def setUp(self):
self.sgd_opt = GradientDescentOptimizer(stepsize)
self.mom_opt = MomentumOptimizer(stepsize, momentum=gamma)
self.nesmom_opt = NesterovMomentumOptimizer(stepsize, momentum=gamma)
self.adag_opt = AdagradOptimizer(stepsize)
self.rms_opt = RMSPropOptimizer(stepsize, decay=gamma)
self.adam_opt = AdamOptimizer(stepsize, beta1=gamma, beta2=delta)
self.fnames = ['test_function_1', 'test_function_2', 'test_function_3']
self.univariate_funcs = [
np.sin, lambda x: np.exp(x / 10.), lambda x: x**2
]
self.grad_uni_fns = [
np.cos, lambda x: np.exp(x / 10.) / 10., lambda x: 2 * x
]
self.multivariate_funcs = [
lambda x: np.sin(x[0]) + np.cos(x[1]),
lambda x: np.exp(x[0] / 3) * np.tanh(x[1]),
lambda x: np.sum([x_**2 for x_ in x])
]
self.grad_multi_funcs = [
lambda x: np.array([np.cos(x[0]), -np.sin(x[1])]),
lambda x: np.array([
np.exp(x[0] / 3) / 3 * np.tanh(x[1]),
np.exp(x[0] / 3) * (1 - np.tanh(x[1])**2)
]), lambda x: np.array([2 * x_ for x_ in x])
]
self.mvar_mdim_funcs = [
lambda x: np.sin(x[0, 0]) + np.cos(x[1, 0]) - np.sin(x[0, 1]) + x[
1, 1], lambda x: np.exp(x[0, 0] / 3) * np.tanh(x[0, 1]),
lambda x: np.sum([x_[0]**2 for x_ in x])
]
self.grad_mvar_mdim_funcs = [
lambda x: np.array([[np.cos(x[0, 0]), -np.cos(x[0, 1])],
[-np.sin(x[1, 0]), 1.]]),
lambda x: np.array([[
np.exp(x[0, 0] / 3) / 3 * np.tanh(x[0, 1]),
np.exp(x[0, 0] / 3) * (1 - np.tanh(x[0, 1])**2)
], [0., 0.]]), lambda x: np.array([[2 * x_[0], 0.] for x_ in x])
]
self.class_fun = class_fun
self.quant_fun = quant_fun
self.hybrid_fun = hybrid_fun
self.hybrid_fun_nested = hybrid_fun_nested
self.hybrid_fun_flat = hybrid_fun_flat
self.hybrid_fun_mdarr = hybrid_fun_mdarr
self.hybrid_fun_mdlist = hybrid_fun_mdlist
self.mixed_list = [(0.2, 0.3), np.array([0.4, 0.2, 0.4]), 0.1]
self.mixed_tuple = (np.array([0.2, 0.3]), [0.4, 0.2, 0.4], 0.1)
self.nested_list = [[[0.2], 0.3], [0.1, [0.4]], -0.1]
self.flat_list = [0.2, 0.3, 0.1, 0.4, -0.1]
self.multid_array = np.array([[0.1, 0.2], [-0.1, -0.4]])
self.multid_list = [[0.1, 0.2], [-0.1, -0.4]]
class A:
sgd_opt = GradientDescentOptimizer(stepsize)
mom_opt = MomentumOptimizer(stepsize, momentum=gamma)
nesmom_opt = NesterovMomentumOptimizer(stepsize, momentum=gamma)
adag_opt = AdagradOptimizer(stepsize)
rms_opt = RMSPropOptimizer(stepsize, decay=gamma)
adam_opt = AdamOptimizer(stepsize, beta1=gamma, beta2=delta)
def test_momentum_optimizer_univar(self, x_start, tol):
"""Tests that momentum optimizer takes one and two steps correctly
for univariate functions."""
