def test_gradient_UX_HY_wfnsim(simulator, angle, controlled, silent=True):
# same as before just with wavefunction simulation
# case YX
# U = cos(angle/2) + sin(-angle/2)*i*X
# O = cos*sin*i*<0|YX|0> + sin*cos*(-i)<0|XY|0>
# = 0.5*sin(-angle)*i <0|[YX,XY]|0>
# = -sin(angle)
angle_value = angle
angle = Variable(name="angle")
variables = {angle: angle_value}
qubit = 0
H = paulis.Y(qubit=qubit)
if controlled:
control = 1
U = gates.X(target=control) + gates.Rx(
target=qubit, control=control, angle=angle)
else:
U = gates.Rx(target=qubit, angle=angle)
O = ExpectationValue(U=U, H=H)
E = simulate(O, variables=variables, backend=simulator)
dO = grad(objective=O, variable=angle)
dE = simulate(dO, variables=variables)
assert (numpy.isclose(E, -numpy.sin(angle(variables)), atol=0.0001))
assert (numpy.isclose(dE, -numpy.cos(angle(variables)), atol=0.0001))
if not silent:
print("E =", E)
print("-sin(angle)=", -numpy.sin(angle(variables)))
print("dE =", dE)
print("-cos(angle)=", -numpy.cos(angle(variables)))
def test_gradient_UY_HX(simulator, angle_value, controlled, silent=True):
# case X Y
# U = cos(angle/2) + sin(-angle/2)*i*Y
# <0|Ud H U |0> = cos^2(angle/2)*<0|X|0>
# + sin^2(-angle/2) <0|YXY|0>
# + cos(angle/2)*sin(angle/2)*i<0|XY|0>
# + sin(-angle/2)*cos(angle/2)*(-i) <0|YX|0>
# = cos^2*0 + sin^2*0 + cos*sin*i(<0|[XY,YX]|0>)
# = 0.5*sin(-angle)*i <0|[XY,YX]|0> = -0.5*sin(angle)*i * 2 i <0|Z|0>
# = sin(angle)
angle = Variable(name="angle")
variables = {angle: angle_value}
qubit = 0
H = paulis.X(qubit=qubit)
if controlled:
control = 1
U = gates.X(target=control) + gates.Ry(
target=qubit, control=control, angle=angle)
else:
U = gates.X(target=qubit) + gates.X(target=qubit) + gates.Ry(
target=qubit, angle=angle)
O = ExpectationValue(U=U, H=H)
E = simulate(O, variables=variables, backend=simulator)
print("O={type}".format(type=type(O)))
dO = grad(objective=O, variable=angle)
dE = simulate(dO, variables=variables, backend=simulator)
assert (numpy.isclose(E, numpy.sin(angle(variables)), atol=1.e-4))
assert (numpy.isclose(dE, numpy.cos(angle(variables)), atol=1.e-4))
if not silent:
print("E =", E)
print("sin(angle)=", numpy.sin(angle()))
print("dE =", dE)
print("cos(angle)=", numpy.cos(angle()))
def test_gradient_PHASE_HY(simulator, angle_value, controlled, silent=False):
angle = Variable(name="angle")
variables = {angle: angle_value}
qubit = 0
H = paulis.Y(qubit=qubit)
if controlled:
control = 1
U = gates.X(target=control) + gates.H(target=qubit) + gates.Phase(
target=qubit, control=control, phi=angle) + gates.H(target=qubit)
else:
U = gates.H(target=qubit) + gates.Phase(
target=qubit, phi=angle) + gates.H(target=qubit)
O = ExpectationValue(U=U, H=H)
E = simulate(O, variables=variables, backend=simulator)
dO = grad(objective=O, variable='angle')
dE = simulate(dO, variables=variables)
assert (numpy.isclose(E, -numpy.sin(angle(variables)), atol=1.e-4))
assert (numpy.isclose(dE, -numpy.cos(angle(variables)), atol=1.e-4))
if not silent:
print("E =", E)
print("-sin(angle)=", -numpy.sin(angle(variables)))
print("dE =", dE)
print("-cos(angle)=", -numpy.cos(angle(variables)))
def test_gradient():
a = Variable(name='a')
variables = {a: 3.0}
b = a + 2 - 2
c = (b * 5) / 5
d = -(-c)
assert grad(d, a)(variables) == 1.0
def compile_gradient(self,
objective: Objective,
variables: typing.List[Variable],
gradient=None,
*args,
**kwargs) -> typing.Tuple[typing.Dict, typing.Dict]:
"""
convenience function to compile gradient objects and relavant types. For use by inheritors.
