Source code for qutip.control.optimizer

# -*- coding: utf-8 -*-
# This file is part of QuTiP: Quantum Toolbox in Python.
#
#    Copyright (c) 2014 and later, Alexander J G Pitchford
#    All rights reserved.
#
#    Redistribution and use in source and binary forms, with or without
#    modification, are permitted provided that the following conditions are
#    met:
#
#    1. Redistributions of source code must retain the above copyright notice,
#       this list of conditions and the following disclaimer.
#
#    2. Redistributions in binary form must reproduce the above copyright
#       notice, this list of conditions and the following disclaimer in the
#       documentation and/or other materials provided with the distribution.
#
#    3. Neither the name of the QuTiP: Quantum Toolbox in Python nor the names
#       of its contributors may be used to endorse or promote products derived
#       from this software without specific prior written permission.
#
#    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
#    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
#    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
#    PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
#    HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
#    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
#    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
#    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
#    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
#    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
#    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
###############################################################################

# @author: Alexander Pitchford
# @email1: agp1@aber.ac.uk
# @email2: alex.pitchford@gmail.com
# @organization: Aberystwyth University
# @supervisor: Daniel Burgarth

"""
Classes here are expected to implement a run_optimization function
that will use some method for optimising the control pulse, as defined
by the control amplitudes. The system that the pulse acts upon are defined
by the Dynamics object that must be passed in the instantiation.

The methods are typically N dimensional function optimisers that
find the minima of a fidelity error function. Note the number of variables
for the fidelity function is the number of control timeslots,
i.e. n_ctrls x Ntimeslots
The methods will call functions on the Dynamics.fid_computer object,
one or many times per interation,
to get the fidelity error and gradient wrt to the amplitudes.
The optimisation will stop when one of the termination conditions are met,
for example: the fidelity aim has be reached, a local minima has been found,
the maximum time allowed has been exceeded

These function optimisation methods are so far from SciPy.optimize
The two methods implemented are:
    
    BFGS - Broyden–Fletcher–Goldfarb–Shanno algorithm
        
        This a quasi second order Newton method. It uses successive calls to
        the gradient function to make an estimation of the curvature (Hessian)
        and hence direct its search for the function minima
        The SciPy implementation is pure Python and hance is execution speed is
        not high
        use subclass: OptimizerBFGS

    L-BFGS-B - Bounded, limited memory BFGS
        
        This a version of the BFGS method where the Hessian approximation is
        only based on a set of the most recent gradient calls. It generally
        performs better where the are a large number of variables
        The SciPy implementation of L-BFGS-B is wrapper around a well
        established and actively maintained implementation in Fortran
        Its is therefore very fast.
        # See SciPy documentation for credit and details on the
        # scipy.optimize.fmin_l_bfgs_b function
        use subclass: OptimizerLBFGSB

The baseclass Optimizer implements the function wrappers to the
fidelity error, gradient, and iteration callback functions.
These are called from the within the SciPy optimisation functions.
The subclasses implement the algorithm specific pulse optimisation function.
"""

import os
import numpy as np
import timeit
import scipy.optimize as spopt
import copy
import collections
# QuTiP
from qutip import Qobj
import qutip.logging_utils as logging
logger = logging.get_logger()
# QuTiP control modules
import qutip.control.optimresult as optimresult
import qutip.control.termcond as termcond
import qutip.control.errors as errors
import qutip.control.dynamics as dynamics
import qutip.control.pulsegen as pulsegen
import qutip.control.dump as qtrldump

def _is_string(var):
    try:
        if isinstance(var, basestring):
            return True
    except NameError:
        try:
            if isinstance(var, str):
                return True
        except:
            return False
    except:
        return False

    return False