stepsize, gamma = 0.1, 0.5
mom_opt = MomentumOptimizer(stepsize, momentum=gamma)
univariate_funcs = [np.sin, lambda x: np.exp(x / 10.0), lambda x: x**2]
grad_uni_fns = [
lambda x: (np.cos(x), ),
lambda x: (np.exp(x / 10.0) / 10.0, ),
lambda x: (2 * x, ),
]
for gradf, f in zip(grad_uni_fns, univariate_funcs):
mom_opt.reset()
x_onestep = mom_opt.step(f, x_start)
x_onestep_target = x_start - gradf(x_start)[0] * stepsize
assert np.allclose(x_onestep, x_onestep_target, atol=tol)
x_twosteps = mom_opt.step(f, x_onestep)
momentum_term = gamma * gradf(x_start)[0]
x_twosteps_target = x_onestep - (gradf(x_onestep)[0] +
momentum_term) * stepsize
assert np.allclose(x_twosteps, x_twosteps_target, atol=tol)
def test_apply_grad(self, grad, args, tol):
"""
Test that the gradient can be applied correctly to a set of parameters
and that momentum accumulation works correctly.
"""
stepsize, gamma = 0.1, 0.5
sgd_opt = MomentumOptimizer(stepsize, momentum=gamma)
grad, args = np.array(grad), np.array(args, requires_grad=True)
a1 = stepsize * grad
expected = args - a1
res = sgd_opt.apply_grad(grad, args)
assert np.allclose(res, expected)
# Simulate a new step
grad = grad + args
args = expected
a2 = gamma * a1 + stepsize * grad
expected = args - a2
res = sgd_opt.apply_grad(grad, args)
assert np.allclose(res, expected, atol=tol)
def opt(opt_name):
stepsize, gamma, delta = 0.1, 0.5, 0.8
if opt_name == "gd":
return GradientDescentOptimizer(stepsize)
if opt_name == "nest":
return NesterovMomentumOptimizer(stepsize, momentum=gamma)
if opt_name == "moment":
return MomentumOptimizer(stepsize, momentum=gamma)
if opt_name == "ada":
return AdagradOptimizer(stepsize)
if opt_name == "rms":
return RMSPropOptimizer(stepsize, decay=gamma)
if opt_name == "adam":
return AdamOptimizer(stepsize, beta1=gamma, beta2=delta)
def test_momentum_optimizer_multivar(self, tol):
"""Tests that momentum optimizer takes one and two steps correctly
for multivariate functions."""
stepsize, gamma = 0.1, 0.5
mom_opt = MomentumOptimizer(stepsize, momentum=gamma)
multivariate_funcs = [
lambda x: np.sin(x[0]) + np.cos(x[1]),
lambda x: np.exp(x[0] / 3) * np.tanh(x[1]),
lambda x: np.sum([x_**2 for x_ in x]),
]
grad_multi_funcs = [
lambda x: (np.array([np.cos(x[0]), -np.sin(x[1])]), ),
lambda x: (np.array([
np.exp(x[0] / 3) / 3 * np.tanh(x[1]),
np.exp(x[0] / 3) * (1 - np.tanh(x[1])**2),
]), ),
lambda x: (np.array([2 * x_ for x_ in x]), ),
]
x_vals = np.linspace(-10, 10, 16, endpoint=False)
for gradf, f in zip(grad_multi_funcs, multivariate_funcs):
for jdx in range(len(x_vals[:-1])):
mom_opt.reset()
x_vec = x_vals[jdx:jdx + 2]
x_onestep = mom_opt.step(f, x_vec)
x_onestep_target = x_vec - gradf(x_vec)[0] * stepsize
assert np.allclose(x_onestep, x_onestep_target, atol=tol)
x_twosteps = mom_opt.step(f, x_onestep)
momentum_term = gamma * gradf(x_vec)[0]
x_twosteps_target = x_onestep - (gradf(x_onestep)[0] +
momentum_term) * stepsize
assert np.allclose(x_twosteps, x_twosteps_target, atol=tol)
eta2 = 0.1
opt.update_stepsize(eta2)
assert opt._stepsize == eta2
def reset(opt):
if getattr(opt, "reset", None):
opt.reset()
@pytest.mark.parametrize(
"opt, opt_name",
[
(GradientDescentOptimizer(stepsize), "gd"),
(MomentumOptimizer(stepsize, momentum=gamma), "moment"),
(NesterovMomentumOptimizer(stepsize, momentum=gamma), "nest"),
(AdagradOptimizer(stepsize), "ada"),
(RMSPropOptimizer(stepsize, decay=gamma), "rms"),
(AdamOptimizer(stepsize, beta1=gamma, beta2=delta), "adam"),
(RotosolveOptimizer(), "roto"),
],
)
class TestOverOpts:
"""Tests keywords, multiple arguements, and non-training arguments in relevent optimizers"""
def test_kwargs(self, mocker, opt, opt_name, tol):
"""Test that the keywords get passed and alter the function"""
class func_wrapper:
@staticmethod
def func(x, c=1.0):
return (x - c)**2
class BasicTest(BaseTest):
"""Basic optimizer tests.
