Add `QuantumKernel` class using terra primitives #437

ElePT · 2022-07-11T12:12:55Z

Summary

This PR presents a reference implementation of a QuantumKernel (original code by @gentinettagian) that leverages the newly introduced primitives (so far, this class is based on the terra interface).

Details and comments

The current design is pending on the approval of the fidelity primitive (see this PR), which is currently blocked by the release of the new primitive sessions.

There are a series of items in the to-do list:

Dealing with identical samples in the kernel matrix (see Evaluate quantum kernel matrix elements for identical samples #432)
[Update: final names settled] Deciding on class names -> once the new names are set, we can move the code to the kernels directory instead of creating a new primitives submodule
[Update: there will be an abstract class] Assessing whether the TrainableQuantumKernel should have a mixin/additional base class (as proposed in the latest version of the code). It makes sense for type-hinting purposes (so that we can check in algorithms whether Kernel is trainable), but in terms of functionality I don't see much advantage.
Assessing whether additional batching functionality is needed

…learning into primitives

…achine-learning into primitives

qiskit_machine_learning/primitives/kernels/base_kernel.py

dlasecki · 2022-07-18T05:59:51Z

qiskit_machine_learning/primitives/kernels/base_kernel.py

+        r"""
+        Construct kernel matrix for given data
+
+        If y_vec is None, self inner product is calculated.


Formatting of variables in docs missing.

qiskit_machine_learning/primitives/kernels/base_kernel.py

dlasecki · 2022-07-18T06:17:46Z

qiskit_machine_learning/primitives/kernels/quantum_kernel.py

+        return kernel_matrix
+
+    def _get_symmetric_kernel_matrix(
+        self, left_parameters: np.ndarray, right_parameters: np.ndarray, shape: Tuple


Tuple type missing.

dlasecki · 2022-07-18T06:20:12Z

qiskit_machine_learning/primitives/kernels/quantum_kernel.py

+        kernel_entries = self._fidelity.compute(left_parameters, right_parameters)
+        kernel_matrix = np.zeros(shape)
+        index = 0
+        for i in range(shape[0]):
+            for j in range(i, shape[1]):
+                kernel_matrix[i, j] = kernel_entries[index]
+                index += 1


Code duplication with the previous method. Could parametrize the starting point of the inner for loop - 0 vs i.

dlasecki · 2022-07-18T06:20:38Z

qiskit_machine_learning/primitives/kernels/trainable_quantum_kernel.py

+
+class TrainableQuantumKernel(QuantumKernel):
+    """
+    Trainable overlap kernel.


More info needed.

qiskit_machine_learning/utils/utils.py

dlasecki · 2022-07-18T06:24:08Z

test/primitives/kernels/test_qkernel_primitives.py

+
+
+@ddt
+class TestQuantumKernelClassify(QiskitMachineLearningTestCase):


Usually, a test class should correspond to a class being tested, not a behaviour being tested.

…achine-learning into primitives-kernel

Co-authored-by: dlasecki <dal@zurich.ibm.com>

adekusar-drl · 2022-10-13T15:58:30Z

I have not looked at the feature maps in some while. The idea of the BlueprintCircuit (what I keep calling template circuit) was that base classes can extend this and invalidate it etc over changes requiring it to be rebuilt. So setting feature_dimension as an indirect to num_qubits may be a better option, but not sure they handle that or not. Either way it comes down to use cases. When we first did Aqua we did it as you are talking about where a new instance needed to be created for each use. As more application stack logic got built out this linear progression down the stack building stuff out as you go needed to be revised. We wanted higher level iteration over the algorithms, be able to build out the algorithm, ie create/config it, for later use by the stack, but often while a user might be able to select a circuit up front the number of qubits was not known. Hence the need to be able to update in that regard. So for me it comes down to how one might expect these to be used going forwards.

feature_dimension does not work with all circuits. I'm going to keep as it is now, but this part to be revised.

woodsp-ibm · 2022-10-13T19:15:42Z

qiskit_machine_learning/algorithms/classifiers/pegasos_qsvc.py

+            elif not isinstance(quantum_kernel, BaseKernel):
+                raise TypeError("'quantum_kernel' has to be of type None or BaseKernel")


Typically we do not police types like this especially when only one type is specified by the typehint. If you wish to keep it then it might be nice to indicate the failing type in the error message too.

