, self.func, self.startparams, args=(), method='Nelder-mead', options=opts) else: assert_raises(ValueError, optimize.fmin, self.func, self.startparams, args=(), maxiter=self.maxiter, full_output=True, disp=False, retall=False, initial_simplex=simplex) def test_ncg_negative_maxiter(self): # Regression test for gh-8241 opts = {'maxiter': -1} result = optimize.minimize(self.func, self.startparams, method='Newton-CG', jac=self.grad, args=(), options=opts) assert_(result.status == 1) def test_ncg(self): # line-search Newton conjugate gradient optimization routine if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': self.disp, 'return_all': False} retval = optimize.minimize(self.func, self.startparams, method='Newton-CG', jac=self.grad, args=(), options=opts)['x'] else: retval = optimize.fmin_ncg(self.func, self.startparams, self.grad, args=(), maxiter=self.maxiter, full_output=False, disp=self.disp, retall=False) params = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.7.0. Don't allow them to increase. assert_(self.funccalls == 7, self.funccalls) assert_(self.gradcalls <= 22, self.gradcalls) # 0.13.0 # assert_(self.gradcalls <= 18, self.gradcalls) # 0.9.0 # assert_(self.gradcalls == 18, self.gradcalls) # 0.8.0 # assert_(self.gradcalls == 22, self.gradcalls) # 0.7.0 # Ensure that the function behaves the same; this is from SciPy 0.7.0 assert_allclose(self.trace[3:5], [[-4.35700753e-07, -5.24869435e-01, 4.87527480e-01], [-4.35700753e-07, -5.24869401e-01, 4.87527774e-01]], atol=1e-6, rtol=1e-7) def test_ncg_hess(self): # Newton conjugate gradient with Hessian if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': self.disp, 'return_all': False} retval = optimize.minimize(self.func, self.startparams, method='Newton-CG', jac=self.grad, hess=self.hess, args=(), options=opts)['x'] else: retval = optimize.fmin_ncg(self.func, self.startparams, self.grad, fhess=self.hess, args=(), maxiter=self.maxiter, full_output=False, disp=self.disp, retall=False) params = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.7.0. Don't allow them to increase. assert_(self.funccalls <= 7, self.funccalls) # gh10673 assert_(self.gradcalls <= 18, self.gradcalls) # 0.9.0 # assert_(self.gradcalls == 18, self.gradcalls) # 0.8.0 # assert_(self.gradcalls == 22, self.gradcalls) # 0.7.0 # Ensure that the function behaves the same; this is from SciPy 0.7.0 assert_allclose(self.trace[3:5], [[-4.35700753e-07, -5.24869435e-01, 4.87527480e-01], [-4.35700753e-07, -5.24869401e-01, 4.87527774e-01]], atol=1e-6, rtol=1e-7) def test_ncg_hessp(self): # Newton conjugate gradient with Hessian times a vector p. if self.use_wrapper: opts = {'maxiter': self.maxiter, 'disp': self.disp, 'return_all': False} retval = optimize.minimize(self.func, self.startparams, method='Newton-CG', jac=self.grad, hessp=self.hessp, args=(), options=opts)['x'] else: retval = optimize.fmin_ncg(self.func, self.startparams, self.grad, fhess_p=self.hessp, args=(), maxiter=self.maxiter, full_output=False, disp=self.disp, retall=False) params = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.7.0. Don't allow them to increase. assert_(self.funccalls <= 7, self.funccalls) # gh10673 assert_(self.gradcalls <= 18, self.gradcalls) # 0.9.0 # assert_(self.gradcalls == 18, self.gradcalls) # 0.8.0 # assert_(self.gradcalls == 22, self.gradcalls) # 0.7.0 # Ensure that the function behaves the same; this is from SciPy 0.7.0 assert_allclose(self.trace[3:5], [[-4.35700753e-07, -5.24869435e-01, 4.87527480e-01], [-4.35700753e-07, -5.24869401e-01, 4.87527774e-01]], atol=1e-6, rtol=1e-7) def test_obj_func_returns_scalar(): match = ("The user-provided " "objective function must " "return a scalar value.") with assert_raises(ValueError, match=match): optimize.minimize(lambda x: x, np.array([1, 1]), method='BFGS') def test_neldermead_xatol_fatol(): # gh4484 # test we can call with fatol, xatol specified func = lambda x: x[0]**2 + x[1]**2 optimize._minimize._minimize_neldermead(func, [1, 1], maxiter=2, xatol=1e-3, fatol=1e-3) assert_warns(DeprecationWarning, optimize._minimize._minimize_neldermead, func, [1, 1], xtol=1e-3, ftol=1e-3, maxiter=2) def test_neldermead_adaptive(): func = lambda x: np.sum(x**2) p0 = [0.15746215, 0.48087031, 0.44519198, 0.4223638, 0.61505159, 0.32308456, 0.9692297, 0.4471682, 0.77411992, 0.80441652, 0.35994957, 0.75487856, 0.99973421, 0.65063887, 0.09626474] res = optimize.minimize(func, p0, method='Nelder-Mead') assert_equal(res.success, False) res = optimize.minimize(func, p0, method='Nelder-Mead', options={'adaptive': True}) assert_equal(res.success, True) def test_bounded_powell_outsidebounds(): # With the bounded Powell method if you start outside the bounds the final # should still be within the bounds (provided that the user doesn't make a # bad choice for the `direc` argument). func = lambda x: np.sum(x**2) bounds = (-1, 1), (-1, 1), (-1, 1) x0 = [-4, .5, -.8] # we're starting outside the bounds, so we should get a warning with assert_warns(optimize.OptimizeWarning): res = optimize.minimize(func, x0, bounds=bounds, method="Powell") assert_allclose(res.x, np.array([0.] * len(x0)), atol=1e-6) assert_equal(res.success, True) assert_equal(res.status, 0) # However, now if we change the `direc` argument such that the # set of vectors does not span the parameter space, then we may # not end up back within the bounds. Here we see that the first # parameter cannot be updated! direc = [[0, 0, 0], [0, 1, 0], [0, 0, 1]] # we're starting outside the bounds, so we should get a warning with assert_warns(optimize.OptimizeWarning): res = optimize.minimize(func, x0, bounds=bounds, method="Powell", options={'direc': direc}) assert_allclose(res.x, np.array([-4., 0, 0]), atol=1e-6) assert_equal(res.success, False) assert_equal(res.status, 4) def test_bounded_powell_vs_powell(): # here we test an example where the bounded Powell method # will return a different result than the standard Powell # method. # first we test a simple example where the minimum is at # the origin and the minimum that is within the bounds is # larger than the minimum at the origin. func = lambda x: np.sum(x**2) bounds = (-5, -1), (-10, -0.1), (1, 9.2), (-4, 7.6), (-15.9, -2) x0 = [-2.1, -5.2, 1.9, 0, -2] options = {'ftol': 1e-10, 'xtol': 1e-10} res_powell = optimize.minimize(func, x0, method="Powell", options=options) assert_allclose(res_powell.x, 0., atol=1e-6) assert_allclose(res_powell.fun, 0., atol=1e-6) res_bounded_powell = optimize.minimize(func, x0, options=options, bounds=bounds, method="Powell") p = np.array([-1, -0.1, 1, 0, -2]) assert_allclose(res_bounded_powell.x, p, atol=1e-6) assert_allclose(res_bounded_powell.fun, func(p), atol=1e-6) # now we test bounded Powell but with a mix of inf bounds. bounds = (None, -1), (-np.inf, -.1), (1, np.inf), (-4, None), (-15.9, -2) res_bounded_powell = optimize.minimize(func, x0, options=options, bounds=bounds, method="Powell") p = np.array([-1, -0.1, 1, 0, -2]) assert_allclose(res_bounded_powell.x, p, atol=1e-6) assert_allclose(res_bounded_powell.fun, func(p), atol=1e-6) # next we test an example where the global minimum is within # the bounds, but the bounded Powell method performs better # than the standard Powell method. def func(x): t = np.sin(-x[0]) * np.cos(x[1]) * np.sin(-x[0] * x[1]) * np.cos(x[1]) t -= np.cos(np.sin(x[1] * x[2]) * np.cos(x[2])) return t**2 bounds = [(-2, 5)] * 3 x0 = [-0.5, -0.5, -0.5] res_powell = optimize.minimize(func, x0, method="Powell") res_bounded_powell = optimize.minimize(func, x0, bounds=bounds, method="Powell") assert_allclose(res_powell.fun, 0.007136253919761627, atol=1e-6) assert_allclose(res_bounded_powell.fun, 0, atol=1e-6) # next we test the previous example where the we provide Powell # with (-inf, inf) bounds, and compare it to providing Powell # with no bounds. They should end up the same. bounds = [(-np.inf, np.inf)] * 3 res_bounded_powell = optimize.minimize(func, x0, bounds=bounds, method="Powell") assert_allclose(res_powell.fun, res_bounded_powell.fun, atol=1e-6) assert_allclose(res_powell.nfev, res_bounded_powell.nfev, atol=1e-6) assert_allclose(res_powell.x, res_bounded_powell.x, atol=1e-6) # now test when x0 starts outside of the bounds. x0 = [45.46254415, -26.52351498, 31.74830248] bounds = [(-2, 5)] * 3 # we're starting outside the bounds, so we should get a warning with assert_warns(optimize.OptimizeWarning): res_bounded_powell = optimize.minimize(func, x0, bounds=bounds, method="Powell") assert_allclose(res_bounded_powell.fun, 0, atol=1e-6) def test_onesided_bounded_powell_stability(): # When the Powell method is bounded on only one side, a # np.tan transform is done in order to convert it into a # completely bounded problem. Here we do some simple tests # of one-sided bounded Powell where the optimal solutions # are large to test the stability of the transformation. kwargs = {'method': 'Powell', 'bounds': [(-np.inf, 1e6)] * 3, 'options': {'ftol': 1e-8, 'xtol': 1e-8}} x0 = [1, 1, 1] # df/dx is constant. f = lambda x: -np.sum(x) res = optimize.minimize(f, x0, **kwargs) assert_allclose(res.fun, -3e6, atol=1e-4) # df/dx gets smaller and smaller. def f(x): return -np.abs(np.sum(x)) ** (0.1) * (1 if np.all(x > 0) else -1) res = optimize.minimize(f, x0, **kwargs) assert_allclose(res.fun, -(3e6) ** (0.1)) # df/dx gets larger and larger. def f(x): return -np.abs(np.sum(x)) ** 10 * (1 if np.all(x > 0) else -1) res = optimize.minimize(f, x0, **kwargs) assert_allclose(res.fun, -(3e6) ** 10, rtol=1e-7) # df/dx gets larger for some of the variables and smaller for others. def f(x): t = -np.abs(np.sum(x[:2])) ** 5 - np.abs(np.sum(x[2:])) ** (0.1) t *= (1 if np.all(x > 0) else -1) return t kwargs['bounds'] = [(-np.inf, 1e3)] * 3 res = optimize.minimize(f, x0, **kwargs) assert_allclose(res.fun, -(2e3) ** 5 - (1e6) ** (0.1), rtol=1e-7) class TestOptimizeWrapperDisp(CheckOptimizeParameterized): use_wrapper = True disp = True class TestOptimizeWrapperNoDisp(CheckOptimizeParameterized): use_wrapper = True disp = False class TestOptimizeNoWrapperDisp(CheckOptimizeParameterized): use_wrapper = False disp = True class TestOptimizeNoWrapperNoDisp(CheckOptimizeParameterized): use_wrapper = False disp = False class TestOptimizeSimple(CheckOptimize): def test_bfgs_nan(self): # Test corner case where nan is fed to optimizer. See gh-2067. func = lambda x: x fprime = lambda x: np.ones_like(x) x0 = [np.nan] with np.errstate(over='ignore', invalid='ignore'): x = optimize.fmin_bfgs(func, x0, fprime, disp=False) assert_(np.isnan(func(x))) def test_bfgs_nan_return(self): # Test corner cases where fun returns NaN. See gh-4793. # First case: NaN from first call. func = lambda x: np.nan with np.errstate(invalid='ignore'): result = optimize.minimize(func, 0) assert_(np.isnan(result['fun'])) assert_(result['success'] is False) # Second case: NaN from second call. func = lambda x: 0 if x == 0 else np.nan fprime = lambda x: np.ones_like(x) # Steer away from zero. with np.errstate(invalid='ignore'): result = optimize.minimize(func, 0, jac=fprime) assert_(np.isnan(result['fun'])) assert_(result['success'] is False) def test_bfgs_numerical_jacobian(self): # BFGS with numerical Jacobian and a vector epsilon parameter. # define the epsilon parameter using a random vector epsilon = np.sqrt(np.spacing(1.)) * np.random.rand(len(self.solution)) params = optimize.fmin_bfgs(self.func, self.startparams, epsilon=epsilon, args=(), maxiter=self.maxiter, disp=False) assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) def test_finite_differences(self): methods = ['BFGS', 'CG', 'TNC'] jacs = ['2-point', '3-point', None] for method, jac in itertools.product(methods, jacs): result = optimize.minimize(self.func, self.startparams, method=method, jac=jac) assert_allclose(self.func(result.x), self.func(self.solution), atol=1e-6) def test_bfgs_gh_2169(self): def f(x): if x < 0: return 1.79769313e+308 else: return x + 1./x xs = optimize.fmin_bfgs(f, [10.], disp=False) assert_allclose(xs, 1.0, rtol=1e-4, atol=1e-4) def test_bfgs_double_evaluations(self): # check BFGS does not evaluate twice in a row at same point def f(x): xp = float(x) assert xp not in seen seen.add(xp) return 10*x**2, 20*x seen = set() optimize.minimize(f, -100, method='bfgs', jac=True, tol=1e-7) def test_l_bfgs_b(self): # limited-memory bound-constrained BFGS algorithm retval = optimize.fmin_l_bfgs_b(self.func, self.startparams, self.grad, args=(), maxiter=self.maxiter) (params, fopt, d) = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) # Ensure that function call counts are 'known good'; these are from # SciPy 0.7.0. Don't allow them to increase. assert_(self.funccalls == 7, self.funccalls) assert_(self.gradcalls == 5, self.gradcalls) # Ensure that the function behaves the same; this is from SciPy 0.7.0 # test fixed in gh10673 assert_allclose(self.trace[3:5], [[8.117083e-16, -5.196198e-01, 4.897617e-01], [0., -0.52489628, 0.48753042]], atol=1e-14, rtol=1e-7) def