"""
def setUp(self):
self.sgd_opt = GradientDescentOptimizer(stepsize)
self.mom_opt = MomentumOptimizer(stepsize, momentum=gamma)
self.nesmom_opt = NesterovMomentumOptimizer(stepsize, momentum=gamma)
self.adag_opt = AdagradOptimizer(stepsize)
self.rms_opt = RMSPropOptimizer(stepsize, decay=gamma)
self.adam_opt = AdamOptimizer(stepsize, beta1=gamma, beta2=delta)
self.fnames = ['test_function_1', 'test_function_2', 'test_function_3']
self.univariate_funcs = [
np.sin, lambda x: np.exp(x / 10.), lambda x: x**2
]
self.grad_uni_fns = [
np.cos, lambda x: np.exp(x / 10.) / 10., lambda x: 2 * x
]
self.multivariate_funcs = [
lambda x: np.sin(x[0]) + np.cos(x[1]),
lambda x: np.exp(x[0] / 3) * np.tanh(x[1]),
lambda x: np.sum([x_**2 for x_ in x])
]
self.grad_multi_funcs = [
lambda x: np.array([np.cos(x[0]), -np.sin(x[1])]),
lambda x: np.array([
np.exp(x[0] / 3) / 3 * np.tanh(x[1]),
np.exp(x[0] / 3) * (1 - np.tanh(x[1])**2)
]), lambda x: np.array([2 * x_ for x_ in x])
]
self.mvar_mdim_funcs = [
lambda x: np.sin(x[0, 0]) + np.cos(x[1, 0]) - np.sin(x[0, 1]) + x[
1, 1], lambda x: np.exp(x[0, 0] / 3) * np.tanh(x[0, 1]),
lambda x: np.sum([x_[0]**2 for x_ in x])
]
self.grad_mvar_mdim_funcs = [
lambda x: np.array([[np.cos(x[0, 0]), -np.cos(x[0, 1])],
[-np.sin(x[1, 0]), 1.]]),
lambda x: np.array([[
np.exp(x[0, 0] / 3) / 3 * np.tanh(x[0, 1]),
np.exp(x[0, 0] / 3) * (1 - np.tanh(x[0, 1])**2)
], [0., 0.]]), lambda x: np.array([[2 * x_[0], 0.] for x_ in x])
]
self.class_fun = class_fun
self.quant_fun = quant_fun
self.hybrid_fun = hybrid_fun
self.hybrid_fun_nested = hybrid_fun_nested
self.hybrid_fun_flat = hybrid_fun_flat
self.hybrid_fun_mdarr = hybrid_fun_mdarr
self.hybrid_fun_mdlist = hybrid_fun_mdlist
self.mixed_list = [(0.2, 0.3), np.array([0.4, 0.2, 0.4]), 0.1]
self.mixed_tuple = (np.array([0.2, 0.3]), [0.4, 0.2, 0.4], 0.1)
self.nested_list = [[[0.2], 0.3], [0.1, [0.4]], -0.1]
self.flat_list = [0.2, 0.3, 0.1, 0.4, -0.1]
self.multid_array = np.array([[0.1, 0.2], [-0.1, -0.4]])
self.multid_list = [[0.1, 0.2], [-0.1, -0.4]]
def test_mixed_inputs_for_hybrid_optimization(self):
"""Tests that gradient descent optimizer treats parameters of mixed types the same
for hybrid optimization tasks."""