[docs]class Optimizer(object): """ Base class for all control pulse optimisers. This class should not be instantiated, use its subclasses This class implements the fidelity, gradient and interation callback functions. All subclass objects must be initialised with a OptimConfig instance - various configuration options Dynamics instance - describes the dynamics of the (quantum) system to be control optimised Attributes ---------- log_level : integer level of messaging output from the logger. Options are attributes of qutip.logging_utils, in decreasing levels of messaging, are: DEBUG_INTENSE, DEBUG_VERBOSE, DEBUG, INFO, WARN, ERROR, CRITICAL Anything WARN or above is effectively 'quiet' execution, assuming everything runs as expected. The default NOTSET implies that the level will be taken from the QuTiP settings file, which by default is WARN params: Dictionary The key value pairs are the attribute name and value Note: attributes are created if they do not exist already, and are overwritten if they do. alg : string Algorithm to use in pulse optimisation. Options are: 'GRAPE' (default) - GRadient Ascent Pulse Engineering 'CRAB' - Chopped RAndom Basis alg_params : Dictionary options that are specific to the pulse optim algorithm that is GRAPE or CRAB disp_conv_msg : bool Set true to display a convergence message (for scipy.optimize.minimize methods anyway) optim_method : string a scipy.optimize.minimize method that will be used to optimise the pulse for minimum fidelity error method_params : Dictionary Options for the optim_method. Note that where there is an equivalent attribute of this instance or the termination_conditions (for example maxiter) it will override an value in these options approx_grad : bool If set True then the method will approximate the gradient itself (if it has requirement and facility for this) This will mean that the fid_err_grad_wrapper will not get called Note it should be left False when using the Dynamics to calculate approximate gradients Note it is set True automatically when the alg is CRAB amp_lbound : float or list of floats lower boundaries for the control amplitudes Can be a scalar value applied to all controls or a list of bounds for each control amp_ubound : float or list of floats upper boundaries for the control amplitudes Can be a scalar value applied to all controls or a list of bounds for each control bounds : List of floats Bounds for the parameters. If not set before the run_optimization call then the list is built automatically based on the amp_lbound and amp_ubound attributes. Setting this attribute directly allows specific bounds to be set for individual parameters. Note: Only some methods use bounds dynamics : Dynamics (subclass instance) describes the dynamics of the (quantum) system to be control optimised (see Dynamics classes for details) config : OptimConfig instance various configuration options (see OptimConfig for details) termination_conditions : TerminationCondition instance attributes determine when the optimisation will end pulse_generator : PulseGen (subclass instance) (can be) used to create initial pulses not used by the class, but set by pulseoptim.create_pulse_optimizer stats : Stats attributes of which give performance stats for the optimisation set to None to reduce overhead of calculating stats. Note it is (usually) shared with the Dynamics instance dump : :class:`dump.OptimDump` Container for data dumped during the optimisation. Can be set by specifying the dumping level or set directly. Note this is mainly intended for user and a development debugging but could be used for status information during a long optimisation. dumping : string level of data dumping: NONE, SUMMARY, FULL or CUSTOM See property docstring for details dump_to_file : bool If set True then data will be dumped to file during the optimisation dumping will be set to SUMMARY during init_optim if dump_to_file is True and dumping not set. Default is False dump_dir : string Basically a link to dump.dump_dir. Exists so that it can be set through optim_params. If dump is None then will return None or will set dumping to SUMMARY when setting a path iter_summary : :class:`OptimIterSummary` Summary of the most recent iteration. Note this is only set if dummping is on """ def __init__(self, config, dyn, params=None): self.dynamics = dyn self.config = config self.params = params self.reset() dyn.parent = self def reset(self): self.log_level = self.config.log_level self.id_text = 'OPTIM' self.termination_conditions = None self.pulse_generator = None self.disp_conv_msg = False self.iteration_steps = None self.record_iteration_steps=False self.alg = 'GRAPE' self.alg_params = None self.method = 'l_bfgs_b' self.method_params = None self.method_options = None self.approx_grad = False self.amp_lbound = None self.amp_ubound = None self.bounds = None self.num_iter = 0 self.num_fid_func_calls = 0 self.num_grad_func_calls = 0 self.stats = None self.wall_time_optim_start = 0.0 self.dump_to_file = False self.dump = None self.iter_summary = None # AJGP 2015-04-21: # These (copying from config) are here for backward compatibility if hasattr(self.config, 'amp_lbound'): if self.config.amp_lbound: self.amp_lbound = self.config.amp_lbound if hasattr(self.config, 'amp_ubound'): if self.config.amp_ubound: self.amp_ubound = self.config.amp_ubound self.apply_params() @property def log_level(self): return logger.level @log_level.setter def log_level(self, lvl): """ Set the log_level attribute and set the level of the logger that is call logger.setLevel(lvl) """ logger.setLevel(lvl)
