def test_strip_slap():
import numpy as np
from qoc.models.dummy import Dummy
from qoc.core.common import (
slap_controls,
strip_controls,
)
big = 100
pstate = Dummy()
pstate.complex_controls = True
shape_range = np.arange(big) + 1
for step_count in shape_range:
for control_count in shape_range:
pstate.controls_shape = controls_shape = (step_count,
control_count)
pstate.max_control_norms = np.ones(control_count) * 2
controls = np.random.rand(
*controls_shape) + 1j * np.random.rand(*controls_shape)
stripped_controls = strip_controls(pstate, controls)
assert (stripped_controls.ndim == 1)
assert (not (stripped_controls.dtype
in (np.complex64, np.complex128)))
transformed_controls = slap_controls(pstate.complex_controls,
stripped_controls,
pstate.controls_shape)
assert (np.allclose(controls, transformed_controls))
assert (controls.shape == transformed_controls.shape)
#ENDFOR
#ENDFOR
pstate.complex_controls = False
for step_count in shape_range:
for control_count in shape_range:
pstate.controls_shape = controls_shape = (step_count,
control_count)
pstate.max_control_norms = np.ones(control_count)
controls = np.random.rand(*controls_shape)
stripped_controls = strip_controls(pstate.complex_controls,
controls)
assert (stripped_controls.ndim == 1)
assert (not (stripped_controls.dtype
in (np.complex64, np.complex128)))
transformed_controls = slap_controls(pstate.complex_controls,
stripped_controls,
pstate.controls_shape)
assert (np.allclose(controls, transformed_controls))
assert (controls.shape == transformed_controls.shape)
def _esdj_wrap(controls, pstate, reporter, result):
"""
Do intermediary work between the optimizer feeding controls to
the jacobian of _evaluate_schroedinger_discrete.
Args:
controls
pstate
reporter
result
Returns:
grads
"""
# Convert the controls from optimizer format to cost function format.
controls = slap_controls(pstate.complex_controls, controls,
pstate.controls_shape)
# Rescale the controls to their maximum norm.
clip_control_norms(controls, pstate.max_control_norms)
# Impose user boundary conditions.
if pstate.impose_control_conditions is not None:
controls = pstate.impose_control_conditions(controls)
# Evaluate the jacobian.
error, grads = (ans_jacobian(_evaluate_schroedinger_discrete,
0)(controls, pstate, reporter))
# Autograd defines the derivative of a function of complex inputs as
# df_dz = du_dx - i * du_dy for z = x + iy, f(z) = u(x, y) + iv(x, y).
# For optimization, we care about df_dz = du_dx + i * du_dy.
if pstate.complex_controls:
grads = np.conjugate(grads)
# The states need to be unwrapped from their autograd box.
if isinstance(reporter.final_states, Box):
final_states = reporter.final_states._value
else:
final_states = reporter.final_states
# Update best configuration.
if error < result.best_error:
result.best_controls = controls
result.best_error = error
result.best_final_states = final_states
result.best_iteration = reporter.iteration
# Save and log optimization progress.
pstate.log_and_save(controls, error, final_states, grads,
reporter.iteration)
reporter.iteration += 1
# Convert the gradients from cost function to optimizer format.
grads = strip_controls(pstate.complex_controls, grads)
# Determine if optimization should terminate.
if error <= pstate.min_error:
terminate = True
else:
terminate = False
return grads, terminate
def _esd_wrap(controls, pstate, reporter, result):
"""
Do intermediary work between the optimizer feeding controls
to _evaluate_schroedinger_discrete.