self.logTestName()
hybrid_list = self.sgd_opt.step(self.hybrid_fun, self.mixed_list)
hybrid_tuple = self.sgd_opt.step(self.hybrid_fun, self.mixed_tuple)
self.assertAllAlmostEqual(hybrid_list[0],
hybrid_tuple[0],
delta=self.tol)
self.assertAllAlmostEqual(hybrid_list[1],
hybrid_tuple[1],
delta=self.tol)
self.assertAllAlmostEqual(hybrid_list[2],
hybrid_tuple[2],
delta=self.tol)
def test_mixed_inputs_for_classical_optimization(self):
"""Tests that gradient descent optimizer treats parameters of mixed types the same
for purely classical optimization tasks."""
self.logTestName()
class_list = self.sgd_opt.step(self.class_fun, self.mixed_list)
class_tuple = self.sgd_opt.step(self.class_fun, self.mixed_tuple)
self.assertAllAlmostEqual(class_list[0],
class_tuple[0],
delta=self.tol)
self.assertAllAlmostEqual(class_list[1],
class_tuple[1],
delta=self.tol)
self.assertAllAlmostEqual(class_list[2],
class_tuple[2],
delta=self.tol)
def test_mixed_inputs_for_quantum_optimization(self):
"""Tests that gradient descent optimizer treats parameters of mixed types the same
for purely quantum optimization tasks."""
self.logTestName()
quant_list = self.sgd_opt.step(self.quant_fun, self.mixed_list)
quant_tuple = self.sgd_opt.step(self.quant_fun, self.mixed_tuple)
self.assertAllAlmostEqual(quant_list[0],
quant_tuple[0],
delta=self.tol)
self.assertAllAlmostEqual(quant_list[1],
quant_tuple[1],
delta=self.tol)
self.assertAllAlmostEqual(quant_list[2],
quant_tuple[2],
delta=self.tol)
def test_nested_and_flat_returns_same_update(self):
"""Tests that gradient descent optimizer has the same output for
nested and flat lists."""
self.logTestName()
nested = self.sgd_opt.step(self.hybrid_fun_nested, self.nested_list)
flat = self.sgd_opt.step(self.hybrid_fun_flat, self.flat_list)
self.assertAllAlmostEqual(flat, list(_flatten(nested)), delta=self.tol)
def test_array_and_list_return_same_update(self):
"""Tests that gradient descent optimizer has the same output for
lists and arrays."""
self.logTestName()
array = self.sgd_opt.step(self.hybrid_fun_mdarr, self.multid_array)
list = self.sgd_opt.step(self.hybrid_fun_mdlist, self.multid_list)
self.assertAllAlmostEqual(array, list, delta=self.tol)
def test_gradient_descent_optimizer_univar(self):
"""Tests that basic stochastic gradient descent takes gradient-descent steps correctly
for uni-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_uni_fns, self.univariate_funcs,
self.fnames):
with self.subTest(i=name):
for x_start in x_vals:
x_new = self.sgd_opt.step(f, x_start)
x_correct = x_start - gradf(x_start) * stepsize
self.assertAlmostEqual(x_new, x_correct, delta=self.tol)
def test_gradient_descent_optimizer_multivar(self):
"""Tests that basic stochastic gradient descent takes gradient-descent steps correctly
for multi-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_multi_funcs,
self.multivariate_funcs, self.fnames):
with self.subTest(i=name):
for jdx in range(len(x_vals[:-1])):
x_vec = x_vals[jdx:jdx + 2]
x_new = self.sgd_opt.step(f, x_vec)
x_correct = x_vec - gradf(x_vec) * stepsize
self.assertAllAlmostEqual(x_new, x_correct, delta=self.tol)
def test_gradient_descent_optimizer_multivar_multidim(self):
"""Tests that basic stochastic gradient descent takes gradient-descent steps correctly
for multi-variate functions and with higher dimensional inputs."""