test_l_bfgs_b_numjac(self): # L-BFGS-B with numerical Jacobian retval = optimize.fmin_l_bfgs_b(self.func, self.startparams, approx_grad=True, maxiter=self.maxiter) (params, fopt, d) = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) def test_l_bfgs_b_funjac(self): # L-BFGS-B with combined objective function and Jacobian def fun(x): return self.func(x), self.grad(x) retval = optimize.fmin_l_bfgs_b(fun, self.startparams, maxiter=self.maxiter) (params, fopt, d) = retval assert_allclose(self.func(params), self.func(self.solution), atol=1e-6) def test_l_bfgs_b_maxiter(self): # gh7854 # Ensure that not more than maxiters are ever run. class Callback(object): def __init__(self): self.nit = 0 self.fun = None self.x = None def __call__(self, x): self.x = x self.fun = optimize.rosen(x) self.nit += 1 c = Callback() res = optimize.minimize(optimize.rosen, [0., 0.], method='l-bfgs-b', callback=c, options={'maxiter': 5}) assert_equal(res.nit, 5) assert_almost_equal(res.x, c.x) assert_almost_equal(res.fun, c.fun) assert_equal(res.status, 1) assert_(res.success is False) assert_equal(res.message.decode(), 'STOP: TOTAL NO. of ITERATIONS REACHED LIMIT') def test_minimize_l_bfgs_b(self): # Minimize with L-BFGS-B method opts = {'disp': False, 'maxiter': self.maxiter} r = optimize.minimize(self.func, self.startparams, method='L-BFGS-B', jac=self.grad, options=opts) assert_allclose(self.func(r.x), self.func(self.solution), atol=1e-6) assert self.gradcalls == r.njev self.funccalls = self.gradcalls = 0 # approximate jacobian ra = optimize.minimize(self.func, self.startparams, method='L-BFGS-B', options=opts) # check that function evaluations in approximate jacobian are counted # assert_(ra.nfev > r.nfev) assert self.funccalls == ra.nfev assert_allclose(self.func(ra.x), self.func(self.solution), atol=1e-6) self.funccalls = self.gradcalls = 0 # approximate jacobian ra = optimize.minimize(self.func, self.startparams, jac='3-point', method='L-BFGS-B', options=opts) assert self.funccalls == ra.nfev assert_allclose(self.func(ra.x), self.func(self.solution), atol=1e-6) def test_minimize_l_bfgs_b_ftol(self): # Check that the `ftol` parameter in l_bfgs_b works as expected v0 = None for tol in [1e-1, 1e-4, 1e-7, 1e-10]: opts = {'disp': False, 'maxiter': self.maxiter, 'ftol': tol} sol = optimize.minimize(self.func, self.startparams, method='L-BFGS-B', jac=self.grad, options=opts) v = self.func(sol.x) if v0 is None: v0 = v else: assert_(v < v0) assert_allclose(v, self.func(self.solution), rtol=tol) def test_minimize_l_bfgs_maxls(self): # check that the maxls is passed down to the Fortran routine sol = optimize.minimize(optimize.rosen, np.array([-1.2, 1.0]), method='L-BFGS-B', jac=optimize.rosen_der, options={'disp': False, 'maxls': 1}) assert_(not sol.success) def test_minimize_l_bfgs_b_maxfun_interruption(self): # gh-6162 f = optimize.rosen g = optimize.rosen_der values = [] x0 = np.full(7, 1000) def objfun(x): value = f(x) values.append(value) return value # Look for an interesting test case. # Request a maxfun that stops at a particularly bad function # evaluation somewhere between 100 and 300 evaluations. low, medium, high = 30, 100, 300 optimize.fmin_l_bfgs_b(objfun, x0, fprime=g, maxfun=high) v, k = max((y, i) for i, y in enumerate(values[medium:])) maxfun = medium + k # If the minimization strategy is reasonable, # the minimize() result should not be worse than the best # of the first 30 function evaluations. target = min(values[:low]) xmin, fmin, d = optimize.fmin_l_bfgs_b(f, x0, fprime=g, maxfun=maxfun) assert_array_less(fmin, target) def test_custom(self): # This function comes from the documentation example. def custmin(fun, x0, args=(), maxfev=None, stepsize=0.1, maxiter=100, callback=None, **options): bestx = x0 besty = fun(x0) funcalls = 1 niter = 0 improved = True stop = False while improved and not stop and niter < maxiter: improved = False niter += 1 for dim in range(np.size(x0)): for s in [bestx[dim] - stepsize, bestx[dim] + stepsize]: testx = np.copy(bestx) testx[dim] = s testy = fun(testx, *args) funcalls += 1 if testy < besty: besty = testy bestx = testx improved = True if callback is not None: callback(bestx) if maxfev is not None and funcalls >= maxfev: stop = True break return optimize.OptimizeResult(fun=besty, x=bestx, nit=niter, nfev=funcalls, success=(niter > 1)) x0 = [1.35, 0.9, 0.8, 1.1, 1.2] res = optimize.minimize(optimize.rosen, x0, method=custmin, options=dict(stepsize=0.05)) assert_allclose(res.x, 1.0, rtol=1e-4, atol=1e-4) def test_gh10771(self): # check that minimize passes bounds and constraints to a custom # minimizer without altering them. bounds = [(-2, 2), (0, 3)] constraints = 'constraints' def custmin(fun, x0, **options): assert options['bounds'] is bounds assert options['constraints'] is constraints return optimize.OptimizeResult() x0 = [1, 1] optimize.minimize(optimize.rosen, x0, method=custmin, bounds=bounds, constraints=constraints) def test_minimize_tol_parameter(self): # Check that the minimize() tol= argument does something def func(z): x, y = z return x**2*y**2 + x**4 + 1 def dfunc(z): x, y = z return np.array([2*x*y**2 + 4*x**3, 2*x**2*y]) for method in ['nelder-mead', 'powell', 'cg', 'bfgs', 'newton-cg', 'l-bfgs-b', 'tnc', 'cobyla', 'slsqp']: if method in ('nelder-mead', 'powell', 'cobyla'): jac = None else: jac = dfunc sol1 = optimize.minimize(func, [1, 1], jac=jac, tol=1e-10, method=method) sol2 = optimize.minimize(func, [1, 1], jac=jac, tol=1.0, method=method) assert_(func(sol1.x) < func(sol2.x), "%s: %s vs. %s" % (method, func(sol1.x), func(sol2.x))) @pytest.mark.parametrize('method', ['fmin', 'fmin_powell', 'fmin_cg', 'fmin_bfgs', 'fmin_ncg', 'fmin_l_bfgs_b', 'fmin_tnc', 'fmin_slsqp'] + MINIMIZE_METHODS) def test_minimize_callback_copies_array(self, method): # Check that arrays passed to callbacks are not modified # inplace by the optimizer afterward # cobyla doesn't have callback if method == 'cobyla': return if method in ('fmin_tnc', 'fmin_l_bfgs_b'): func = lambda x: (optimize.rosen(x), optimize.rosen_der(x)) else: func = optimize.rosen jac = optimize.rosen_der hess = optimize.rosen_hess x0 = np.zeros(10) # Set options kwargs = {} if method.startswith('fmin'): routine = getattr(optimize, method) if method == 'fmin_slsqp': kwargs['iter'] = 5 elif method == 'fmin_tnc': kwargs['maxfun'] = 100 else: kwargs['maxiter'] = 5 else: def routine(*a, **kw): kw['method'] = method return optimize.minimize(*a, **kw) if