Args:
controls
pstate
reporter
result
Returns:
error
"""
# Convert the controls from optimizer format to cost function format.
controls = slap_controls(pstate.complex_controls, controls,
pstate.controls_shape)
# Rescale the controls to their maximum norm.
clip_control_norms(controls, pstate.max_control_norms)
# Impose user boundary conditions.
if pstate.impose_control_conditions:
controls = pstate.impose_control_conditions(controls)
# Evaluate the cost function.
error = _evaluate_schroedinger_discrete(controls, pstate, reporter)
# Determine if optimization should terminate.
if error <= pstate.min_error:
terminate = True
else:
terminate = False
return error, terminate
def _eldj_wrap(controls, pstate, reporter, result):
"""
Do intermediary work between the optimizer feeding controls to
the jacobian of _evaluate_indblad_discrete.
Args:
controls
pstate
reporter
result
Returns:
grads
"""
# Convert the controls from optimizer format to cost function format.
controls = slap_controls(pstate.complex_controls, controls,
pstate.controls_shape)
clip_control_norms(pstate.max_control_norms, controls)
# Evaluate the jacobian.
total_error, grads = (ans_jacobian(_evaluate_lindblad_discrete,
0)(controls, pstate, reporter))
# Autograd defines the derivative of a function of complex inputs as
# df_dz = du_dx - i * du_dy for z = x + iy, f(z) = u(x, y) + iv(x, y).
# For optimization, we care about df_dz = du_dx + i * du_dy.
if pstate.complex_controls:
grads = conjugate(grads)
# The states need to be unwrapped from their autograd box.
if isinstance(reporter.final_densities, Box):
final_densities = reporter.final_densities._value
# Update best configuration.
if total_error < result.best_total_error:
result.best_controls = controls
result.best_final_densities = final_densities
result.best_iteration = reporter.iteration
result.best_total_error = total_error
# Save and log optimization progress.
pstate.log_and_save(
controls,
final_densities,
total_error,
grads,
reporter.iteration,
)
reporter.iteration += 1
# Convert the gradients from cost function to optimizer format.
grads = strip_controls(pstate.complex_controls, grads)
return grads
def test_adam():
import numpy as np
from qoc.core.common import (strip_controls, slap_controls)
from qoc.models.dummy import Dummy
from qoc.standard.optimizers.adam import Adam
# Check that the update method was implemented correctly
# using hand-checked values.
adam = Adam()
grads = np.array([[0, 1], [2, 3]])
params = np.array([[0, 1], [2, 3]], dtype=np.float64)
params1 = np.array([[0, 0.999], [1.999, 2.999]])
params2 = np.array([[0, 0.99900003], [1.99900001, 2.99900001]])
adam.run(None, 0, params, None, None)
params1_test = adam.update(params, grads)
params2_test = adam.update(params1, grads)
assert (np.allclose(params1_test, params1))
assert (np.allclose(params2_test, params2))
# Check that complex mapping works and params
# without gradients are unaffected.
gstate = Dummy()
gstate.complex_controls = True
grads = np.array([[1 + 1j, 0 + 0j], [0 + 0j, -1 - 1j]])
params = np.array([[1 + 2j, 3 + 4j], [5 + 6j, 7 + 8j]])
gstate.controls_shape = params.shape
gstate.max_param_norms = np.ones(gstate.controls_shape[0]) * 10
flat_controls = strip_controls(gstate.complex_controls, params)
flat_grads = strip_controls(gstate.complex_controls, grads)
adam.run(None, 0, flat_controls, None, None)
params1 = adam.update(flat_grads, flat_controls)
params1 = slap_controls(gstate.complex_controls, params1,
gstate.controls_shape)
assert (np.allclose(params1[0][1], params[0][1]))
assert (np.allclose(params1[1][0], params[1][0]))
def _eld_wrap(controls, pstate, reporter, result):
"""
Do intermediary work between the optimizer feeding controls
to _evaluate_lindblad_discrete.
Args:
controls
pstate
reporter
result
Returns:
total_error
"""
# Convert the controls from optimizer format to cost function format.
controls = slap_controls(pstate.complex_controls, controls,
pstate.controls_shape)
clip_control_norms(pstate.max_control_norms, controls)
# Evaluate the cost function.
return _evaluate_lindblad_discrete(controls, pstate, reporter)