self.logTestName()
for gradf, f, name in zip(self.grad_mvar_mdim_funcs,
self.mvar_mdim_funcs, self.fnames):
with self.subTest(i=name):
for jdx in range(len(x_vals[:-3])):
x_vec = x_vals[jdx:jdx + 4]
x_vec_multidim = np.reshape(x_vec, (2, 2))
x_new = self.sgd_opt.step(f, x_vec_multidim)
x_correct = x_vec_multidim - gradf(
x_vec_multidim) * stepsize
x_new_flat = x_new.flatten()
x_correct_flat = x_correct.flatten()
self.assertAllAlmostEqual(x_new_flat,
x_correct_flat,
delta=self.tol)
def test_gradient_descent_optimizer_usergrad(self):
"""Tests that basic stochastic gradient descent takes gradient-descent steps correctly
using user-provided gradients."""
self.logTestName()
for gradf, f, name in zip(self.grad_uni_fns[::-1],
self.univariate_funcs, self.fnames):
with self.subTest(i=name):
for x_start in x_vals:
x_new = self.sgd_opt.step(f, x_start, grad_fn=gradf)
x_correct = x_start - gradf(x_start) * stepsize
self.assertAlmostEqual(x_new, x_correct, delta=self.tol)
def test_momentum_optimizer_univar(self):
"""Tests that momentum optimizer takes one and two steps correctly
for uni-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_uni_fns, self.univariate_funcs,
self.fnames):
with self.subTest(i=name):
for x_start in x_vals:
self.mom_opt.reset()
x_onestep = self.mom_opt.step(f, x_start)
x_onestep_target = x_start - gradf(x_start) * stepsize
self.assertAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.mom_opt.step(f, x_onestep)
momentum_term = gamma * gradf(x_start)
x_twosteps_target = x_onestep - (gradf(x_onestep) +
momentum_term) * stepsize
self.assertAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_momentum_optimizer_multivar(self):
"""Tests that momentum optimizer takes one and two steps correctly
for multi-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_multi_funcs,
self.multivariate_funcs, self.fnames):
with self.subTest(i=name):
for jdx in range(len(x_vals[:-1])):
self.mom_opt.reset()
x_vec = x_vals[jdx:jdx + 2]
x_onestep = self.mom_opt.step(f, x_vec)
x_onestep_target = x_vec - gradf(x_vec) * stepsize
self.assertAllAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.mom_opt.step(f, x_onestep)
momentum_term = gamma * gradf(x_vec)
x_twosteps_target = x_onestep - (gradf(x_onestep) +
momentum_term) * stepsize
self.assertAllAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_nesterovmomentum_optimizer_univar(self):
"""Tests that nesterov momentum optimizer takes one and two steps correctly
for uni-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_uni_fns, self.univariate_funcs,
self.fnames):
with self.subTest(i=name):
for x_start in x_vals:
self.nesmom_opt.reset()
x_onestep = self.nesmom_opt.step(f, x_start)
x_onestep_target = x_start - gradf(x_start) * stepsize
self.assertAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.nesmom_opt.step(f, x_onestep)
momentum_term = gamma * gradf(x_start)
shifted_grad_term = gradf(x_onestep -
stepsize * momentum_term)
x_twosteps_target = x_onestep - (shifted_grad_term +
momentum_term) * stepsize
self.assertAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_nesterovmomentum_optimizer_multivar(self):
"""Tests that nesterov momentum optimizer takes one and two steps correctly
for multi-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_multi_funcs,
self.multivariate_funcs, self.fnames):
with self.subTest(i=name):
for jdx in range(len(x_vals[:-1])):
self.nesmom_opt.reset()
x_vec = x_vals[jdx:jdx + 2]
x_onestep = self.nesmom_opt.step(f, x_vec)
x_onestep_target = x_vec - gradf(x_vec) * stepsize
self.assertAllAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.nesmom_opt.step(f, x_onestep)
momentum_term = gamma * gradf(x_vec)
shifted_grad_term = gradf(x_onestep -
stepsize * momentum_term)
x_twosteps_target = x_onestep - (shifted_grad_term +
momentum_term) * stepsize
self.assertAllAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_nesterovmomentum_optimizer_usergrad(self):
"""Tests that nesterov momentum optimizer takes gradient-descent steps correctly
using user-provided gradients."""