method == 'tnc': kwargs['options'] = dict(maxfun=100) else: kwargs['options'] = dict(maxiter=5) if method in ('fmin_ncg',): kwargs['fprime'] = jac elif method in ('newton-cg',): kwargs['jac'] = jac elif method in ('trust-krylov', 'trust-exact', 'trust-ncg', 'dogleg', 'trust-constr'): kwargs['jac'] = jac kwargs['hess'] = hess # Run with callback results = [] def callback(x, *args, **kwargs): results.append((x, np.copy(x))) routine(func, x0, callback=callback, **kwargs) # Check returned arrays coincide with their copies # and have no memory overlap assert_(len(results) > 2) assert_(all(np.all(x == y) for x, y in results)) assert_(not any(np.may_share_memory(x[0], y[0]) for x, y in itertools.combinations(results, 2))) @pytest.mark.parametrize('method', ['nelder-mead', 'powell', 'cg', 'bfgs', 'newton-cg', 'l-bfgs-b', 'tnc', 'cobyla', 'slsqp']) def test_no_increase(self, method): # Check that the solver doesn't return a value worse than the # initial point. def func(x): return (x - 1)**2 def bad_grad(x): # purposefully invalid gradient function, simulates a case # where line searches start failing return 2*(x - 1) * (-1) - 2 x0 = np.array([2.0]) f0 = func(x0) jac = bad_grad if method in ['nelder-mead', 'powell', 'cobyla']: jac = None sol = optimize.minimize(func, x0, jac=jac, method=method, options=dict(maxiter=20)) assert_equal(func(sol.x), sol.fun) if method == 'slsqp': pytest.xfail("SLSQP returns slightly worse") assert_(func(sol.x) <= f0) def test_slsqp_respect_bounds(self): # Regression test for gh-3108 def f(x): return sum((x - np.array([1., 2., 3., 4.]))**2) def cons(x): a = np.array([[-1, -1, -1, -1], [-3, -3, -2, -1]]) return np.concatenate([np.dot(a, x) + np.array([5, 10]), x]) x0 = np.array([0.5, 1., 1.5, 2.]) res = optimize.minimize(f, x0, method='slsqp', constraints={'type': 'ineq', 'fun': cons}) assert_allclose(res.x, np.array([0., 2, 5, 8])/3, atol=1e-12) @pytest.mark.parametrize('method', ['Nelder-Mead', 'Powell', 'CG', 'BFGS', 'Newton-CG', 'L-BFGS-B', 'SLSQP', 'trust-constr', 'dogleg', 'trust-ncg', 'trust-exact', 'trust-krylov']) def test_respect_maxiter(self, method): # Check that the number of iterations equals max_iter, assuming # convergence doesn't establish before MAXITER = 4 x0 = np.zeros(10) sf = ScalarFunction(optimize.rosen, x0, (), optimize.rosen_der, optimize.rosen_hess, None, None) # Set options kwargs = {'method': method, 'options': dict(maxiter=MAXITER)} if method in ('Newton-CG',): kwargs['jac'] = sf.grad elif method in ('trust-krylov', 'trust-exact', 'trust-ncg', 'dogleg', 'trust-constr'): kwargs['jac'] = sf.grad kwargs['hess'] = sf.hess sol = optimize.minimize(sf.fun, x0, **kwargs) assert sol.nit == MAXITER assert sol.nfev >= sf.nfev if hasattr(sol, 'njev'): assert sol.njev >= sf.ngev # method specific tests if method == 'SLSQP': assert sol.status == 9 # Iteration limit reached def test_respect_maxiter_trust_constr_ineq_constraints(self): # special case of minimization with trust-constr and inequality # constraints to check maxiter limit is obeyed when using internal # method 'tr_interior_point' MAXITER = 4 f = optimize.rosen jac = optimize.rosen_der hess = optimize.rosen_hess fun = lambda x: np.array([0.2 * x[0] - 0.4 * x[1] - 0.33 * x[2]]) cons = ({'type': 'ineq', 'fun': fun},) x0 = np.zeros(10) sol = optimize.minimize(f, x0, constraints=cons, jac=jac, hess=hess, method='trust-constr', options=dict(maxiter=MAXITER)) assert sol.nit == MAXITER def test_minimize_automethod(self): def f(x): return x**2 def cons(x): return x - 2 x0 = np.array([10.]) sol_0 = optimize.minimize(f, x0) sol_1 = optimize.minimize(f, x0, constraints=[{'type': 'ineq', 'fun': cons}]) sol_2 = optimize.minimize(f, x0, bounds=[(5, 10)]) sol_3 = optimize.minimize(f, x0, constraints=[{'type': 'ineq', 'fun': cons}], bounds=[(5, 10)]) sol_4 = optimize.minimize(f, x0, constraints=[{'type': 'ineq', 'fun': cons}], bounds=[(1, 10)]) for sol in [sol_0, sol_1, sol_2, sol_3, sol_4]: assert_(sol.success) assert_allclose(sol_0.x, 0, atol=1e-7) assert_allclose(sol_1.x, 2, atol=1e-7) assert_allclose(sol_2.x, 5, atol=1e-7) assert_allclose(sol_3.x, 5, atol=1e-7) assert_allclose(sol_4.x, 2, atol=1e-7) def test_minimize_coerce_args_param(self): # Regression test for gh-3503 def Y(x, c): return np.sum((x-c)**2) def dY_dx(x, c=None): return 2*(x-c) c = np.array([3, 1, 4, 1, 5, 9, 2, 6, 5, 3, 5]) xinit = np.random.randn(len(c)) optimize.minimize(Y, xinit, jac=dY_dx, args=(c), method="BFGS") def test_initial_step_scaling(self): # Check that optimizer initial step is not huge even if the # function and gradients are scales = [1e-50, 1, 1e50] methods = ['CG', 'BFGS', 'L-BFGS-B', 'Newton-CG'] def f(x): if first_step_size[0] is None and x[0] != x0[0]: first_step_size[0] = abs(x[0] - x0[0]) if abs(x).max() > 1e4: raise AssertionError("Optimization stepped far away!") return scale*(x[0] - 1)**2 def g(x): return np.array([scale*(x[0] - 1)]) for scale, method in itertools.product(scales, methods): if method in ('CG', 'BFGS'): options = dict(gtol=scale*1e-8) else: options = dict() if scale < 1e-10 and method in ('L-BFGS-B', 'Newton-CG'): # XXX: return initial point if they see small gradient continue x0 = [-1.0] first_step_size = [None] res = optimize.minimize(f, x0, jac=g, method=method, options=options) err_msg = "{0} {1}: {2}: {3}".format(method, scale, first_step_size, res) assert_(res.success, err_msg) assert_allclose(res.x, [1.0], err_msg=err_msg) assert_(res.nit <= 3, err_msg) if scale > 1e-10: if method in ('CG', 'BFGS'): assert_allclose(first_step_size[0], 1.01, err_msg=err_msg) else: # Newton-CG and L-BFGS-B use different logic for the first # step, but are both scaling invariant with step sizes ~ 1 assert_(first_step_size[0] > 0.5 and first_step_size[0] < 3, err_msg) else: # step size has upper bound of ||grad||, so line # search makes many small steps pass @pytest.mark.parametrize('method', ['nelder-mead', 'powell', 'cg', 'bfgs', 'newton-cg', 'l-bfgs-b', 'tnc', 'cobyla', 'slsqp', 'trust-constr', 'dogleg', 'trust-ncg', 'trust-exact', 'trust-krylov']) def test_nan_values(self, method): # Check nan values result to failed exit status np.random.seed(1234) count = [0] def func(x): return np.nan def func2(x): count[0] += 1 if count[0] > 2: return np.nan else: return np.random.rand() def grad(x): return np.array([1.0]) def hess(x): return np.array([[1.0]]) x0 = np.array([1.0]) needs_grad = method in ('newton-cg', 'trust-krylov', 'trust-exact', 'trust-ncg', 'dogleg') needs_hess = method in ('trust-krylov', 'trust-exact', 'trust-ncg', 'dogleg') funcs = [func, func2] grads = [grad] if needs_grad else [grad, None] hesss = [hess] if needs_hess else [hess, None] with np.errstate(invalid='ignore'), suppress_warnings() as sup: sup.filter(UserWarning, "delta_grad == 0.