Parameters
----------
objective: Objective:
the objective whose gradient is to be calculated.
variables: list:
the variables to take gradients with resepct to.
gradient, optional:
special argument to change what structure is used to calculate the gradient, like numerical, or QNG.
Default: use regular, analytic gradients.
args
kwargs
Returns
-------
tuple:
both the uncompiled and compiled gradients of objective, w.r.t variables.
"""
if gradient is None:
dO = {
k: grad(objective=objective, variable=k, *args, **kwargs)
for k in variables
}
compiled_grad = {
k: self.compile_objective(objective=dO[k], *args, **kwargs)
for k in variables
}
elif isinstance(gradient, dict):
if all([isinstance(x, Objective) for x in gradient.values()]):
dO = gradient
compiled_grad = {
k: self.compile_objective(objective=dO[k], *args, **kwargs)
for k in variables
}
else:
dO = None
compiled = self.compile_objective(objective=objective)
compiled_grad = {
k: _NumGrad(objective=compiled, variable=k, **gradient)
for k in variables
}
else:
raise TequilaOptimizerException(
"unknown gradient instruction of type {} : {}".format(
type(gradient), gradient))
return dO, compiled_grad
def test_qubit_excitations():
H = paulis.Projector("1.0*|100>")
U1 = gates.X(0) + gates.QubitExcitation(
target=[0, 1], angle="a", assume_real=True)
U2 = gates.X(0) + gates.Trotterized(
generators=[U1.gates[1].make_generator()], angles=["a"], steps=1)
E1 = ExpectationValue(H=H, U=U1)
E2 = ExpectationValue(H=H, U=U2)
dE1 = grad(E1, "a")
dE2 = grad(E2, "a")
for a in numpy.random.uniform(-numpy.pi, numpy.pi, 5):
a = float(a)
variables = {"a": a}
wfn1 = simulate(U1, variables=variables)
wfn2 = simulate(U2, variables=variables)
F = numpy.abs(wfn1.inner(wfn2))**2
assert numpy.isclose(F, 1.0, 1.e-4)
eval1 = simulate(dE1, variables=variables)
eval2 = simulate(dE2, variables=variables)
assert numpy.isclose(eval1, eval2, 1.e-4)
H = paulis.Projector("1.0*|0110>")
U1 = gates.X([1, 2]) + gates.QubitExcitation(
target=[0, 1, 3, 2], angle="a", assume_real=True)
U2 = gates.X([1, 2]) + gates.Trotterized(
generators=[U1.gates[2].make_generator()], angles=["a"], steps=1)
E1 = ExpectationValue(H=H, U=U1)
E2 = ExpectationValue(H=H, U=U2)
dE1 = grad(E1, "a")
dE2 = grad(E2, "a")
for a in numpy.random.uniform(-numpy.pi, numpy.pi, 5):
a = float(a)
variables = {"a": a}
wfn1 = simulate(U1, variables=variables)
wfn2 = simulate(U2, variables=variables)
F = numpy.abs(wfn1.inner(wfn2))**2
assert numpy.isclose(F, 1.0, 1.e-4)
eval1 = simulate(dE1, variables=variables)
eval2 = simulate(dE2, variables=variables)
assert numpy.isclose(eval1, eval2, 1.e-4)
def test_really_awfull_thing(
simulator,
value1=(numpy.random.randint(0, 1000) / 1000.0 * (numpy.pi / 2.0)),
value2=(numpy.random.randint(0, 1000) / 1000.0 * (numpy.pi / 2.0))):
angle1 = Variable(name="angle1")
angle2 = Variable(name="angle2")
variables = {angle1: value1, angle2: value2}
prod = angle1 * angle2
qubit = 0
control = None
H = paulis.Y(qubit=qubit)
U = gates.Rx(target=qubit, control=control, angle=prod)
Up = gates.Rx(target=qubit, control=control, angle=prod + np.pi / 2)
Down = gates.Rx(target=qubit, control=control, angle=prod - np.pi / 2)
e1 = ExpectationValue(U=U, H=H)
en1 = simulate(e1, variables=variables, backend=simulator)
uen = simulate(0.5 * ExpectationValue(Up, H),
variables=variables,
backend=simulator)
den = simulate(-0.5 * ExpectationValue(Down, H),