[docs] def apply_params(self, params=None): """ Set object attributes based on the dictionary (if any) passed in the instantiation, or passed as a parameter This is called during the instantiation automatically. The key value pairs are the attribute name and value Note: attributes are created if they do not exist already, and are overwritten if they do. """ if not params: params = self.params if isinstance(params, dict): self.params = params for key in params: setattr(self, key, params[key])
@property def dumping(self): """ The level of data dumping that will occur during the optimisation - NONE : No processing data dumped (Default) - SUMMARY : A summary at each iteration will be recorded - FULL : All logs will be generated and dumped - CUSTOM : Some customised level of dumping When first set to CUSTOM this is equivalent to SUMMARY. It is then up to the user to specify which logs are dumped """ if self.dump is None: lvl = 'NONE' else: lvl = self.dump.level return lvl @dumping.setter def dumping(self, value): if value is None: self.dump = None else: if not _is_string(value): raise TypeError("Value must be string value") lvl = value.upper() if lvl == 'NONE': self.dump = None else: if not isinstance(self.dump, qtrldump.OptimDump): self.dump = qtrldump.OptimDump(self, level=lvl) else: self.dump.level = lvl @property def dump_dir(self): if self.dump: return self.dump.dump_dir else: return None @dump_dir.setter def dump_dir(self, value): if not self.dump: self.dumping = 'SUMMARY' self.dump.dump_dir = value def _create_result(self): """ create the result object and set the initial_amps attribute as the current amplitudes """ result = optimresult.OptimResult() result.initial_fid_err = self.dynamics.fid_computer.get_fid_err() result.initial_amps = self.dynamics.ctrl_amps.copy() result.evo_full_initial = self.dynamics.full_evo.copy() result.time = self.dynamics.time.copy() result.optimizer = self return result
[docs] def init_optim(self, term_conds): """ Check optimiser attribute status and passed parameters before running the optimisation. This is called by run_optimization, but could called independently to check the configuration. """ if term_conds is not None: self.termination_conditions = term_conds term_conds = self.termination_conditions if not isinstance(term_conds, termcond.TerminationConditions): raise errors.UsageError("No termination conditions for the " "optimisation function") if not isinstance(self.dynamics, dynamics.Dynamics): raise errors.UsageError("No dynamics object attribute set") self.dynamics.check_ctrls_initialized() self.apply_method_params() if term_conds.fid_err_targ is None and term_conds.fid_goal is None: raise errors.UsageError("Either the goal or the fidelity " "error tolerance must be set") if term_conds.fid_err_targ is None: term_conds.fid_err_targ = np.abs(1 - term_conds.fid_goal) if term_conds.fid_goal is None: term_conds.fid_goal = 1 - term_conds.fid_err_targ if self.alg == 'CRAB': self.approx_grad = True if self.stats is not None: self.stats.clear() if self.dump_to_file: if self.dump is None: self.dumping = 'SUMMARY' self.dump.write_to_file = True self.dump.create_dump_dir() logger.info("Optimiser dump will be written to:\n{}".format( self.dump.dump_dir)) if self.dump: self.iter_summary = OptimIterSummary() else: self.iter_summary = None self.num_iter = 0 self.num_fid_func_calls = 0 self.num_grad_func_calls = 0 self.iteration_steps = None
def _build_method_options(self): """ Creates the method_options dictionary for the scipy.optimize.minimize function based on the attributes of this object and the termination_conditions It assumes that apply_method_params has already been run and hence the method_options attribute may already contain items. These values will NOT be overridden """ tc = self.termination_conditions if self.method_options is None: self.method_options = {} mo = self.method_options if 'max_metric_corr' in mo and not 'maxcor' in mo: mo['maxcor'] = mo['max_metric_corr'] elif hasattr(self, 'max_metric_corr') and not 'maxcor' in mo: mo['maxcor'] = self.max_metric_corr if 'accuracy_factor' in mo and not 'ftol' in mo: mo['ftol'] = mo['accuracy_factor'] elif hasattr(tc, 'accuracy_factor') and not 'ftol' in mo: mo['ftol'] = tc.accuracy_factor if tc.max_iterations > 0 and not 'maxiter' in mo: mo['maxiter'] = tc.max_iterations if tc.max_fid_func_calls > 0 and not 'maxfev' in mo: mo['maxfev'] = tc.max_fid_func_calls if tc.min_gradient_norm > 0 and not 'gtol' in mo: mo['gtol'] = tc.min_gradient_norm if not 'disp' in mo: mo['disp'] = self.disp_conv_msg return mo