*") sup.filter(RuntimeWarning, ".*does not use Hessian.*") sup.filter(RuntimeWarning, ".*does not use gradient.*") for f, g, h in itertools.product(funcs, grads, hesss): count = [0] sol = optimize.minimize(f, x0, jac=g, hess=h, method=method, options=dict(maxiter=20)) assert_equal(sol.success, False) @pytest.mark.parametrize('method', ['nelder-mead', 'cg', 'bfgs', 'l-bfgs-b', 'tnc', 'cobyla', 'slsqp', 'trust-constr', 'dogleg', 'trust-ncg', 'trust-exact', 'trust-krylov']) def test_duplicate_evaluations(self, method): # check that there are no duplicate evaluations for any methods jac = hess = None if method in ('newton-cg', 'trust-krylov', 'trust-exact', 'trust-ncg', 'dogleg'): jac = self.grad if method in ('trust-krylov', 'trust-exact', 'trust-ncg', 'dogleg'): hess = self.hess with np.errstate(invalid='ignore'), suppress_warnings() as sup: # for trust-constr sup.filter(UserWarning, "delta_grad == 0.*") optimize.minimize(self.func, self.startparams, method=method, jac=jac, hess=hess) for i in range(1, len(self.trace)): if np.array_equal(self.trace[i - 1], self.trace[i]): raise RuntimeError( "Duplicate evaluations made by {}".format(method)) class TestLBFGSBBounds(object): def setup_method(self): self.bounds = ((1, None), (None, None)) self.solution = (1, 0) def fun(self, x, p=2.0): return 1.0 / p * (x[0]**p + x[1]**p) def jac(self, x, p=2.0): return x**(p - 1) def fj(self, x, p=2.0): return self.fun(x, p), self.jac(x, p) def test_l_bfgs_b_bounds(self): x, f, d = optimize.fmin_l_bfgs_b(self.fun, [0, -1], fprime=self.jac, bounds=self.bounds) assert_(d['warnflag'] == 0, d['task']) assert_allclose(x, self.solution, atol=1e-6) def test_l_bfgs_b_funjac(self): # L-BFGS-B with fun and jac combined and extra arguments x, f, d = optimize.fmin_l_bfgs_b(self.fj, [0, -1], args=(2.0, ), bounds=self.bounds) assert_(d['warnflag'] == 0, d['task']) assert_allclose(x, self.solution, atol=1e-6) def test_minimize_l_bfgs_b_bounds(self): # Minimize with method='L-BFGS-B' with bounds res = optimize.minimize(self.fun, [0, -1], method='L-BFGS-B', jac=self.jac, bounds=self.bounds) assert_(res['success'], res['message']) assert_allclose(res.x, self.solution, atol=1e-6) @pytest.mark.parametrize('bounds', [ ([(10, 1), (1, 10)]), ([(1, 10), (10, 1)]), ([(10, 1), (10, 1)]) ]) def test_minimize_l_bfgs_b_incorrect_bounds(self, bounds): with pytest.raises(ValueError, match='.*bounds.*'): optimize.minimize(self.fun, [0, -1], method='L-BFGS-B', jac=self.jac, bounds=bounds) def test_minimize_l_bfgs_b_bounds_FD(self): # test that initial starting value outside bounds doesn't raise # an error (done with clipping). # test all different finite differences combos, with and without args jacs = ['2-point', '3-point', None] argss = [(2.,), ()] for jac, args in itertools.product(jacs, argss): res = optimize.minimize(self.fun, [0, -1], args=args, method='L-BFGS-B', jac=jac, bounds=self.bounds, options={'finite_diff_rel_step': None}) assert_(res['success'], res['message']) assert_allclose(res.x, self.solution, atol=1e-6) class TestOptimizeScalar(object): def setup_method(self): self.solution = 1.5 def fun(self, x, a=1.5): """Objective function""" return (x - a)**2 - 0.8 def test_brent(self): x = optimize.brent(self.fun) assert_allclose(x, self.solution, atol=1e-6) x = optimize.brent(self.fun, brack=(-3, -2)) assert_allclose(x, self.solution, atol=1e-6) x = optimize.brent(self.fun, full_output=True) assert_allclose(x[0], self.solution, atol=1e-6) x = optimize.brent(self.fun, brack=(-15, -1, 15)) assert_allclose(x, self.solution, atol=1e-6) def test_golden(self): x = optimize.golden(self.fun) assert_allclose(x, self.solution, atol=1e-6) x = optimize.golden(self.fun, brack=(-3, -2)) assert_allclose(x, self.solution, atol=1e-6) x = optimize.golden(self.fun, full_output=True) assert_allclose(x[0], self.solution, atol=1e-6) x = optimize.golden(self.fun, brack=(-15, -1, 15)) assert_allclose(x, self.solution, atol=1e-6) x = optimize.golden(self.fun, tol=0) assert_allclose(x, self.solution) maxiter_test_cases = [0, 1, 5] for maxiter in maxiter_test_cases: x0 = optimize.golden(self.fun, maxiter=0, full_output=True) x = optimize.golden(self.fun, maxiter=maxiter, full_output=True) nfev0, nfev = x0[2], x[2] assert_equal(nfev - nfev0, maxiter) def test_fminbound(self): x = optimize.fminbound(self.fun, 0, 1) assert_allclose(x, 1, atol=1e-4) x = optimize.fminbound(self.fun, 1, 5) assert_allclose(x, self.solution, atol=1e-6) x = optimize.fminbound(self.fun, np.array([1]), np.array([5])) assert_allclose(x, self.solution, atol=1e-6) assert_raises(ValueError, optimize.fminbound, self.fun, 5, 1) def test_fminbound_scalar(self): with pytest.raises(ValueError, match='.*must be scalar.*'): optimize.fminbound(self.fun, np.zeros((1, 2)), 1) x = optimize.fminbound(self.fun, 1, np.array(5)) assert_allclose(x, self.solution, atol=1e-6) def test_gh11207(self): def fun(x): return x**2 optimize.fminbound(fun, 0, 0) def test_minimize_scalar(self): # combine all tests above for the minimize_scalar wrapper x = optimize.minimize_scalar(self.fun).x assert_allclose(x, self.solution, atol=1e-6) x = optimize.minimize_scalar(self.fun, method='Brent') assert_(x.success) x = optimize.minimize_scalar(self.fun, method='Brent', options=dict(maxiter=3)) assert_(not x.success) x = optimize.minimize_scalar(self.fun, bracket=(-3, -2), args=(1.5, ), method='Brent').x assert_allclose(x, self.solution, atol=1e-6) x = optimize.minimize_scalar(self.fun, method='Brent', args=(1.5,)).x assert_allclose(x, self.solution, atol=1e-6) x = optimize.minimize_scalar(self.fun, bracket=(-15, -1, 15), args=(1.5, ), method='Brent').x assert_allclose(x, self.solution, atol=1e-6) x = optimize.minimize_scalar(self.fun, bracket=(-3, -2), args=(1.5, ), method='golden').x assert_allclose(x, self.solution, atol=1e-6) x = optimize.minimize_scalar(self.fun, method='golden', args=(1.5,)).x assert_allclose(x, self.solution, atol=1e-6) x = optimize.minimize_scalar(self.fun, bracket=(-15, -1, 15), args=(1.5, ), method='golden').x assert_allclose(x, self.solution, atol=1e-6) x = optimize.minimize_scalar(self.fun, bounds=(0, 1), args=(1.5,), method='Bounded').x