variables=variables,
backend=simulator)
an1 = -np.sin(prod(variables=variables))
anval = prod(variables=variables)
an2 = angle2(variables=variables)
added = angle1 * e1
raised = added.wrap(np.sin)
dO = grad(raised, 'angle1')
dE = grad(e1, 'angle1')
dA = grad(added, 'angle1')
val = simulate(added, variables=variables, backend=simulator)
dave = simulate(dA, variables=variables, backend=simulator)
deval = simulate(dE, variables=variables, backend=simulator)
doval = simulate(dO, variables=variables, backend=simulator)
dtrue = np.cos(val) * dave
assert np.isclose(en1, an1, atol=1.e-4)
assert np.isclose(deval, an2 * (uen + den), atol=1.e-4)
assert np.isclose(doval, dtrue, atol=1.e-4)
def test_heterogeneous_gradient_r_div(simulator):
### the reason we don't test float power here is that it keeps coming up NAN, because the argument is too small
angle1 = Variable(name="angle1")
value = (numpy.random.randint(100, 1000) / 1000.0 * (numpy.pi / 2.0))
variables = {angle1: value}
qubit = 0
control = 1
H1 = paulis.Y(qubit=qubit)
U1 = gates.X(target=control) + gates.Rx(target=qubit, control=control, angle=angle1)
e1 = ExpectationValue(U=U1, H=H1)
added = Objective(args=[e1.args[0], angle1], transformation=np.true_divide)
val = simulate(added, variables=variables, backend=simulator)
en1 = simulate(e1, variables=variables, backend=simulator)
an1 = -np.sin(angle1(variables=variables))
anval = angle1(variables=variables)
dO = grad(added, 'angle1')
dE = grad(e1, 'angle1')
deval = simulate(dE, variables=variables, backend=simulator)
doval = simulate(dO, variables=variables, backend=simulator)
dtrue = deval / anval - en1 / (anval ** 2)
assert np.isclose(float(val), float(np.true_divide(en1, anval)))
assert np.isclose(en1, an1, atol=1.e-4)
assert np.isclose(doval, dtrue, atol=1.e-4)
def test_total_type_jumble(
simulator,
value1=(numpy.random.randint(0, 1000) / 1000.0 * (numpy.pi / 2.0)),
value2=(numpy.random.randint(0, 1000) / 1000.0 * (numpy.pi / 2.0))):
a = Variable('a')
b = Variable('b')
values = {a: value1, b: value2}
H1 = tq.paulis.X(0)
H2 = tq.paulis.Y(0)
U1 = tq.gates.Ry(angle=a, target=0)
U2 = tq.gates.Rx(angle=b, target=0)
e1 = ExpectationValue(U1, H1)
e2 = ExpectationValue(U2, H2)
stacked = tq.objective.vectorize([e1, e2])
stacked = stacked * a * e2
out = simulate(stacked, variables=values, backend=simulator)
v1 = out[0]
v2 = out[1]
appendage = a(values) * -np.sin(b(values))
an1 = np.sin(a(values)) * appendage
an2 = -np.sin(b(values)) * appendage
assert np.isclose(v1 + v2, an1 + an2)
# not gonna contract, lets make gradient do some real work
ga = grad(stacked, a)
gb = grad(stacked, b)
la = [tq.simulate(x, variables=values) for x in ga]
print(la)
lb = [tq.simulate(x, variables=values) for x in gb]
print(lb)
tota = np.sum(np.array(la))
totb = np.sum(np.array(lb))
gan1 = np.cos(a(values)) * appendage + (
np.sin(a(values)) * -np.sin(b(values))) - (np.sin(b(values)) *
-np.sin(b(values)))
gan2 = np.sin(a(values)) * a(values) * -np.cos(b(values)) + 2 * (
-np.cos(b(values)) * appendage)
assert np.isclose(tota + totb, gan1 + gan2)
def test_exotic_gradients(gradvar):
# a and b will fail for autograd not with jax
a = Variable('a')
b = Variable('b')
c = Variable('c')
d = Variable('d')
e = Variable('e')
f = Variable('f')
variables = {a: 2.0, b: 3.0, c: 4.0, d: 5.0, e: 6.0, f: 7.0}
t = c * a**b + b / c - Objective(
