Gradients Framework (qiskit-community/qiskit-aqua#1293)

Zoufalc · Cryoris · t-imamichi · web-flow · commit 6fe815753ee8 · 2020-10-15T15:26:01.000-04:00
* clean PR with the current gradient state * Update derivatives_base.py * Update hessian.py * Update test_grad.py * add notebooks * Typing, Linting * Update gradient_framework.ipynb * typing qfi * enable gradients for VQAlgorithms * remove old files and notes * enfoce num chars <= 100 * fix an import * re-adjust import * fix other imports * additional base classes * test cases pass and the core logic all works. Doc strings could use another look. Same goes for some of the class method names inside of LinCombQFI and OverlapQFI * Update derivatives_base.py (qiskit-community/qiskit-aqua#75) * Update derivatives_base.py Added some doc-strings to several methods inside of derivatives_base.py. Also simplified the logic of `erase_operator_coeffs` by removing some leftover logic from previous implementations. * Update qiskit/aqua/operators/gradients/derivatives_base.py Co-authored-by: Julien Gacon <jul@zurich.ibm.com> * update ordering/naming * reshuffle, rename * add VQE gradient unittest * add notebook B2Meeting This notebook will be removed after today - sorry but this is the best way to share this notebook with @bryce * Update gradient_framework_reduced.ipynb * removed a leftover notebook file * removed the notebooks (and am pushing them to a cleaned up QGradients repo) * Update gradient.py This small change prevents trivial/zero operators from appearing in SummedOps during product rule computations. * will be deleted need this to pull sry * fix qfi docstring * Add retworkx version check * delete notebooks, fixed seed fin_diff circSampler, fix kwargs import error * Delete test_natural_gradient.py * fix symbol map in parameter expression grad * typing * Update qfi.py * typing * Update test_grad.py * added a docstring to unroll_operator * Update hessian.py * Update gradient.py * Fix statevector check (not always a QuantumInstance) * Update hessian.py * fix lint of qiskit/aqua/operators/gradients/*.py * fix lint of qiskit/aqua/operators/gradients/circuit_gradients/*.py * fix lint of qiskit/aqua/operators/gradients/circuit_qfis/*.py * fix test_grad * Copy folder * remove unknown aqua changes * fix ListOp, VQAlgorithms for gradients * VQE Gradients import * update operators init * rm some changes not related to gradients * mypy fixes * skip jax test if not installed * try to fix cyclic import * fix lint in isinstance * uncomment vqe tests + consistent HAS_JAX * mypy fixes * ignore mypy warning from opbase.coeff * fix isinstance used w/ typehints * try fixing sphinx * ignore cases mypy doesn't understand correctly * fix unnecessary else after raise * Add jax to github actions * fix spell * add jax dependency to actions lint * fix spell * Reduce VQE Iterations in unittest * Use ImportError * fix lint * Add gradient framework release note * update gradient framework release note * fix reno formatting errors * Update to documentation * Fix doctring to avoid warnings * update gradient init * update qaoa docstring * remover retworkx check * update docstrings * lint fixes * update docstrings to hint max_evals_grouped deprecation if gradient is given * Update test_grad.py Several high-level function calls still used a `method=...` argument in the initialization; however, we changed this to `grad_method`, `hess_method`, `qfi_method`. So I updated these arguments. This was causing some tests to not be carried out as the default argument was inferred when the correct argument name was not assigned a method type. * Update test_grad.py Added a test case for Hessian.py that forces it to differentiate an operator with a custom combo_fn * Update lin_comb circuit gradients to fix hessian evaluation * Move renamed gradient test into operator file * update lin_comb * update lin_comb gradients * fix lint * fix lin_comb * fixed the bugs that were causing Hessian logic to fail with custom combo_fns * increase coverage of natural gradient * merge zoufalc gradients * try all combinations of grad_method and qfi_method * style fix * updates for hessians * remove redundant line * make fixes * remove hessian from grad * relax VQE test * remove redundant lambda * fix identation * fix spelling * fix whitespace * revert qfi_method in test_natural_gradient * revert param name * fix lint * remove redundant comment * update identation * fix np.random.seed * increase coverage for lin_comb * fix lint * lint * whitespace fix * docstring * lint * disable misspelling * add yy, zz to pylintdict * Update __init__.py updated the gradients __init__ file with some of the content from the release note * fixed linting * update init file * fix spell * fix spelling * remove redundant import * attempt to fix sphinx qiskit-community/qiskit-aqua#1 * Update qiskit/aqua/operators/gradients/qfi.py Co-authored-by: Julien Gacon <gaconju@gmail.com> * try to fix sphinx no.2 * add words to pylintdict * fix cryoris comments * Update qiskit/aqua/operators/gradients/circuit_qfis/lin_comb_full.py Co-authored-by: Julien Gacon <gaconju@gmail.com> * Update qiskit/aqua/operators/gradients/circuit_qfis/lin_comb_full.py Co-authored-by: Julien Gacon <gaconju@gmail.com> * Update qiskit/aqua/operators/gradients/circuit_qfis/lin_comb_full.py Co-authored-by: Julien Gacon <gaconju@gmail.com> * cryoris comments * delete unneccessary * Update qiskit/aqua/operators/gradients/hessian.py Co-authored-by: Julien Gacon <gaconju@gmail.com> * raise warning hessians * include docstring TypeError * Update qiskit/aqua/operators/gradients/circuit_qfis/overlap_diag.py Co-authored-by: Julien Gacon <gaconju@gmail.com> * Update qiskit/aqua/operators/gradients/gradient.py Co-authored-by: Julien Gacon <gaconju@gmail.com> * remove unnecessary os comment * Update qiskit/aqua/operators/gradients/circuit_gradients/lin_comb.py Co-authored-by: Julien Gacon <gaconju@gmail.com> * Insert proper Error * apply changes from code review Co-authored-by: Cryoris <jules.gacon@googlemail.com> Co-authored-by: Takashi Imamichi <imamichi@jp.ibm.com> Co-authored-by: Bryce-Fuller <Bryce.Fuller@ibm.com> Co-authored-by: Julien Gacon <jul@zurich.ibm.com> Co-authored-by: Manoel Marques <manoel.marques@ibm.com> Co-authored-by: Manoel Marques <manoel@us.ibm.com> Co-authored-by: woodsp <woodsp@us.ibm.com> Co-authored-by: Steve Wood <40241007+woodsp-ibm@users.noreply.github.com> Co-authored-by: Julien Gacon <gaconju@gmail.com>