self.logTestName()
for gradf, f, name in zip(self.grad_uni_fns[::-1],
self.univariate_funcs, self.fnames):
with self.subTest(i=name):
for x_start in x_vals:
self.nesmom_opt.reset()
x_onestep = self.nesmom_opt.step(f, x_start, grad_fn=gradf)
x_onestep_target = x_start - gradf(x_start) * stepsize
self.assertAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.nesmom_opt.step(f,
x_onestep,
grad_fn=gradf)
momentum_term = gamma * gradf(x_start)
shifted_grad_term = gradf(x_onestep -
stepsize * momentum_term)
x_twosteps_target = x_onestep - (shifted_grad_term +
momentum_term) * stepsize
self.assertAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_adagrad_optimizer_univar(self):
"""Tests that adagrad optimizer takes one and two steps correctly
for uni-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_uni_fns, self.univariate_funcs,
self.fnames):
with self.subTest(i=name):
for x_start in x_vals:
self.adag_opt.reset()
x_onestep = self.adag_opt.step(f, x_start)
past_grads = gradf(x_start) * gradf(x_start)
adapt_stepsize = stepsize / np.sqrt(past_grads + 1e-8)
x_onestep_target = x_start - gradf(
x_start) * adapt_stepsize
self.assertAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.adag_opt.step(f, x_onestep)
past_grads = gradf(x_start) * gradf(x_start) + gradf(
x_onestep) * gradf(x_onestep)
adapt_stepsize = stepsize / np.sqrt(past_grads + 1e-8)
x_twosteps_target = x_onestep - gradf(
x_onestep) * adapt_stepsize
self.assertAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_adagrad_optimizer_multivar(self):
"""Tests that adagrad optimizer takes one and two steps correctly
for multi-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_multi_funcs,
self.multivariate_funcs, self.fnames):
with self.subTest(i=name):
for jdx in range(len(x_vals[:-1])):
self.adag_opt.reset()
x_vec = x_vals[jdx:jdx + 2]
x_onestep = self.adag_opt.step(f, x_vec)
past_grads = gradf(x_vec) * gradf(x_vec)
adapt_stepsize = stepsize / np.sqrt(past_grads + 1e-8)
x_onestep_target = x_vec - gradf(x_vec) * adapt_stepsize
self.assertAllAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.adag_opt.step(f, x_onestep)
past_grads = gradf(x_vec) * gradf(x_vec) + gradf(
x_onestep) * gradf(x_onestep)
adapt_stepsize = stepsize / np.sqrt(past_grads + 1e-8)
x_twosteps_target = x_onestep - gradf(
x_onestep) * adapt_stepsize
self.assertAllAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_rmsprop_optimizer_univar(self):
"""Tests that rmsprop optimizer takes one and two steps correctly
for uni-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_uni_fns, self.univariate_funcs,
self.fnames):
with self.subTest(i=name):
for x_start in x_vals:
self.rms_opt.reset()
x_onestep = self.rms_opt.step(f, x_start)
past_grads = (1 - gamma) * gradf(x_start) * gradf(x_start)
adapt_stepsize = stepsize / np.sqrt(past_grads + 1e-8)
x_onestep_target = x_start - gradf(
x_start) * adapt_stepsize
self.assertAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.rms_opt.step(f, x_onestep)
past_grads = (1 - gamma) * gamma * gradf(x_start)*gradf(x_start) \
+ (1 - gamma) * gradf(x_onestep)*gradf(x_onestep)
adapt_stepsize = stepsize / np.sqrt(past_grads + 1e-8)
x_twosteps_target = x_onestep - gradf(
x_onestep) * adapt_stepsize
self.assertAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_rmsprop_optimizer_multivar(self):
"""Tests that rmsprop optimizer takes one and two steps correctly
for multi-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_multi_funcs,
self.multivariate_funcs, self.fnames):
with self.subTest(i=name):
for jdx in range(len(x_vals[:-1])):
self.rms_opt.reset()
x_vec = x_vals[jdx:jdx + 2]
x_onestep = self.rms_opt.step(f, x_vec)