[docs] def apply_method_params(self, params=None): """ Loops through all the method_params (either passed here or the method_params attribute) If the name matches an attribute of this object or the termination conditions object, then the value of this attribute is set. Otherwise it is assumed to a method_option for the scipy.optimize.minimize function """ if not params: params = self.method_params if isinstance(params, dict): self.method_params = params unused_params = {} for key in params: val = params[key] if hasattr(self, key): setattr(self, key, val) if hasattr(self.termination_conditions, key): setattr(self.termination_conditions, key, val) else: unused_params[key] = val if len(unused_params) > 0: if not isinstance(self.method_options, dict): self.method_options = unused_params else: self.method_options.update(unused_params)
def _build_bounds_list(self): cfg = self.config dyn = self.dynamics n_ctrls = dyn.num_ctrls self.bounds = [] for t in range(dyn.num_tslots): for c in range(n_ctrls): if isinstance(self.amp_lbound, list): lb = self.amp_lbound[c] else: lb = self.amp_lbound if isinstance(self.amp_ubound, list): ub = self.amp_ubound[c] else: ub = self.amp_ubound if not lb is None and np.isinf(lb): lb = None if not ub is None and np.isinf(ub): ub = None self.bounds.append((lb, ub))
[docs] def run_optimization(self, term_conds=None): """ This default function optimisation method is a wrapper to the scipy.optimize.minimize function. It will attempt to minimise the fidelity error with respect to some parameters, which are determined by _get_optim_var_vals (see below) The optimisation end when one of the passed termination conditions has been met, e.g. target achieved, wall time, or function call or iteration count exceeded. Note these conditions include gradient minimum met (local minima) for methods that use a gradient. The function minimisation method is taken from the optim_method attribute. Note that not all of these methods have been tested. Note that some of these use a gradient and some do not. See the scipy documentation for details. Options specific to the method can be passed setting the method_params attribute. If the parameter term_conds=None, then the termination_conditions attribute must already be set. It will be overwritten if the parameter is not None The result is returned in an OptimResult object, which includes the final fidelity, time evolution, reason for termination etc """ self.init_optim(term_conds) term_conds = self.termination_conditions dyn = self.dynamics cfg = self.config self.optim_var_vals = self._get_optim_var_vals() st_time = timeit.default_timer() self.wall_time_optimize_start = st_time if self.stats is not None: self.stats.wall_time_optim_start = st_time self.stats.wall_time_optim_end = 0.0 self.stats.num_iter = 0 if self.bounds is None: self._build_bounds_list() self._build_method_options() result = self._create_result() if self.approx_grad: jac=None else: jac=self.fid_err_grad_wrapper if self.log_level <= logging.INFO: msg = ("Optimising pulse(s) using {} with " "minimise '{}' method").format(self.alg, self.method) if self.approx_grad: msg += " (approx grad)" logger.info(msg) try: opt_res = spopt.minimize( self.fid_err_func_wrapper, self.optim_var_vals, method=self.method, jac=jac, bounds=self.bounds, options=self.method_options, callback=self.iter_step_callback_func) amps = self._get_ctrl_amps(opt_res.x) dyn.update_ctrl_amps(amps) result.termination_reason = opt_res.message # Note the iterations are counted in this object as well # so there are compared here for interest sake only if self.num_iter != opt_res.nit: logger.info("The number of iterations counted {} " " does not match the number reported {} " "by {}".format(self.num_iter, opt_res.nit, self.method)) result.num_iter = opt_res.nit except errors.OptimizationTerminate as except_term: self._interpret_term_exception(except_term, result) end_time = timeit.default_timer() self._add_common_result_attribs(result, st_time, end_time) return result
def _get_optim_var_vals(self): """ Generate the 1d array that holds the current variable values of the function to be optimised By default (as used in GRAPE) these are the control amplitudes in each timeslot """ return self.dynamics.ctrl_amps.reshape([-1]) def _get_ctrl_amps(self, optim_var_vals): """ Get the control amplitudes from the current variable values of the function to be optimised. that is the 1d array that is passed from the optimisation method Note for GRAPE these are the function optimiser parameters (and this is the default) Returns ------- float array[dynamics.num_tslots, dynamics.num_ctrls] """ amps = optim_var_vals.reshape(self.dynamics.ctrl_amps.shape) return amps