args=[c], transformation=np.cos) + f / (d * e) + a * Objective(
args=[d], transformation=np.exp) / (f + b) + Objective(
args=[e], transformation=np.tanh) + Objective(
args=[f], transformation=np.sinc)
g = grad(t, gradvar)
if gradvar == 'a':
assert np.isclose(
g(variables),
c(variables) * b(variables) * (a(variables)**(b(variables) - 1.)) +
np.exp(d(variables)) / (f(variables) + b(variables)))
if gradvar == 'b':
assert np.isclose(
g(variables),
(c(variables) * a(variables)**b(variables)) * np.log(a(variables))
+ 1. / c(variables) - a(variables) * np.exp(d(variables)) /
(f(variables) + b(variables))**2.0)
if gradvar == 'c':
assert np.isclose(
g(variables),
a(variables)**b(variables) - b(variables) / c(variables)**2. +
np.sin(c(variables)))
if gradvar == 'd':
assert np.isclose(
g(variables),
-f(variables) / (np.square(d(variables)) * e(variables)) +
a(variables) * np.exp(d(variables)) /
(f(variables) + b(variables)))
if gradvar == 'e':
assert np.isclose(
g(variables), 2. / (1. + np.cosh(2 * e(variables))) -
f(variables) / (d(variables) * e(variables)**2.))
if gradvar == 'f':
assert np.isclose(
g(variables), 1. / (d(variables) * e(variables)) -
a(variables) * np.exp(d(variables)) /
(f(variables) + b(variables))**2. +
np.cos(np.pi * f(variables)) / f(variables) -
np.sin(np.pi * f(variables)) / (np.pi * f(variables)**2.))
def test_gradient_deep_controlled_Y(simulator, power, controls):
if controls > 2 and simulator == "qiskit":
# does not work yet
return
qubit = 0
control = [i for i in range(1, controls + 1)]
angle = Variable(name="angle")
U = gates.X(target=control) + gates.Y(target=qubit, power=angle, control=control)
angle = Variable(name="angle")
variables = {angle: power}
H = paulis.X(qubit=qubit)
O = ExpectationValue(U=U, H=H)
E = simulate(O, variables=variables, backend=simulator)
dO = grad(objective=O, variable=angle)
dE = simulate(dO, variables=variables, backend=simulator)
assert (numpy.isclose(E, numpy.sin(angle(variables) * (numpy.pi)), atol=1.e-4))
assert (numpy.isclose(dE, numpy.pi * numpy.cos(angle(variables) * (numpy.pi)), atol=1.e-4))
def test_gradient_X(simulator, power, controlled):
qubit = 0
control = 1
angle = Variable(name="angle")
if controlled:
U = gates.X(target=control) + gates.X(target=qubit, power=angle, control=control)
else:
U = gates.X(target=qubit, power=angle)
angle = Variable(name="angle")
variables = {angle: power}
H = paulis.Y(qubit=qubit)
O = ExpectationValue(U=U, H=H)
E = simulate(O, variables=variables, backend=simulator)
dO = grad(objective=O, variable=angle)
dE = simulate(dO, variables=variables, backend=simulator)
assert (numpy.isclose(E, -numpy.sin(angle(variables) * (numpy.pi)), atol=1.e-4))
assert (numpy.isclose(dE, -numpy.pi * numpy.cos(angle(variables) * (numpy.pi)), atol=1.e-4))
def compile_hessian(self,
variables: typing.List[Variable],
grad_obj: typing.Dict[Variable, Objective],
comp_grad_obj: typing.Dict[Variable, Objective],
hessian: dict = None,
*args,
**kwargs) -> tuple:
dO = grad_obj
cdO = comp_grad_obj
if hessian is None:
if dO is None:
raise TequilaOptimizerException("Can not combine analytical Hessian with numerical Gradient\n"
"hessian instruction was: {}".format(hessian))
compiled_hessian = {}
ddO = {}
for k in variables:
dOk = dO[k]
for l in variables:
ddO[(k, l)] = grad(objective=dOk, variable=l)
compiled_hessian[(k, l)] = self.compile_objective(ddO[(k, l)])
ddO[(l, k)] = ddO[(k, l)]