assert_allclose(x, 1, atol=1e-4) x = optimize.minimize_scalar(self.fun, bounds=(1, 5), args=(1.5, ), method='bounded').x assert_allclose(x, self.solution, atol=1e-6) x = optimize.minimize_scalar(self.fun, bounds=(np.array([1]), np.array([5])), args=(np.array([1.5]), ), method='bounded').x assert_allclose(x, self.solution, atol=1e-6) assert_raises(ValueError, optimize.minimize_scalar, self.fun, bounds=(5, 1), method='bounded', args=(1.5, )) assert_raises(ValueError, optimize.minimize_scalar, self.fun, bounds=(np.zeros(2), 1), method='bounded', args=(1.5, )) x = optimize.minimize_scalar(self.fun, bounds=(1, np.array(5)), method='bounded').x assert_allclose(x, self.solution, atol=1e-6) def test_minimize_scalar_custom(self): # This function comes from the documentation example. def custmin(fun, bracket, args=(), maxfev=None, stepsize=0.1, maxiter=100, callback=None, **options): bestx = (bracket[1] + bracket[0]) / 2.0 besty = fun(bestx) funcalls = 1 niter = 0 improved = True stop = False while improved and not stop and niter < maxiter: improved = False niter += 1 for testx in [bestx - stepsize, bestx + stepsize]: testy = fun(testx, *args) funcalls += 1 if testy < besty: besty = testy bestx = testx improved = True if callback is not None: callback(bestx) if maxfev is not None and funcalls >= maxfev: stop = True break return optimize.OptimizeResult(fun=besty, x=bestx, nit=niter, nfev=funcalls, success=(niter > 1)) res = optimize.minimize_scalar(self.fun, bracket=(0, 4), method=custmin, options=dict(stepsize=0.05)) assert_allclose(res.x, self.solution, atol=1e-6) def test_minimize_scalar_coerce_args_param(self): # Regression test for gh-3503 optimize.minimize_scalar(self.fun, args=1.5) @pytest.mark.parametrize('method', ['brent', 'bounded', 'golden']) def test_nan_values(self, method): # Check nan values result to failed exit status np.random.seed(1234) count = [0] def func(x): count[0] += 1 if count[0] > 4: return np.nan else: return x**2 + 0.1 * np.sin(x) bracket = (-1, 0, 1) bounds = (-1, 1) with np.errstate(invalid='ignore'), suppress_warnings() as sup: sup.filter(UserWarning, "delta_grad == 0.*") sup.filter(RuntimeWarning, ".*does not use Hessian.*") sup.filter(RuntimeWarning, ".*does not use gradient.*") count = [0] sol = optimize.minimize_scalar(func, bracket=bracket, bounds=bounds, method=method, options=dict(maxiter=20)) assert_equal(sol.success, False) def test_brent_negative_tolerance(): assert_raises(ValueError, optimize.brent, np.cos, tol=-.01) class TestNewtonCg(object): def test_rosenbrock(self): x0 = np.array([-1.2, 1.0]) sol = optimize.minimize(optimize.rosen, x0, jac=optimize.rosen_der, hess=optimize.rosen_hess, tol=1e-5, method='Newton-CG') assert_(sol.success, sol.message) assert_allclose(sol.x, np.array([1, 1]), rtol=1e-4) def test_himmelblau(self): x0 = np.array(himmelblau_x0) sol = optimize.minimize(himmelblau, x0, jac=himmelblau_grad, hess=himmelblau_hess, method='Newton-CG', tol=1e-6) assert_(sol.success, sol.message) assert_allclose(sol.x, himmelblau_xopt, rtol=1e-4) assert_allclose(sol.fun, himmelblau_min, atol=1e-4) def test_line_for_search(): # _line_for_search is only used in _linesearch_powell, which is also # tested below. Thus there are more tests of _line_for_search in the # test_linesearch_powell_bounded function. line_for_search = optimize.optimize._line_for_search # args are x0, alpha, lower_bound, upper_bound # returns lmin, lmax lower_bound = np.array([-5.3, -1, -1.5, -3]) upper_bound = np.array([1.9, 1, 2.8, 3]) # test when starting in the bounds x0 = np.array([0., 0, 0, 0]) # and when starting outside of the bounds x1 = np.array([0., 2, -3, 0]) all_tests = ( (x0, np.array([1., 0, 0, 0]), -5.3, 1.9), (x0, np.array([0., 1, 0, 0]), -1, 1), (x0, np.array([0., 0, 1, 0]), -1.5, 2.8), (x0, np.array([0., 0, 0, 1]), -3, 3), (x0, np.array([1., 1, 0, 0]), -1, 1), (x0, np.array([1., 0, -1, 2]), -1.5, 1.5), (x0, np.array([2., 0, -1, 2]), -1.5, 0.95), (x1, np.array([1., 0, 0, 0]), -5.3, 1.9), (x1, np.array([0., 1, 0, 0]), -3, -1), (x1, np.array([0., 0, 1, 0]), 1.5, 5.8), (x1, np.array([0., 0, 0, 1]), -3, 3), (x1, np.array([1., 1, 0, 0]), -3, -1), (x1, np.array([1., 0, -1, 0]), -5.3, -1.5), ) for x, alpha, lmin, lmax in all_tests: mi, ma = line_for_search(x, alpha, lower_bound, upper_bound) assert_allclose(mi, lmin, atol=1e-6) assert_allclose(ma, lmax, atol=1e-6) # now with infinite bounds lower_bound = np.array([-np.inf, -1, -np.inf, -3]) upper_bound = np.array([np.inf, 1, 2.8, np.inf]) all_tests = ( (x0, np.array([1., 0, 0, 0]), -np.inf, np.inf), (x0, np.array([0., 1, 0, 0]), -1, 1), (x0, np.array([0., 0, 1, 0]), -np.inf, 2.8), (x0, np.array([0., 0, 0, 1]), -3, np.inf), (x0, np.array([1., 1, 0, 0]), -1, 1), (x0, np.array([1., 0, -1, 2]), -1.5, np.inf), (x1, np.array([1., 0, 0, 0]), -np.inf, np.inf), (x1, np.array([0., 1, 0, 0]), -3, -1), (x1, np.array([0., 0, 1, 0]), -np.inf, 5.8), (x1, np.array([0., 0, 0, 1]), -3, np.inf), (x1, np.array([1., 1, 0, 0]), -3, -1), (x1, np.array([1., 0, -1, 0]), -5.8, np.inf), ) for x, alpha, lmin, lmax in all_tests: mi, ma = line_for_search(x, alpha, lower_bound, upper_bound) assert_allclose(mi, lmin, atol=1e-6) assert_allclose(ma, lmax, atol=1e-6) def test_linesearch_powell(): # helper function in optimize.py, not a public function. linesearch_powell = optimize.optimize._linesearch_powell # args are func, p, xi, fval, lower_bound=None, upper_bound=None, tol=1e-3 # returns new_fval, p + direction, direction func = lambda x: np.sum((x - np.array([-1., 2., 1.5, -.4]))**2) p0 = np.array([0., 0, 0, 0]) fval = func(p0) lower_bound = np.array([-np.inf] * 4) upper_bound = np.array([np.inf] * 4) all_tests = ( (np.array([1., 0, 0, 0]), -1), (np.array([0., 1, 0, 0]), 2), (np.array([0., 0, 1, 0]), 1.5), (np.array([0., 0, 0, 1]), -.4), (np.array([-1., 0, 1, 0]), 1.25), (np.array([0., 0, 1, 1]), .55), (np.array([2., 0, -1, 1]), -.65), ) for xi, l in all_tests: f, p, direction = linesearch_powell(func, p0, xi, fval=fval, tol=1e-5) assert_allclose(f, func(l * xi), atol=1e-6) assert_allclose(p, l * xi, atol=1e-6) assert_allclose(direction, l * xi, atol=1e-6) f, p, direction = linesearch_powell(func, p0, xi, tol=1e-5, lower_bound=lower_bound, upper_bound=upper_bound, fval=fval) assert_allclose(f, func(l * xi), atol=1e-6) assert_allclose(p, l * xi, atol=1e-6) assert_allclose(direction, l * xi, atol=1e-6) def test_linesearch_powell_bounded(): # helper function in optimize.py, not a public function. linesearch_powell = optimize.optimize._linesearch_powell # args are func, p, xi, fval, lower_bound=None, upper_bound=None, tol=1e-3 # returns new_fval, p+direction, direction func = lambda x: np.sum((x-np.array([-1., 2., 1.5, -.4]))**2) p0 = np.array([0., 0, 0, 0]) fval = func(p0) # first choose bounds such that the same tests from # test_linesearch_powell should pass. lower_bound = np.array([-2.]