[docs] def fid_err_func_wrapper(self, *args): """ Get the fidelity error achieved using the ctrl amplitudes passed in as the first argument. This is called by generic optimisation algorithm as the func to the minimised. The argument is the current variable values, i.e. control amplitudes, passed as a flat array. Hence these are reshaped as [nTimeslots, n_ctrls] and then used to update the stored ctrl values (if they have changed) The error is checked against the target, and the optimisation is terminated if the target has been achieved. """ self.num_fid_func_calls += 1 # *** update stats *** if self.stats is not None: self.stats.num_fidelity_func_calls = self.num_fid_func_calls if self.log_level <= logging.DEBUG: logger.debug("fidelity error call {}".format( self.stats.num_fidelity_func_calls)) amps = self._get_ctrl_amps(args[0].copy()) self.dynamics.update_ctrl_amps(amps) tc = self.termination_conditions err = self.dynamics.fid_computer.get_fid_err() if self.iter_summary: self.iter_summary.fid_func_call_num = self.num_fid_func_calls self.iter_summary.fid_err = err if self.dump and self.dump.dump_fid_err: self.dump.update_fid_err_log(err) if err <= tc.fid_err_targ: raise errors.GoalAchievedTerminate(err) if self.num_fid_func_calls > tc.max_fid_func_calls: raise errors.MaxFidFuncCallTerminate() return err
[docs] def fid_err_grad_wrapper(self, *args): """ Get the gradient of the fidelity error with respect to all of the variables, i.e. the ctrl amplidutes in each timeslot This is called by generic optimisation algorithm as the gradients of func to the minimised wrt the variables. The argument is the current variable values, i.e. control amplitudes, passed as a flat array. Hence these are reshaped as [nTimeslots, n_ctrls] and then used to update the stored ctrl values (if they have changed) Although the optimisation algorithms have a check within them for function convergence, i.e. local minima, the sum of the squares of the normalised gradient is checked explicitly, and the optimisation is terminated if this is below the min_gradient_norm condition """ # *** update stats *** self.num_grad_func_calls += 1 if self.stats is not None: self.stats.num_grad_func_calls = self.num_grad_func_calls if self.log_level <= logging.DEBUG: logger.debug("gradient call {}".format( self.stats.num_grad_func_calls)) amps = self._get_ctrl_amps(args[0].copy()) self.dynamics.update_ctrl_amps(amps) fid_comp = self.dynamics.fid_computer # gradient_norm_func is a pointer to the function set in the config # that returns the normalised gradients grad = fid_comp.get_fid_err_gradient() if self.iter_summary: self.iter_summary.grad_func_call_num = self.num_grad_func_calls self.iter_summary.grad_norm = fid_comp.grad_norm if self.dump: if self.dump.dump_grad_norm: self.dump.update_grad_norm_log(fid_comp.grad_norm) if self.dump.dump_grad: self.dump.update_grad_log(grad) tc = self.termination_conditions if fid_comp.grad_norm < tc.min_gradient_norm: raise errors.GradMinReachedTerminate(fid_comp.grad_norm) return grad.flatten()
[docs] def iter_step_callback_func(self, *args): """ Check the elapsed wall time for the optimisation run so far. Terminate if this has exceeded the maximum allowed time """ self.num_iter += 1 if self.log_level <= logging.DEBUG: logger.debug("Iteration callback {}".format(self.num_iter)) wall_time = timeit.default_timer() - self.wall_time_optimize_start if self.iter_summary: self.iter_summary.iter_num = self.num_iter self.iter_summary.wall_time = wall_time if self.dump and self.dump.dump_summary: self.dump.add_iter_summary() tc = self.termination_conditions if wall_time > tc.max_wall_time: raise errors.MaxWallTimeTerminate() # *** update stats *** if self.stats is not None: self.stats.num_iter = self.num_iter
def _interpret_term_exception(self, except_term, result): """ Update the result object based on the exception that occurred during the optimisation """ result.termination_reason = except_term.reason if isinstance(except_term, errors.GoalAchievedTerminate): result.goal_achieved = True elif isinstance(except_term, errors.MaxWallTimeTerminate): result.wall_time_limit_exceeded = True elif isinstance(except_term, errors.GradMinReachedTerminate): result.grad_norm_min_reached = True elif isinstance(except_term, errors.MaxFidFuncCallTerminate): result.max_fid_func_exceeded = True def _add_common_result_attribs(self, result, st_time, end_time): """ Update the result object attributes which are common to all optimisers and outcomes """ dyn = self.dynamics result.num_iter = self.num_iter result.num_fid_func_calls = self.num_fid_func_calls result.wall_time = end_time - st_time result.fid_err = dyn.fid_computer.get_fid_err() result.grad_norm_final = dyn.fid_computer.grad_norm result.final_amps = dyn.ctrl_amps final_evo = dyn.full_evo if isinstance(final_evo, Qobj): result.evo_full_final = final_evo else: result.evo_full_final = Qobj(final_evo, dims=dyn.sys_dims) # *** update stats *** if self.stats is not None: self.stats.wall_time_optim_end = end_time self.stats.calculate() result.stats = copy.copy(self.stats)
[docs]class OptimizerBFGS(Optimizer): """ Implements the run_optimization method using the BFGS algorithm """ def reset(self): Optimizer.reset(self) self.id_text = 'BFGS'