compiled_hessian[(l, k)] = compiled_hessian[(k, l)]
elif isinstance(hessian, dict):
if all([isinstance(x, Objective) for x in hessian.values()]):
ddO = hessian
compiled_hessian = {k: self.compile_objective(objective=ddO[k], *args, **kwargs) for k in
hessian.keys()}
else:
ddO = None
compiled_hessian = {}
for k in variables:
for l in variables:
compiled_hessian[(k, l)] = _NumGrad(objective=cdO[k], variable=l, **hessian)
compiled_hessian[(l, k)] = _NumGrad(objective=cdO[l], variable=k, **hessian)
else:
raise TequilaOptimizerException("unknown hessian instruction: {}".format(hessian))
return ddO, compiled_hessian
def test_gradient_UY_HX_wfnsim(simulator,
angle_value,
controlled,
silent=True):
# same as before just with wavefunction simulation
# case X Y
# U = cos(angle/2) + sin(-angle/2)*i*Y
# <0|Ud H U |0> = cos^2(angle/2)*<0|X|0>
# + sin^2(-angle/2) <0|YXY|0>
# + cos(angle/2)*sin(angle/2)*i<0|XY|0>
# + sin(-angle/2)*cos(angle/2)*(-i) <0|YX|0>
# = cos^2*0 + sin^2*0 + cos*sin*i(<0|[XY,YX]|0>)
# = 0.5*sin(-angle)*i <0|[XY,YX]|0> = -0.5*sin(angle)*i * 2 i <0|Z|0>
# = sin(angle)
angle = Variable(name="angle")
variables = {angle: angle_value}
qubit = 0
H = paulis.X(qubit=qubit)
if controlled:
control = 1
U = gates.X(target=control) + gates.Ry(
target=qubit, control=control, angle=angle)
else:
U = gates.Ry(target=qubit, angle=angle)
O = ExpectationValue(U=U, H=H)
E = simulate(O, variables=variables, backend=simulator)
dO = grad(objective=O, variable='angle')
dE = simulate(dO, variables=variables, backend=simulator)
E = numpy.float(E) # for isclose
dE = numpy.float(dE) # for isclose
assert (numpy.isclose(E, numpy.sin(angle(variables)), atol=0.0001))
assert (numpy.isclose(dE, numpy.cos(angle(variables)), atol=0.0001))
if not silent:
print("E =", E)
print("sin(angle)=", numpy.sin(angle(variables)))
print("dE =", dE)
print("cos(angle)=", numpy.cos(angle(variables)))
def compile_gradient(self, objective: Objective,
variables: typing.List[Variable],
gradient=None,
*args, **kwargs) -> typing.Tuple[
typing.Dict, typing.Dict]:
if gradient is None:
dO = {k: grad(objective=objective, variable=k, *args, **kwargs) for k in variables}
compiled_grad = {k: self.compile_objective(objective=dO[k], *args, **kwargs) for k in variables}
elif isinstance(gradient, dict):
if all([isinstance(x, Objective) for x in gradient.values()]):
dO = gradient
compiled_grad = {k: self.compile_objective(objective=dO[k], *args, **kwargs) for k in variables}
else:
dO = None
compiled = self.compile_objective(objective=objective)
compiled_grad = {k: _NumGrad(objective=compiled, variable=k, **gradient) for k in variables}
else:
raise TequilaOptimizerException(
"unknown gradient instruction of type {} : {}".format(type(gradient), gradient))
return dO, compiled_grad
def __call__(self,
objective: Objective,
maxiter,
lr: float = .01,
method: str = 'sgd',
qng: bool = False,
stop_count: int = None,
initial_values: typing.Dict[Variable, numbers.Real] = None,
variables: typing.List[Variable] = None,
samples: int = None,
backend: str = None,
noise: NoiseModel = None,
reset_history: bool = True,
*args,
**kwargs) -> GDReturnType:
"""
Optimizes with a variation of gradient descent and gives back the optimized angles
Get the optimized energies over the history
:param objective: The tequila Objective to minimize
:param maxiter: how many iterations to run, at maximum.