Removed this elif clause.

woodsp-ibm · 2022-10-13T19:18:23Z

qiskit_machine_learning/algorithms/classifiers/pegasos_qsvc.py

@@ -340,14 +345,15 @@ def precomputed(self) -> bool:
    @precomputed.setter
    def precomputed(self, precomputed: bool):
        """Sets the pre-computed kernel flag. If ``True`` is passed then the previous kernel is
-        cleared. If ``False`` is passed then a new instance of ``QuantumKernel`` is created."""
+        cleared. If ``False`` is passed then a new instance of
+        :class:`~qiskit_machine_learning.kernels.BaseKernel` is created."""


Should this be FidelityQuantumKernel - that is a BaseKernel, but is that too abstract that a user needs to know the actual type?

That's right, updated.

qiskit_machine_learning/algorithms/classifiers/qsvc.py

qiskit_machine_learning/algorithms/regressors/qsvr.py

qiskit_machine_learning/kernels/algorithms/quantum_kernel_trainer.py

qiskit_machine_learning/kernels/quantum_kernel.py

qiskit_machine_learning/kernels/trainable_fidelity_quantum_kernel.py

qiskit_machine_learning/kernels/trainable_kernel.py

releasenotes/notes/add-fidelity-quantum-kernel-d40278abb49e19b5.yaml

woodsp-ibm · 2022-10-13T20:09:28Z

Looking at the coverage its very good - BaseKernel is a fair bit lower. Error cases and properties seem to get missed often in testing where they end up not being covered.

Co-authored-by: Steve Wood <40241007+woodsp-ibm@users.noreply.github.com>

…5.yaml Co-authored-by: Steve Wood <40241007+woodsp-ibm@users.noreply.github.com>

adekusar-drl · 2022-10-14T10:03:58Z

Looking at the coverage its very good - BaseKernel is a fair bit lower. Error cases and properties seem to get missed often in testing where they end up not being covered.

Added more tests for the base class and for fidelity kernels.

woodsp-ibm

LGTM!