diff --git a/qiskit/aqua/algorithms/minimum_eigen_solvers/qaoa/qaoa.py b/qiskit/aqua/algorithms/minimum_eigen_solvers/qaoa/qaoa.py
@@ -20,6 +20,7 @@
 from qiskit.providers import Backend
 from qiskit.aqua import QuantumInstance
 from qiskit.aqua.operators import OperatorBase, ExpectationBase, LegacyBaseOperator
+from qiskit.aqua.operators.gradients import GradientBase
 from qiskit.aqua.components.initial_states import InitialState
 from qiskit.aqua.components.optimizers import Optimizer
 from qiskit.aqua.utils.validation import validate_min
@@ -69,6 +70,8 @@ def __init__(self,
                  initial_state: Optional[InitialState] = None,
                  mixer: Union[OperatorBase, LegacyBaseOperator] = None,
                  initial_point: Optional[np.ndarray] = None,
+                 gradient: Optional[Union[GradientBase, Callable[[Union[np.ndarray, List]],
+                                                                 List]]] = None,
                  expectation: Optional[ExpectationBase] = None,
                  include_custom: bool = False,
                  max_evals_grouped: int = 1,
@@ -88,6 +91,8 @@ def __init__(self,
                 constrained subspaces as per https://arxiv.org/abs/1709.03489
             initial_point: An optional initial point (i.e. initial parameter values)
                 for the optimizer. If ``None`` then it will simply compute a random one.
+            gradient: An optional gradient operator respectively a gradient function used for
+                      optimization.
             expectation: The Expectation converter for taking the average value of the
                 Observable over the var_form state function. When None (the default) an
                 :class:`~qiskit.aqua.operators.expectations.ExpectationFactory` is used to select
@@ -105,7 +110,8 @@ def __init__(self,
                 potentially the expectation values can be computed in parallel. Typically this is
                 possible when a finite difference gradient is used by the optimizer such that
                 multiple points to compute the gradient can be passed and if computed in parallel
-                improve overall execution time.
+                improve overall execution time. Ignored if a gradient operator or function is
+                given.
             aux_operators: Optional list of auxiliary operators to be evaluated with the eigenstate
                 of the minimum eigenvalue main result and their expectation values returned.
                 For instance in chemistry these can be dipole operators, total particle count
@@ -129,6 +135,7 @@ def __init__(self,
                          None,
                          optimizer,
                          initial_point=initial_point,
+                         gradient=gradient,
                          expectation=expectation,
                          include_custom=include_custom,
                          max_evals_grouped=max_evals_grouped,
diff --git a/qiskit/aqua/algorithms/minimum_eigen_solvers/vqe.py b/qiskit/aqua/algorithms/minimum_eigen_solvers/vqe.py
@@ -30,6 +30,7 @@
 from qiskit.aqua.algorithms import QuantumAlgorithm
 from qiskit.aqua.operators import (OperatorBase, ExpectationBase, ExpectationFactory, StateFn,
                                    CircuitStateFn, LegacyBaseOperator, ListOp, I, CircuitSampler)
+from qiskit.aqua.operators.gradients import GradientBase
 from qiskit.aqua.components.optimizers import Optimizer, SLSQP
 from qiskit.aqua.components.variational_forms import VariationalForm
 from qiskit.aqua.utils.validation import validate_min
@@ -91,6 +92,7 @@ def __init__(self,
                  var_form: Optional[Union[QuantumCircuit, VariationalForm]] = None,
                  optimizer: Optional[Optimizer] = None,
                  initial_point: Optional[np.ndarray] = None,
+                 gradient: Optional[Union[GradientBase, Callable]] = None,
                  expectation: Optional[ExpectationBase] = None,
                  include_custom: bool = False,
                  max_evals_grouped: int = 1,
@@ -108,6 +110,7 @@ def __init__(self,
             initial_point: An optional initial point (i.e. initial parameter values)
                 for the optimizer. If ``None`` then VQE will look to the variational form for a
                 preferred point and if not will simply compute a random one.
+            gradient: An optional gradient function or operator for optimizer.
             expectation: The Expectation converter for taking the average value of the
                 Observable over the var_form state function. When ``None`` (the default) an
                 :class:`~qiskit.aqua.operators.expectations.ExpectationFactory` is used to select
@@ -125,7 +128,8 @@ def __init__(self,
                 potentially the expectation values can be computed in parallel. Typically this is
                 possible when a finite difference gradient is used by the optimizer such that
                 multiple points to compute the gradient can be passed and if computed in parallel
-                improve overall execution time.
+                improve overall execution time. Deprecated if a gradient operator or function is
+                given.
             aux_operators: Optional list of auxiliary operators to be evaluated with the
                 eigenstate of the minimum eigenvalue main result and their expectation values
                 returned. For instance in chemistry these can be dipole operators, total particle
@@ -159,6 +163,7 @@ def __init__(self,
         super().__init__(var_form=var_form,
                          optimizer=optimizer,
                          cost_fn=self._energy_evaluation,
+                         gradient=gradient,
                          initial_point=initial_point,
                          quantum_instance=quantum_instance)
         self._ret = None  # type: Dict[str, Any]
@@ -412,9 +417,19 @@ def _run(self) -> 'VQEResult':
         self._quantum_instance.circuit_summary = True
 