:param method: what method to optimize via.
:param qng: whether or not to use the QNG to calculate gradients.
:param stop_count: how many steps after which to abort if no improvement occurs.
:param initial_values: initial values for the objective
:param variables: which variables to optimize over. Default None: all the variables of the objective.
:param samples: the number of samples to use. Default None: Wavefunction simulation used instead.
:param backend: which simulation backend to use. Default None: let Tequila Pick!
:param noise: the NoiseModel to apply to sampling. Default None. Affects chosen simulator.
:param reset_history: reset the history before optimization starts (has no effect if self.save_history is False)
:return: tuple of optimized energy ,optimized angles and scipy output
"""
if self.save_history and reset_history:
self.reset_history()
active_angles = {}
for v in variables:
active_angles[v] = initial_values[v]
passive_angles = {}
for k, v in initial_values.items():
if k not in active_angles.keys():
passive_angles[k] = v
# Transform the initial value directory into (ordered) arrays
comp = compile(objective=objective,
variables=initial_values,
backend=backend,
noise=noise,
samples=samples)
if not qng:
g_list = []
for k in active_angles.keys():
g = grad(objective, k)
g_comp = compile(objective=g,
variables=initial_values,
backend=backend,
noise=noise,
samples=samples)
g_list.append(g_comp)
gradients = CallableVector(g_list)
else:
if method.lower() == 'adagrad':
print(
'Warning! you have chosen to use QNG with adagrad ; convergence is not likely.'
.format(method))
gradients = QNGVector(
get_qng_combos(objective=objective,
initial_values=initial_values,
backend=backend,
noise=noise,
samples=samples))
if not self.silent:
print("backend: {}".format(comp.backend))
print("samples: {}".format(samples))
print("{} active variables".format(len(active_angles)))
print("qng: {}".format(str(qng)))
### prefactor. Early stopping, initialization, etc. handled here
if maxiter is None:
maxiter = self.maxiter
if stop_count == None:
stop_count = maxiter
### the actual algorithm acts here:
f = self.method_dict[method.lower()]
v = initial_values
vec_len = len(active_angles)
best = None
best_angles = None
first = numpy.zeros(vec_len)
second = numpy.zeros(vec_len)
moments = [first, second]
all_moments = [moments]
tally = 0
for step in range(maxiter):
e = comp(v, samples=samples)
self.history.energies.append(e)
self.history.angles.append(v)
### saving best performance and counting the stop tally.
if step == 0:
best = e
best_angles = v
tally = 0
else:
if e < best:
best = e
best_angles = v
tally = 0
else:
tally += 1
if not self.silent:
string = "Iteration: {} , Energy: {}, angles: {}".format(
str(step), str(e), v)
print(string)
### check if its time to stop!
if tally == stop_count:
if not self.silent:
print(
'no improvement after {} epochs. Stopping optimization.'