* Add primitives branch to CI workflow * added pocs * init fixed * more init * Add qnns PoC * lint fixed * Migrate qnns to new branch * some unittests * more unittest * adapted new design without unittests * removed compute * removed compute * Change inputs, assign parameters method name * docstrings * working on dict version * support dictionarys * Update qiskit_machine_learning/primitives/kernels/base_kernel.py Co-authored-by: dlasecki <dal@zurich.ibm.com> * Update qiskit_machine_learning/primitives/kernels/base_kernel.py Co-authored-by: dlasecki <dal@zurich.ibm.com> * Small fixes * Remove changes from trailing commit * Update qiskit_machine_learning/utils/utils.py Co-authored-by: dlasecki <dal@zurich.ibm.com> * raise error if not all parameters bound * typo * unit tests * training unit tests * feature params order fixex * lint * Create trainable mixin * Update qiskit_machine_learning/primitives/kernels/base_kernel.py Co-authored-by: dlasecki <dal@zurich.ibm.com> * Update qiskit_machine_learning/primitives/kernels/base_kernel.py Co-authored-by: dlasecki <dal@zurich.ibm.com> * Update qiskit_machine_learning/primitives/kernels/pseudo_kernel.py Co-authored-by: dlasecki <dal@zurich.ibm.com> * Fix fidelity imports (new location) * Remove pseudo kernel * evaluate duplicates * merge * black * fix some issues * update to the recent interfaces * update tests * fix fidelity parameterization * skip some tests * fix pylint, mypy * fix copyright * more tweaks * rename methods * refactor * clean up * some docstrings * more tests * streamline base class * fix spell * trainable qk tweaks * fix trainability * revert workflows * revert the notebook * renamings, refactorings, code review * code review * remove commented code * bump terra requirements * code review * code review * fix pylint * enforce psd test * fix docstring * add reno * fix documentation reference * update qsvc, more tests * update qsvr, even more tests * fix qsvr tests * fix lint * Change Terra requirements to accept rc * update pegasos * fix test * Update qiskit_machine_learning/algorithms/classifiers/pegasos_qsvc.py Co-authored-by: Gian Gentinetta <31244916+gentinettagian@users.noreply.github.com> * fix complex eigenvalues * pass parameters as keywords * fix trainability, code review * mixin -> abstract class, keywords * fix docs * update docs * code review * fix tests, pylint * update reno * update reno * fix reno * Update qiskit_machine_learning/algorithms/classifiers/qsvc.py Co-authored-by: Steve Wood <40241007+woodsp-ibm@users.noreply.github.com> * Update qiskit_machine_learning/algorithms/regressors/qsvr.py Co-authored-by: Steve Wood <40241007+woodsp-ibm@users.noreply.github.com> * Update qiskit_machine_learning/kernels/quantum_kernel.py Co-authored-by: Steve Wood <40241007+woodsp-ibm@users.noreply.github.com> * Update releasenotes/notes/add-fidelity-quantum-kernel-d40278abb49e19b5.yaml Co-authored-by: Steve Wood <40241007+woodsp-ibm@users.noreply.github.com> * code review * fix tests * terra 0.22 * more tests Co-authored-by: Manoel Marques <manoel.marques@ibm.com> Co-authored-by: Gian Gentinetta <gian.gentinetta@gmx.ch> Co-authored-by: dlasecki <dal@zurich.ibm.com> Co-authored-by: Anton Dekusar <adekusar@ie.ibm.com> Co-authored-by: Anton Dekusar <62334182+adekusar-drl@users.noreply.github.com> Co-authored-by: Gian Gentinetta <31244916+gentinettagian@users.noreply.github.com> Co-authored-by: Steve Wood <40241007+woodsp-ibm@users.noreply.github.com>

manoelmarques and others added 13 commits June 29, 2022 10:34

Add primitives branch to CI workflow

71e1bf1

added pocs

ae9c043

init fixed

3aae2df

more init

fa98e91

Add qnns PoC

5106333

lint fixed

de53484

Merge branch 'primitives' of https://github.com/ElePT/qiskit-machine-…

28d578c

…learning into primitives

Migrate qnns to new branch

cdfcfe8

Merge branch 'Qiskit:main' into primitives-kernel

f65b5f4

some unittests

5368534

Merge branch 'primitives-kernel' of https://github.com/ElePT/qiskit-m…

a1a2a7f

…achine-learning into primitives

more unittest

586c6a6

Merge branch 'Qiskit:main' into primitives-kernel

8f68be7

ElePT changed the base branch from main to primitives July 11, 2022 12:13

gentinettagian mentioned this pull request Jul 12, 2022

QuantumKernel refactor #433

Merged

manoelmarques added 2 commits July 14, 2022 09:31

Merge from main

c67ed15

Merge branch 'main'

d2342e0

dlasecki suggested changes Jul 18, 2022

View reviewed changes

gentinettagian and others added 12 commits July 19, 2022 10:18

Merge branch 'primitives-kernel' of https://github.com/ElePT/qiskit-m…

a092545

…achine-learning into primitives-kernel

adapted new design without unittests

d255e02

removed compute

52c928e

removed compute

a3cf674

Change inputs, assign parameters method name

0636e2d

docstrings

c3b0692

working on dict version

fcc670f

pulled

ba575ce

support dictionarys

5de6d92

Update qiskit_machine_learning/primitives/kernels/base_kernel.py

fef72ba

Co-authored-by: dlasecki <dal@zurich.ibm.com>

Update qiskit_machine_learning/primitives/kernels/base_kernel.py

a855070

Co-authored-by: dlasecki <dal@zurich.ibm.com>

Small fixes

a724193

fix reno

80c1954

Merge branch 'main' into primitives-kernel

7e53c38

woodsp-ibm added this to the 0.5.0 milestone Oct 13, 2022