[docs] def run_optimization(self, term_conds=None): """ Optimise the control pulse amplitudes to minimise the fidelity error using the BFGS (Broyden–Fletcher–Goldfarb–Shanno) algorithm The optimisation end when one of the passed termination conditions has been met, e.g. target achieved, gradient minimum met (local minima), wall time / iteration count exceeded. Essentially this is wrapper to the: scipy.optimize.fmin_bfgs function If the parameter term_conds=None, then the termination_conditions attribute must already be set. It will be overwritten if the parameter is not None The result is returned in an OptimResult object, which includes the final fidelity, time evolution, reason for termination etc """ self.init_optim(term_conds) term_conds = self.termination_conditions dyn = self.dynamics self.optim_var_vals = self._get_optim_var_vals() self._build_method_options() st_time = timeit.default_timer() self.wall_time_optimize_start = st_time if self.stats is not None: self.stats.wall_time_optim_start = st_time self.stats.wall_time_optim_end = 0.0 self.stats.num_iter = 1 if self.approx_grad: fprime = None else: fprime = self.fid_err_grad_wrapper if self.log_level <= logging.INFO: msg = ("Optimising pulse(s) using {} with " "'fmin_bfgs' method").format(self.alg) if self.approx_grad: msg += " (approx grad)" logger.info(msg) result = self._create_result() try: optim_var_vals, cost, grad, invHess, nFCalls, nGCalls, warn = \ spopt.fmin_bfgs(self.fid_err_func_wrapper, self.optim_var_vals, fprime=fprime, # approx_grad=self.approx_grad, callback=self.iter_step_callback_func, gtol=term_conds.min_gradient_norm, maxiter=term_conds.max_iterations, full_output=True, disp=True) amps = self._get_ctrl_amps(optim_var_vals) dyn.update_ctrl_amps(amps) if warn == 1: result.max_iter_exceeded = True result.termination_reason = "Iteration count limit reached" elif warn == 2: result.grad_norm_min_reached = True result.termination_reason = "Gradient normal minimum reached" except errors.OptimizationTerminate as except_term: self._interpret_term_exception(except_term, result) end_time = timeit.default_timer() self._add_common_result_attribs(result, st_time, end_time) return result
[docs]class OptimizerLBFGSB(Optimizer): """ Implements the run_optimization method using the L-BFGS-B algorithm Attributes ---------- max_metric_corr : integer The maximum number of variable metric corrections used to define the limited memory matrix. That is the number of previous gradient values that are used to approximate the Hessian see the scipy.optimize.fmin_l_bfgs_b documentation for description of m argument """ def reset(self): Optimizer.reset(self) self.id_text = 'LBFGSB' self.max_metric_corr = 10 self.msg_level = None
[docs] def init_optim(self, term_conds): """ Check optimiser attribute status and passed parameters before running the optimisation. This is called by run_optimization, but could called independently to check the configuration. """ if term_conds is None: term_conds = self.termination_conditions # AJGP 2015-04-21: # These (copying from config) are here for backward compatibility if hasattr(self.config, 'max_metric_corr'): if self.config.max_metric_corr: self.max_metric_corr = self.config.max_metric_corr if hasattr(self.config, 'accuracy_factor'): if self.config.accuracy_factor: term_conds.accuracy_factor = \ self.config.accuracy_factor Optimizer.init_optim(self, term_conds) if not isinstance(self.msg_level, int): if self.log_level < logging.DEBUG: self.msg_level = 2 elif self.log_level <= logging.DEBUG: self.msg_level = 1 else: self.msg_level = 0
[docs] def run_optimization(self, term_conds=None): """ Optimise the control pulse amplitudes to minimise the fidelity error using the L-BFGS-B algorithm, which is the constrained (bounded amplitude values), limited memory, version of the Broyden–Fletcher–Goldfarb–Shanno algorithm. The optimisation end when one of the passed termination conditions has been met, e.g. target achieved, gradient minimum met (local minima), wall time / iteration count exceeded. Essentially this is wrapper to the: scipy.optimize.fmin_l_bfgs_b function This in turn is a warpper for well established implementation of the L-BFGS-B algorithm written in Fortran, which is therefore very fast. See SciPy documentation for credit and details on this function. If the parameter term_conds=None, then the termination_conditions attribute must already be set. It will be overwritten if the parameter is not None The result is returned in an OptimResult object, which includes the final fidelity, time evolution, reason for termination etc """ self.init_optim(term_conds) term_conds = self.termination_conditions dyn = self.dynamics cfg = self.config self.optim_var_vals = self._get_optim_var_vals() self._build_method_options() st_time = timeit.default_timer() self.wall_time_optimize_start = st_time if self.stats is not None: self.stats.wall_time_optim_start = st_time self.stats.wall_time_optim_end = 0.0 self.stats.num_iter = 1 bounds = self._build_bounds_list() result = self._create_result() if self.approx_grad: fprime = None else: fprime = self.fid_err_grad_wrapper if 'accuracy_factor' in self.method_options: factr = self.method_options['accuracy_factor'] elif 'ftol' in self.method_options: factr = self.method_options['ftol'] elif hasattr(term_conds, 'accuracy_factor'): factr = term_conds.accuracy_factor else: factr = 1e7 if 'max_metric_corr' in self.method_options: m = self.method_options['max_metric_corr'] elif 'maxcor' in self.method_options: m = self.method_options['maxcor'] elif hasattr(self, 'max_metric_corr'): m = self.max_metric_corr else: m = 10 if self.log_level <= logging.INFO: msg = ("Optimising pulse(s) using {} with " "'fmin_l_bfgs_b' method").format(self.alg) if self.approx_grad: msg += " (approx grad)" logger.info(msg) try: optim_var_vals, fid, res_dict = spopt.fmin_l_bfgs_b( self.fid_err_func_wrapper, self.optim_var_vals, fprime=fprime, approx_grad=self.approx_grad, callback=self.iter_step_callback_func, bounds=self.bounds, m=m, factr=factr, pgtol=term_conds.min_gradient_norm, disp=self.msg_level, maxfun=term_conds.max_fid_func_calls, maxiter=term_conds.max_iterations) amps = self._get_ctrl_amps(optim_var_vals) dyn.update_ctrl_amps(amps) warn = res_dict['warnflag'] if warn == 0: result.grad_norm_min_reached = True result.termination_reason = "function converged" elif warn == 1: result.max_iter_exceeded = True result.termination_reason = ("Iteration or fidelity " "function call limit reached") elif warn == 2: result.termination_reason = res_dict['task'] result.num_iter = res_dict['nit'] except errors.OptimizationTerminate as except_term: self._interpret_term_exception(except_term, result) end_time = timeit.default_timer() self._add_common_result_attribs(result, st_time, end_time) return result
[docs]class OptimizerCrab(Optimizer): """ Optimises the pulse using the CRAB algorithm [1]. It uses the scipy.optimize.minimize function with the method specified by the optim_method attribute. See Optimizer.run_optimization for details It minimises the fidelity error function with respect to the CRAB basis function coefficients. AJGP ToDo: Add citation here """ def reset(self): Optimizer.reset(self) self.id_text = 'CRAB' self.num_optim_vars = 0
[docs] def init_optim(self, term_conds): """ Check optimiser attribute status and passed parameters before running the optimisation. This is called by run_optimization, but could called independently to check the configuration. """ Optimizer.init_optim(self, term_conds) dyn = self.dynamics self.num_optim_vars = 0 pulse_gen_valid = True # check the pulse generators match the ctrls # (in terms of number) # and count the number of parameters if self.pulse_generator is None: pulse_gen_valid = False err_msg = "pulse_generator attribute is None" elif not isinstance(self.pulse_generator, collections.Iterable): pulse_gen_valid = False err_msg = "pulse_generator is not iterable" elif len(self.pulse_generator) != dyn.num_ctrls: pulse_gen_valid = False err_msg = ("the number of pulse generators {} does not equal " "the number of controls {}".format( len(self.pulse_generator), dyn.num_ctrls)) if pulse_gen_valid: for p_gen in self.pulse_generator: if not isinstance(p_gen, pulsegen.PulseGenCrab): pulse_gen_valid = False err_msg = ( "pulse_generator contained object of type '{}'".format( p_gen.__class__.__name__)) break self.num_optim_vars += p_gen.num_optim_vars if not pulse_gen_valid: raise errors.UsageError( "The pulse_generator attribute must be set to a list of " "PulseGenCrab - one for each control. Here " + err_msg)
def _build_bounds_list(self): """ No bounds necessary here, as the bounds for the CRAB parameters do not have much physical meaning. This needs to override the default method, otherwise the shape will be wrong """ return None def _get_optim_var_vals(self): """ Generate the 1d array that holds the current variable values of the function to be optimised For CRAB these are the basis coefficients Returns ------- ndarray (1d) of float """ pvals = [] for pgen in self.pulse_generator: pvals.extend(pgen.get_optim_var_vals()) return np.array(pvals) def _get_ctrl_amps(self, optim_var_vals): """ Get the control amplitudes from the current variable values of the function to be optimised. that is the 1d array that is passed from the optimisation method For CRAB the amplitudes will need to calculated by expanding the series Returns ------- float array[dynamics.num_tslots, dynamics.num_ctrls] """ dyn = self.dynamics if self.log_level <= logging.DEBUG: changed_params = self.optim_var_vals != optim_var_vals logger.debug( "{} out of {} optimisation parameters changed".format( changed_params.sum(), len(optim_var_vals))) amps = np.empty([dyn.num_tslots, dyn.num_ctrls]) j = 0 param_idx_st = 0 for p_gen in self.pulse_generator: param_idx_end = param_idx_st + p_gen.num_optim_vars pg_pvals = optim_var_vals[param_idx_st:param_idx_end] p_gen.set_optim_var_vals(pg_pvals) amps[:, j] = p_gen.gen_pulse() param_idx_st = param_idx_end j += 1 #print("param_idx_end={}".format(param_idx_end)) self.optim_var_vals = optim_var_vals return amps