         self._eval_count = 0
+
+        # Convert the gradient operator into a callable function that is compatible with the
+        # optimization routine.
+        if self._gradient:
+            if isinstance(self._gradient, GradientBase):
+                self._gradient = self._gradient.gradient_wrapper(
+                    ~StateFn(self._operator) @ StateFn(self._var_form),
+                    bind_params=self._var_form_params,
+                    backend=self._quantum_instance)
         vqresult = self.find_minimum(initial_point=self.initial_point,
                                      var_form=self.var_form,
                                      cost_fn=self._energy_evaluation,
+                                     gradient_fn=self._gradient,
                                      optimizer=self.optimizer)
 
         # TODO remove all former dictionary logic
diff --git a/qiskit/aqua/algorithms/vq_algorithm.py b/qiskit/aqua/algorithms/vq_algorithm.py
@@ -34,6 +34,7 @@
 from qiskit.aqua.algorithms import AlgorithmResult, QuantumAlgorithm
 from qiskit.aqua.components.optimizers import Optimizer, SLSQP
 from qiskit.aqua.components.variational_forms import VariationalForm
+from qiskit.aqua.operators.gradients import GradientBase
 
 logger = logging.getLogger(__name__)
 
@@ -47,6 +48,7 @@ def __init__(self,
                  var_form: Union[QuantumCircuit, VariationalForm],
                  optimizer: Optimizer,
                  cost_fn: Optional[Callable] = None,
+                 gradient: Optional[Union[GradientBase, Callable]] = None,
                  initial_point: Optional[np.ndarray] = None,
                  quantum_instance: Optional[
                      Union[QuantumInstance, BaseBackend, Backend]] = None) -> None:
@@ -56,6 +58,7 @@ def __init__(self,
             optimizer: A classical optimizer.
             cost_fn: An optional cost function for optimizer. If not supplied here must be
                 supplied on :meth:`find_minimum`.
+            gradient: An optional gradient operator or function for optimizer.
             initial_point: An optional initial point (i.e. initial parameter values)
                 for the optimizer.
             quantum_instance: Quantum Instance or Backend
@@ -70,6 +73,7 @@ def __init__(self,
             optimizer = SLSQP()
 
         self._optimizer = optimizer
+        self._gradient = gradient
         self._cost_fn = cost_fn
         self._initial_point = initial_point
         self._var_form = var_form
@@ -197,6 +201,9 @@ def find_minimum(self,
         start = time.time()
         if not optimizer.is_gradient_supported:  # ignore the passed gradient function
             gradient_fn = None
+        else:
+            if not gradient_fn:
+                gradient_fn = self._gradient
 
         logger.info('Starting optimizer.\nbounds=%s\ninitial point=%s', bounds, initial_point)
         opt_params, opt_val, num_optimizer_evals = optimizer.optimize(nparms,
diff --git a/test/aqua/operators/test_gradients.py b/test/aqua/operators/test_gradients.py