past_grads = (1 - gamma) * gradf(x_vec) * gradf(x_vec)
adapt_stepsize = stepsize / np.sqrt(past_grads + 1e-8)
x_onestep_target = x_vec - gradf(x_vec) * adapt_stepsize
self.assertAllAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.rms_opt.step(f, x_onestep)
past_grads = (1 - gamma) * gamma * gradf(x_vec) * gradf(x_vec) \
+ (1 - gamma) * gradf(x_onestep) * gradf(x_onestep)
adapt_stepsize = stepsize / np.sqrt(past_grads + 1e-8)
x_twosteps_target = x_onestep - gradf(
x_onestep) * adapt_stepsize
self.assertAllAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_adam_optimizer_univar(self):
"""Tests that adam optimizer takes one and two steps correctly
for uni-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_uni_fns, self.univariate_funcs,
self.fnames):
with self.subTest(i=name):
for x_start in x_vals:
self.adam_opt.reset()
x_onestep = self.adam_opt.step(f, x_start)
adapted_stepsize = stepsize * np.sqrt(1 - delta) / (1 -
gamma)
firstmoment = gradf(x_start)
secondmoment = gradf(x_start) * gradf(x_start)
x_onestep_target = x_start - adapted_stepsize * firstmoment / (
np.sqrt(secondmoment) + 1e-8)
self.assertAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.adam_opt.step(f, x_onestep)
adapted_stepsize = stepsize * np.sqrt(1 - delta**2) / (
1 - gamma**2)
firstmoment = (gamma * gradf(x_start) +
(1 - gamma) * gradf(x_onestep))
secondmoment = (
delta * gradf(x_start) * gradf(x_start) +
(1 - delta) * gradf(x_onestep) * gradf(x_onestep))
x_twosteps_target = x_onestep - adapted_stepsize * firstmoment / (
np.sqrt(secondmoment) + 1e-8)
self.assertAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def test_adam_optimizer_multivar(self):
"""Tests that adam optimizer takes one and two steps correctly
for multi-variate functions."""
self.logTestName()
for gradf, f, name in zip(self.grad_multi_funcs,
self.multivariate_funcs, self.fnames):
with self.subTest(i=name):
for jdx in range(len(x_vals[:-1])):
self.adam_opt.reset()
x_vec = x_vals[jdx:jdx + 2]
x_onestep = self.adam_opt.step(f, x_vec)
adapted_stepsize = stepsize * np.sqrt(1 - delta) / (1 -
gamma)
firstmoment = gradf(x_vec)
secondmoment = gradf(x_vec) * gradf(x_vec)
x_onestep_target = x_vec - adapted_stepsize * firstmoment / (
np.sqrt(secondmoment) + 1e-8)
self.assertAllAlmostEqual(x_onestep,
x_onestep_target,
delta=self.tol)
x_twosteps = self.adam_opt.step(f, x_onestep)
adapted_stepsize = stepsize * np.sqrt(1 - delta**2) / (
1 - gamma**2)
firstmoment = (gamma * gradf(x_vec) +
(1 - gamma) * gradf(x_onestep))
secondmoment = (
delta * gradf(x_vec) * gradf(x_vec) +
(1 - delta) * gradf(x_onestep) * gradf(x_onestep))
x_twosteps_target = x_onestep - adapted_stepsize * firstmoment / (
np.sqrt(secondmoment) + 1e-8)
self.assertAllAlmostEqual(x_twosteps,
x_twosteps_target,
delta=self.tol)
def classify_data(X_train, Y_train, X_test):
"""Develop and train your very own variational quantum classifier.
Use the provided training data to train your classifier. The code you write
for this challenge should be completely contained within this function
between the # QHACK # comment markers. The number of qubits, choice of
variational ansatz, cost function, and optimization method are all to be
developed by you in this function.
Args:
X_train (np.ndarray): An array of floats of size (250, 3) to be used as training data.
Y_train (np.ndarray): An array of size (250,) which are the categorical labels
associated to the training data. The categories are labeled by -1, 0, and 1.
X_test (np.ndarray): An array of floats of (50, 3) to serve as testing data.
Returns:
str: The predicted categories of X_test, converted from a list of ints to a
comma-separated string.