.format(str(stop_count)))
break
new, moments, grads = f(lr=lr,
step=step,
gradients=gradients,
v=v,
moments=moments,
active_angles=active_angles,
samples=samples,
**kwargs)
save_grad = {}
if passive_angles != None:
v = {**new, **passive_angles}
else:
v = new
for i, k in enumerate(active_angles.keys()):
save_grad[k] = grads[i]
self.history.gradients.append(save_grad)
all_moments.append(moments)
E_final, angles_final = best, best_angles
angles_final = {**angles_final, **passive_angles}
return GDReturnType(energy=E_final,
angles=format_variable_dictionary(angles_final),
history=self.history,
moments=all_moments)
def __call__(self, objective: Objective,
initial_values: typing.Dict[Variable, numbers.Real],
variables: typing.List[Variable],
gradient: typing.Dict[Variable, Objective] = None,
qng: bool = False,
hessian: typing.Dict[typing.Tuple[Variable, Variable], Objective] = None,
samples: int = None,
backend: str = None,
backend_options: dict = None,
noise: NoiseModel = None,
reset_history: bool = True,
*args,
**kwargs) -> SciPyReturnType:
"""
Optimizes with scipy and gives back the optimized angles
Get the optimized energies over the history
:param objective: The tequila Objective to minimize
:param initial_valuesxx: initial values for the objective
:param return_scipy_output: chose if the full scipy output shall be returned
:param reset_history: reset the history before optimization starts (has no effect if self.save_history is False)
:return: tuple of optimized energy ,optimized angles and scipy output
"""
infostring = "Starting {method} optimization\n".format(method=self.method)
infostring += "Objective: {} expectationvalues\n".format(objective.count_expectationvalues())
if self.save_history and reset_history:
self.reset_history()
active_angles = {}
for v in variables:
active_angles[v] = initial_values[v]
passive_angles = {}
for k, v in initial_values.items():
if k not in active_angles.keys():
passive_angles[k] = v
# Transform the initial value directory into (ordered) arrays
param_keys, param_values = zip(*active_angles.items())
param_values = numpy.array(param_values)
bounds = None
if self.method_bounds is not None:
bounds = {k: None for k in active_angles}
for k, v in self.method_bounds.items():
if k in bounds:
bounds[k] = v
infostring += "bounds : {}\n".format(self.method_bounds)
names, bounds = zip(*bounds.items())
assert (names == param_keys) # make sure the bounds are not shuffled
# do the compilation here to avoid costly recompilation during the optimization
compiled_objective = compile(objective=objective, variables=initial_values, backend=backend, noise=noise,
samples=samples, *args, **kwargs)
E = _EvalContainer(objective=compiled_objective,
param_keys=param_keys,
samples=samples,
passive_angles=passive_angles,
save_history=self.save_history,
backend_options = backend_options,
silent=self.silent)
# compile gradients
if self.method in self.gradient_based_methods + self.hessian_based_methods and not isinstance(gradient, str):
compiled_grad_objectives = dict()
if gradient is None:
gradient = {assign_variable(k): grad(objective=objective, variable=k) for k in active_angles.keys()}
else:
gradient = {assign_variable(k): v for k, v in gradient.items()}
grad_exval = []
for k in active_angles.keys():
if k not in gradient:
raise Exception("No gradient for variable {}".format(k))
grad_exval.append(gradient[k].count_expectationvalues())
compiled_grad_objectives[k] = compile(objective=gradient[k], variables=initial_values,
samples=samples, noise=noise, backend=backend, *args, **kwargs)
if qng:
combos = get_qng_combos(objective, samples=samples, backend=backend,
noise=noise, initial_values=initial_values)
dE = _QngContainer(combos=combos,
param_keys=param_keys,
samples=samples,
passive_angles=passive_angles,
save_history=self.save_history,
silent=self.silent,
backend_options=backend_options)
else:
dE = _GradContainer(objective=compiled_grad_objectives,
param_keys=param_keys,
samples=samples,
passive_angles=passive_angles,
save_history=self.save_history,
silent=self.silent,
backend_options=backend_options)
infostring += "Gradients: {} expectationvalues (min={}, max={})\n".format(sum(grad_exval),
min(grad_exval),
max(grad_exval))
else:
# use numerical gradient
dE = gradient
infostring += "Gradients: {}\n".format(gradient)
# compile hessian
if self.method in self.hessian_based_methods and not isinstance(hessian, str):
if isinstance(gradient, str):
raise TequilaScipyException("Can not use numerical gradients for Hessian based methods")
if qng is True:
raise TequilaScipyException('Quantum Natural Hessian not yet well-defined, sorry!')