*4) upper_bound = np.array([2.]*4) all_tests = ( (np.array([1., 0, 0, 0]), -1), (np.array([0., 1, 0, 0]), 2), (np.array([0., 0, 1, 0]), 1.5), (np.array([0., 0, 0, 1]), -.4), (np.array([-1., 0, 1, 0]), 1.25), (np.array([0., 0, 1, 1]), .55), (np.array([2., 0, -1, 1]), -.65), ) for xi, l in all_tests: f, p, direction = linesearch_powell(func, p0, xi, tol=1e-5, lower_bound=lower_bound, upper_bound=upper_bound, fval=fval) assert_allclose(f, func(l * xi), atol=1e-6) assert_allclose(p, l * xi, atol=1e-6) assert_allclose(direction, l * xi, atol=1e-6) # now choose bounds such that unbounded vs bounded gives different results lower_bound = np.array([-.3]*3 + [-1]) upper_bound = np.array([.45]*3 + [.9]) all_tests = ( (np.array([1., 0, 0, 0]), -.3), (np.array([0., 1, 0, 0]), .45), (np.array([0., 0, 1, 0]), .45), (np.array([0., 0, 0, 1]), -.4), (np.array([-1., 0, 1, 0]), .3), (np.array([0., 0, 1, 1]), .45), (np.array([2., 0, -1, 1]), -.15), ) for xi, l in all_tests: f, p, direction = linesearch_powell(func, p0, xi, tol=1e-5, lower_bound=lower_bound, upper_bound=upper_bound, fval=fval) assert_allclose(f, func(l * xi), atol=1e-6) assert_allclose(p, l * xi, atol=1e-6) assert_allclose(direction, l * xi, atol=1e-6) # now choose as above but start outside the bounds p0 = np.array([-1., 0, 0, 2]) fval = func(p0) all_tests = ( (np.array([1., 0, 0, 0]), .7), (np.array([0., 1, 0, 0]), .45), (np.array([0., 0, 1, 0]), .45), (np.array([0., 0, 0, 1]), -2.4), ) for xi, l in all_tests: f, p, direction = linesearch_powell(func, p0, xi, tol=1e-5, lower_bound=lower_bound, upper_bound=upper_bound, fval=fval) assert_allclose(f, func(p0 + l * xi), atol=1e-6) assert_allclose(p, p0 + l * xi, atol=1e-6) assert_allclose(direction, l * xi, atol=1e-6) # now mix in inf p0 = np.array([0., 0, 0, 0]) fval = func(p0) # now choose bounds that mix inf lower_bound = np.array([-.3, -np.inf, -np.inf, -1]) upper_bound = np.array([np.inf, .45, np.inf, .9]) all_tests = ( (np.array([1., 0, 0, 0]), -.3), (np.array([0., 1, 0, 0]), .45), (np.array([0., 0, 1, 0]), 1.5), (np.array([0., 0, 0, 1]), -.4), (np.array([-1., 0, 1, 0]), .3), (np.array([0., 0, 1, 1]), .55), (np.array([2., 0, -1, 1]), -.15), ) for xi, l in all_tests: f, p, direction = linesearch_powell(func, p0, xi, tol=1e-5, lower_bound=lower_bound, upper_bound=upper_bound, fval=fval) assert_allclose(f, func(l * xi), atol=1e-6) assert_allclose(p, l * xi, atol=1e-6) assert_allclose(direction, l * xi, atol=1e-6) # now choose as above but start outside the bounds p0 = np.array([-1., 0, 0, 2]) fval = func(p0) all_tests = ( (np.array([1., 0, 0, 0]), .7), (np.array([0., 1, 0, 0]), .45), (np.array([0., 0, 1, 0]), 1.5), (np.array([0., 0, 0, 1]), -2.4), ) for xi, l in all_tests: f, p, direction = linesearch_powell(func, p0, xi, tol=1e-5, lower_bound=lower_bound, upper_bound=upper_bound, fval=fval) assert_allclose(f, func(p0 + l * xi), atol=1e-6) assert_allclose(p, p0 + l * xi, atol=1e-6) assert_allclose(direction, l * xi, atol=1e-6) class TestRosen(object): def test_hess(self): # Compare rosen_hess(x) times p with rosen_hess_prod(x,p). See gh-1775. x = np.array([3, 4, 5]) p = np.array([2, 2, 2]) hp = optimize.rosen_hess_prod(x, p) dothp = np.dot(optimize.rosen_hess(x), p) assert_equal(hp, dothp) def himmelblau(p): """ R^2 -> R^1 test function for optimization. The function has four local minima where himmelblau(xopt) == 0. """ x, y = p a = x*x + y - 11 b = x + y*y - 7 return a*a + b*b def himmelblau_grad(p): x, y = p return np.array([4*x**3 + 4*x*y - 42*x + 2*y**2 - 14, 2*x**2 + 4*x*y + 4*y**3 - 26*y - 22]) def himmelblau_hess(p): x, y = p return np.array([[12*x**2 + 4*y - 42, 4*x + 4*y], [4*x + 4*y, 4*x + 12*y**2 - 26]]) himmelblau_x0 = [-0.27, -0.9] himmelblau_xopt = [3, 2] himmelblau_min = 0.0 def test_minimize_multiple_constraints(): # Regression test for gh-4240. def func(x): return np.array([25 - 0.2 * x[0] - 0.4 * x[1] - 0.33 * x[2]]) def func1(x): return np.array([x[1]]) def func2(x): return np.array([x[2]]) cons = ({'type': 'ineq', 'fun': func}, {'type': 'ineq', 'fun': func1}, {'type': 'ineq', 'fun': func2}) f = lambda x: -1 * (x[0] + x[1] + x[2]) res = optimize.minimize(f, [0, 0, 0], method='SLSQP', constraints=cons) assert_allclose(res.x, [125, 0, 0], atol=1e-10) class TestOptimizeResultAttributes(object): # Test that all minimizers return an OptimizeResult containing # all the OptimizeResult attributes def setup_method(self): self.x0 = [5, 5] self.func = optimize.rosen self.jac = optimize.rosen_der self.hess = optimize.rosen_hess self.hessp = optimize.rosen_hess_prod self.bounds = [(0., 10.), (0., 10.)] def test_attributes_present(self): attributes = ['nit', 'nfev', 'x', 'success', 'status', 'fun', 'message'] skip = {'cobyla': ['nit']} for method in MINIMIZE_METHODS: with suppress_warnings() as sup: sup.filter(RuntimeWarning, ("Method .+ does not use (gradient|Hessian.*)" " information")) res = optimize.minimize(self.func, self.x0, method=method, jac=self.jac, hess=self.hess, hessp=self.hessp) for attribute in attributes: if method in skip and attribute in skip[method]: continue assert_(hasattr(res, attribute)) assert_(attribute in dir(res)) def f1(z, *params): x, y = z a, b, c, d, e, f, g, h, i, j, k, l, scale = params return (a * x**2 + b * x * y + c * y**2 + d*x + e*y + f) def f2(z, *params): x, y = z a, b, c, d, e, f, g, h, i, j, k, l, scale = params return (-g*np.exp(-((x-h)**2 + (y-i)**2) / scale)) def f3(z, *params): x, y = z a, b, c, d, e, f, g, h, i, j, k, l, scale = params return (-j*np.exp(-((x-k)**2 + (y-l)**2) / scale)) def brute_func(z, *params): return f1(z, *params) + f2(z, *params) + f3(z, *params) class TestBrute: # Test the "brute force" method def setup_method(self): self.params = (2, 3, 7, 8, 9, 10, 44, -1, 2, 26, 1, -2, 0.5) self.rranges = (slice(-4, 4, 0.25), slice(-4, 4, 0.25)) self.solution = np.array([-1.05665192, 1.80834843]) def brute_func(self, z, *params): # an instance method optimizing return brute_func(z, *params) def test_brute(self): # test fmin resbrute = optimize.brute(brute_func, self.rranges, args=self.params, full_output=True, finish=optimize.fmin) assert_allclose(resbrute[0], self.solution, atol=1e-3) assert_allclose(resbrute[1], brute_func(self.solution, *self.params), atol=1e-3) # test minimize resbrute = optimize.brute(brute_func, self.rranges, args=self.params, full_output=True, finish=optimize.minimize) assert_allclose(resbrute[0], self.solution, atol=1e-3) assert_allclose(resbrute[1], brute_func(self.solution, *self.params), atol=1e-3) # test that brute can optimize an instance method (the other tests use # a non-class based function resbrute = optimize.brute(self.brute_func, self.rranges, args=self.params, full_output=True, finish=optimize.minimize) assert_allclose(resbrute[0], self.solution, atol=1e-3) def test_1D(self): # test that for a 1-D problem the test function is passed an array, # not a scalar. def f(x): assert_(len(x.shape) == 1) assert_(x.shape[0] == 1) return x ** 2 optimize.brute(f, [(-1, 1)], Ns=3, finish=None) def test_workers(self): # check that parallel evaluation works resbrute = optimize.brute(brute_func, self.rranges, args=self.params, full_output=True, finish=None) resbrute1 = optimize.brute(brute_func, self.rranges, args=self.params, full_output=True, finish=None, workers=2) assert_allclose(resbrute1[-1], resbrute[-1]) assert_allclose(resbrute1[0], resbrute[0]) def test_cobyla_threadsafe(): # Verify that cobyla is threadsafe. Will segfault if it is not. import concurrent.futures import time def objective1(x): time.sleep(0.1) return x[0]**2 def objective2(x): time.sleep(0.1) return (x[0]-1)**2 min_method = "COBYLA" def minimizer1(): return optimize.minimize(objective1, [0.0], method=min_method) def minimizer2(): return optimize.minimize(objective2, [0.0], method=min_method) with concurrent.futures.ThreadPoolExecutor() as pool: tasks = [] tasks.append(pool.submit(minimizer1)) tasks.append(pool.submit(minimizer2)) for t in tasks: res = t.result() class TestIterationLimits(object): # Tests that optimisation does not give up before trying requested # number of iterations or evaluations. And that it does not succeed # by exceeding the limits. def setup_method(self): self.funcalls = 0 def slow_func(self, v): self.funcalls += 1 r, t = np.sqrt(v[0]**2+v[1]**2), np.arctan2(v[0], v[1]) return np.sin(r*20 + t)+r*0.5 def test_neldermead_limit(self): self.check_limits("Nelder-Mead", 200) def test_powell_limit(self): self.check_limits("powell", 1000) def check_limits(self, method, default_iters): for start_v in [[0.1, 0.1], [1, 1], [2, 2]]: for mfev in [50, 500, 5000]: self.funcalls = 0 res = optimize.minimize(self.slow_func, start_v, method=method, options={"maxfev": mfev}) assert_(self.funcalls == res["nfev"]) if res["success"]: assert_(res["nfev"] < mfev) else: assert_(res["nfev"] >= mfev) for mit in [50, 500, 5000]: res = optimize.minimize(self.slow_func, start_v, method=method, options={"maxiter": mit}) if res["success"]: assert_(res["nit"] <= mit) else: assert_(res["nit"] >= mit) for mfev, mit in [[50, 50], [5000, 5000], [5000, np.inf]]: self.funcalls = 0 res = optimize.minimize(self.slow_func, start_v, method=method, options={"maxiter": mit, "maxfev": mfev}) assert_(self.funcalls == res["nfev"]) if res["success"]: assert_(res["nfev"] < mfev and res["nit"] <= mit) else: assert_(res["nfev"] >= mfev or res["nit"] >= mit) for mfev, mit in [[np.inf, None], [None, np.inf]]: self.funcalls = 0 res = optimize.minimize(self.slow_func, start_v, method=method, options={"maxiter": mit, "maxfev": mfev}) assert_(self.funcalls == res["nfev"]) if res["success"]: if mfev is None: assert_(res["nfev"] < default_iters*2) else: assert_(res["nit"] <= default_iters*2) else: assert_(res["nfev"] >= default_iters*2 or res["nit"] >= default_iters*2) def test_result_x_shape_when_len_x_is_one(): def fun(x): return x * x def jac(x): return 2. * x def hess(x): return np.array([[2.]]) methods = ['Nelder-Mead', 'Powell', 'CG', 'BFGS', 'L-BFGS-B', 'TNC', 'COBYLA', 'SLSQP'] for method in methods: res = optimize.minimize(fun, np.array([0.1]), method=method) assert res.x.shape == (1,) # use jac + hess methods = ['trust-constr', 'dogleg', 'trust-ncg', 'trust-exact', 'trust-krylov', 'Newton-CG'] for method in methods: res = optimize.minimize(fun, np.array([0.1]), method=method, jac=jac, hess=hess) assert res.x.shape == (1,) class FunctionWithGradient(object): def __init__(self): self.number_of_calls = 0 def __call__(self, x): self.number_of_calls += 1 return np.sum(x**2), 2 * x @pytest.fixture def function_with_gradient(): return FunctionWithGradient() def test_memoize_jac_function_before_gradient(function_with_gradient): memoized_function = MemoizeJac(function_with_gradient) x0 = np.array([1.0, 2.0]) assert_allclose(memoized_function(x0), 5.0) assert function_with_gradient.number_of_calls == 1 assert_allclose(memoized_function.derivative(x0), 2 * x0) assert function_with_gradient.number_of_calls == 1, \ "function is not recomputed " \ "if gradient is requested after function value" assert_allclose( memoized_function(2 * x0), 20.0, err_msg="different input triggers new computation") assert function_with_gradient.number_of_calls == 2, \ "different input triggers new computation" def test_memoize_jac_gradient_before_function(function_with_gradient): memoized_function = MemoizeJac(function_with_gradient) x0 = np.array([1.0, 2.0]) assert_allclose(memoized_function.derivative(x0), 2 * x0) assert function_with_gradient.number_of_calls == 1 assert_allclose(memoized_function(x0), 5.0) assert function_with_gradient.number_of_calls == 1, \ "function is not recomputed " \ "if function value is requested after gradient" assert_allclose( memoized_function.derivative(2 * x0), 4 * x0, err_msg="different input triggers new computation") assert function_with_gradient.number_of_calls == 2, \ "different input triggers new computation" def test_memoize_jac_with_bfgs(function_with_gradient): """ Tests that using MemoizedJac in combination with ScalarFunction and BFGS does not lead to repeated function evaluations. Tests changes made in response to GH11868. """ memoized_function = MemoizeJac(function_with_gradient) jac = memoized_function.derivative hess = optimize.BFGS() x0 = np.array([1.0, 0.5]) scalar_function = ScalarFunction( memoized_function, x0, (), jac, hess, None, None) assert function_with_gradient.number_of_calls == 1 scalar_function.fun(x0 + 0.1) assert function_with_gradient.number_of_calls == 2 scalar_function.fun(x0 + 0.2) assert function_with_gradient.number_of_calls == 3