"""
# Use this array to make a prediction for the labels of the data in X_test
predictions = []
# QHACK #
num_qubits = 3
num_layers = 2
batch_size = 5
steps = 162
step_size = 0.15
dev = qml.device("default.qubit", wires=3)
dev2 = qml.device("default.qubit", wires=3)
def layer(W):
qml.Rot(W[0, 0], W[0, 1], W[0, 2], wires=0)
qml.Rot(W[1, 0], W[1, 1], W[1, 2], wires=1)
qml.Rot(W[2, 0], W[2, 1], W[2, 2], wires=2)
qml.CNOT(wires=[0, 1])
qml.CNOT(wires=[1, 2])
qml.CNOT(wires=[2, 0])
def statepreparation(x):
# print("state preparation x: {}" .format(x))
qml.BasisState(x, wires=[0, 1, 2])
@qml.qnode(dev)
def circuit(weights, x):
# statepreparation(x)
AngleEmbedding(x, wires=range(num_qubits))
# print(x)
for W in weights:
layer(W)
# print(qml.expval(qml.PauliZ(0)))
# StronglyEntanglingLayers ( weights , wires = range ( num_qubits ))
return qml.expval(qml.PauliZ(0))
# @qml.qnode(dev2)
# def circuit2(weights, x):
# # statepreparation(x)
# # for W in weights:
# # layer(W)
# AngleEmbedding (x , wires = range ( num_qubits ))
# StronglyEntanglingLayers ( weights , wires = range ( num_qubits ))
# return qml.expval(qml.PauliY(0))
def node_caller(weights, x, flag=False):
res = (circuit(weights, x) + circuit(weights, x))
if flag:
print(res)
return np.round(res)
def variational_classifier(var, x, flag=False):
# print("variational_classifier:\n weights {}, bias {}, \n x {}".format(var[0],var[1], x ))
weights = var[0]
bias = var[1]
return circuit(weights, x) + bias
def square_loss(labels, predictions):
loss = 0
for l, p in zip(labels, predictions):
loss = loss + (l - p)**2
loss = loss / len(labels)
return loss
def accuracy(labels, predictions):
loss = 0
for l, p in zip(labels, predictions):
if abs(l - p) < 1e-5:
loss = loss + 1
loss = loss / len(labels)
return loss
def cost(var, X, Y):
predictions = [(variational_classifier(var, x)) for x in X]
return square_loss(Y, predictions)
# Y_train += 1
# print(Y_train)
X, Y = X_train, Y_train
np.random.seed(0)
var_init = (0.01 * np.random.randn(num_layers, num_qubits, 3), 0.0)
# var_init2 = (0.01 * np.random.randn(num_layers, num_qubits, 3), 0.0)
# print(var_init)
opt = MomentumOptimizer(step_size)
# opt2 = MomentumOptimizer(step_size)
var = var_init
flag = False
for it in range(steps):
# Update the weights by one optimizer step
batch_index = np.random.randint(0, len(X), (batch_size, ))
X_batch = X[batch_index]
Y_batch = Y[batch_index]
var = opt.step(lambda v: cost(v, X_batch, Y_batch), var)
# var = opt2.step(lambda v: cost(v, X_batch, Y_batch), var)
# Compute accuracy
# predictions = [(variational_classifier(var, x,flag)) for x in X]
# acc = accuracy(Y, np.round(predictions))
# flag =False
# if not(it % 10):
# flag =False
# print(array_to_concatenated_string(np.round(predictions)))
# print(
# "Iter: {:5d} | Cost: {:0.7f} | Accuracy: {:0.7f} ".format(
# it + 1, cost(var, X, Y), acc
# )
# )
# X_test += 1
predictions = np.round([(variational_classifier(var, x, flag))
for x in X_test])
# predictions -= 1
labels_test = [
1, 0, -1, 0, -1, 1, -1, -1, 0, -1, 1, -1, 0, 1, 0, -1, -1, 0, 0, 1, 1,
0, -1, 0, 0, -1, 0, -1, 0, 0, 1, 1, -1, -1, -1, 0, -1, 0, 1, 0, -1, 1,
1, 0, -1, -1, -1, -1, 0, 0
]
# print("Prediction Accuracy : ",accuracy(labels_test,predictions))
# QHACK #
return array_to_concatenated_string((np.round(predictions)).astype(int))