compiled_hess_objectives = dict()
hess_exval = []
for i, k in enumerate(active_angles.keys()):
for j, l in enumerate(active_angles.keys()):
if j > i: continue
hess = grad(gradient[k], l)
compiled_hess = compile(objective=hess, variables=initial_values, samples=samples,
noise=noise,
backend=backend, *args, **kwargs)
compiled_hess_objectives[(k, l)] = compiled_hess
compiled_hess_objectives[(l, k)] = compiled_hess
hess_exval.append(compiled_hess.count_expectationvalues())
ddE = _HessContainer(objective=compiled_hess_objectives,
param_keys=param_keys,
samples=samples,
passive_angles=passive_angles,
save_history=self.save_history,
silent=self.silent)
infostring += "Hessian: {} expectationvalues (min={}, max={})\n".format(sum(hess_exval), min(hess_exval),
max(hess_exval))
else:
infostring += "Hessian: {}\n".format(hessian)
if self.method != "TRUST-CONSTR" and hessian is not None:
raise TequilaScipyException("numerical hessians only for trust-constr method")
ddE = hessian
if not self.silent:
print("ObjectiveType is {}".format(type(compiled_objective)))
print(infostring)
print("backend: {}".format(compiled_objective.backend))
print("samples: {}".format(samples))
print("{} active variables".format(len(active_angles)))
# get the number of real scipy iterations for better histories
real_iterations = []
Es = []
callback = lambda x, *args: real_iterations.append(len(E.history) - 1)
res = scipy.optimize.minimize(E, x0=param_values, jac=dE, hess=ddE,
args=(Es,),
method=self.method, tol=self.tol,
bounds=bounds,
constraints=self.method_constraints,
options=self.method_options,
callback=callback)
# failsafe since callback is not implemented everywhere
if len(real_iterations) == 0:
real_iterations = range(len(E.history))
else:
real_iterations = [0] + real_iterations
if self.save_history:
self.history.energies = [E.history[i] for i in real_iterations]
self.history.energy_evaluations = E.history
self.history.angles = [E.history_angles[i] for i in real_iterations]
self.history.angles_evaluations = E.history_angles
if dE is not None and not isinstance(dE, str):
# can currently only save gradients if explicitly evaluated
# and will fail for hessian based approaches
# need better callback functions
try:
if self.method not in self.hessian_based_methods:
self.history.gradients = [dE.history[i] for i in real_iterations]
except:
print("WARNING: History could not assign the stored gradients")
self.history.gradients_evaluations = dE.history
if ddE is not None and not isinstance(ddE, str):
# hessians are not evaluated in the same frequencies as energies
# therefore we can not store the "real" iterations currently
self.history.hessians_evaluations = ddE.history
E_final = res.fun
angles_final = dict((param_keys[i], res.x[i]) for i in range(len(param_keys)))
angles_final = {**angles_final, **passive_angles}
return SciPyReturnType(energy=E_final, angles=format_variable_dictionary(angles_final), history=self.history,
scipy_output=res)
def compile_hessian(self,
variables: typing.List[Variable],
grad_obj: typing.Dict[Variable, Objective],
comp_grad_obj: typing.Dict[Variable, Objective],
hessian: dict = None,
*args,
**kwargs) -> tuple:
"""
convenience function to compile hessians for optimizers which require it.
Parameters
----------
variables:
the variables of the hessian.
grad_obj:
the gradient object, to be differentiated once more
comp_grad_obj:
the compiled gradient object, used for further compilation of the hessian.
hessian: optional:
extra information to modulate compilation of the hessian.
args
kwargs
Returns
-------
tuple:
uncompiled and compiled hessian objects, in that order
"""
dO = grad_obj
cdO = comp_grad_obj
if hessian is None:
if dO is None:
raise TequilaOptimizerException(
"Can not combine analytical Hessian with numerical Gradient\n"
"hessian instruction was: {}".format(hessian))
compiled_hessian = {}
ddO = {}
for k in variables:
dOk = dO[k]
for l in variables:
ddO[(k, l)] = grad(objective=dOk, variable=l)
compiled_hessian[(k, l)] = self.compile_objective(ddO[(k,
l)])
ddO[(l, k)] = ddO[(k, l)]
compiled_hessian[(l, k)] = compiled_hessian[(k, l)]
elif isinstance(hessian, dict):
if all([isinstance(x, Objective) for x in hessian.values()]):
ddO = hessian
compiled_hessian = {
k: self.compile_objective(objective=ddO[k],
*args,
**kwargs)
for k in hessian.keys()
}
else:
ddO = None
compiled_hessian = {}
for k in variables:
for l in variables:
compiled_hessian[(k, l)] = _NumGrad(objective=cdO[k],
variable=l,
**hessian)
compiled_hessian[(l, k)] = _NumGrad(objective=cdO[l],
variable=k,
**hessian)
else:
raise TequilaOptimizerException(
"unknown hessian instruction: {}".format(hessian))
return ddO, compiled_hessian