[docs]class OptimizerCrabFmin(OptimizerCrab): """ Optimises the pulse using the CRAB algorithm [1, 2]. It uses the scipy.optimize.fmin function which is effectively a wrapper for the Nelder-mead method. It minimises the fidelity error function with respect to the CRAB basis function coefficients. This is the default Optimizer for CRAB. Notes ----- [1] P. Doria, T. Calarco & S. Montangero. Phys. Rev. Lett. 106, 190501 (2011). [2] T. Caneva, T. Calarco, & S. Montangero. Phys. Rev. A 84, 022326 (2011). """ def reset(self): OptimizerCrab.reset(self) self.id_text = 'CRAB_FMIN' self.xtol = 1e-4 self.ftol = 1e-4
[docs] def run_optimization(self, term_conds=None): """ This function optimisation method is a wrapper to the scipy.optimize.fmin function. It will attempt to minimise the fidelity error with respect to some parameters, which are determined by _get_optim_var_vals which in the case of CRAB are the basis function coefficients The optimisation end when one of the passed termination conditions has been met, e.g. target achieved, wall time, or function call or iteration count exceeded. Specifically to the fmin method, the optimisation will stop when change parameter values is less than xtol or the change in function value is below ftol. If the parameter term_conds=None, then the termination_conditions attribute must already be set. It will be overwritten if the parameter is not None The result is returned in an OptimResult object, which includes the final fidelity, time evolution, reason for termination etc """ self.init_optim(term_conds) term_conds = self.termination_conditions dyn = self.dynamics cfg = self.config self.optim_var_vals = self._get_optim_var_vals() self._build_method_options() #print("Initial values:\n{}".format(self.optim_var_vals)) st_time = timeit.default_timer() self.wall_time_optimize_start = st_time if self.stats is not None: self.stats.wall_time_optim_start = st_time self.stats.wall_time_optim_end = 0.0 self.stats.num_iter = 1 result = self._create_result() if self.log_level <= logging.INFO: logger.info("Optimising pulse(s) using {} with " "'fmin' (Nelder-Mead) method".format(self.alg)) try: ret = spopt.fmin( self.fid_err_func_wrapper, self.optim_var_vals, xtol=self.xtol, ftol=self.ftol, maxiter=term_conds.max_iterations, maxfun=term_conds.max_fid_func_calls, full_output=True, disp=self.disp_conv_msg, retall=self.record_iteration_steps, callback=self.iter_step_callback_func) final_param_vals = ret[0] num_iter = ret[2] warn_flag = ret[4] if self.record_iteration_steps: self.iteration_steps = ret[5] amps = self._get_ctrl_amps(final_param_vals) dyn.update_ctrl_amps(amps) # Note the iterations are counted in this object as well # so there are compared here for interest sake only if self.num_iter != num_iter: logger.info("The number of iterations counted {} " " does not match the number reported {} " "by {}".format(self.num_iter, num_iter, self.method)) result.num_iter = num_iter if warn_flag == 0: result.termination_reason = \ "Function converged (within tolerance)" elif warn_flag == 1: result.termination_reason = \ "Maximum number of function evaluations reached" result.max_fid_func_exceeded = True elif warn_flag == 2: result.termination_reason = \ "Maximum number of iterations reached" result.max_iter_exceeded = True else: result.termination_reason = \ "Unknown (warn_flag={})".format(warn_flag) except errors.OptimizationTerminate as except_term: self._interpret_term_exception(except_term, result) end_time = timeit.default_timer() self._add_common_result_attribs(result, st_time, end_time) return result
[docs]class OptimIterSummary(qtrldump.DumpSummaryItem): """A summary of the most recent iteration of the pulse optimisation Attributes ---------- iter_num : int Iteration number of the pulse optimisation fid_func_call_num : int Fidelity function call number of the pulse optimisation grad_func_call_num : int Gradient function call number of the pulse optimisation fid_err : float Fidelity error grad_norm : float fidelity gradient (wrt the control parameters) vector norm that is the magnitude of the gradient wall_time : float Time spent computing the pulse optimisation so far (in seconds of elapsed time) """ # Note there is some duplication here with Optimizer attributes # this exists solely to be copied into the summary dump min_col_width = 11 summary_property_names = ( "idx", "iter_num", "fid_func_call_num", "grad_func_call_num", "fid_err", "grad_norm", "wall_time" ) summary_property_fmt_type = ( 'd', 'd', 'd', 'd', 'g', 'g', 'g' ) summary_property_fmt_prec = ( 0, 0, 0, 0, 4, 4, 2 ) def __init__(self): self.reset() def reset(self): qtrldump.DumpSummaryItem.reset(self) self.iter_num = None self.fid_func_call_num = None self.grad_func_call_num = None self.fid_err = None self.grad_norm = None self.wall_time = 0.0