QSVM with train+test data returns second matrix of wrong size

[zoplex]

Code is running with the following simulation: backend = BasicAer.get_backend('qasm_simulator') backend_options = {"max_parallel_threads": 0}. # RHEL has 8 CPUs, I can see 4-5 being used in parallel. quantum_instance = QuantumInstance(backend, shots=1024, seed_simulator=_seed, optimization_level=3, backend_options=backend_options) feature_map = ZZFeatureMap(feature_dimension=feature_dim, reps=2, entanglement='linear')

qsvm = QSVM(feature_map, train_data, test_data) result = qsvm.run(quantum_instance) kernel_matrix_training = result['kernel_matrix_training'] kernel_matrix_testing = result['kernel_matrix_testing']

size of the returned kernel_matrix_testing is sometimes wrong, second dimension smaller than train sample size;

example: train 250x10 (10 features), test 300x10, then returned train matrix: (250, 250) which is ok, but returned test matrix is (300, 199); that prevents using returned train/test kernel matrices to be used in downstream SVM sklearn modeling. It does not happen with every data set, just some.

• Qiskit Aqua version:
• Python version: 3.7.3 on Red Hat Enterprise Linux Server release 7.9 (Maipo)
• Operating system:
• {'qiskit-terra': '0.16.4', 'qiskit-aer': '0.7.6', 'qiskit-ignis': '0.5.2', 'qiskit-ibmq-provider': '0.12.1', 'qiskit-aqua': '0.8.2', 'qiskit': '0.24.0'}

What is the current behavior?

Steps to reproduce the problem

What is the expected behavior?

returning correct test x train kernel matrix

Suggested solutions

[woodsp-ibm]

Hi, Aqua code is now deprecated and only critical fixes are being considered. The QSVM algorithm has been superseded by the QSVC algorithm in machine learning here https://github.com/Qiskit/qiskit-machine-learning/blob/main/qiskit_machine_learning/algorithms/classifiers/qsvc.py It likewise uses a QuantumKernel but is a more direct integration with sklearn in that the quantum kernel is passed to the sklearn SVC; looks like that might be useful to you. You might also like to look at this tutorial https://qiskit.org/documentation/machine-learning/tutorials/03_quantum_kernel.html. The integration with sklearn has other advantages such as access to other tools etc of sklearn that work with the SVC etc.

[zoplex]

Thank you, in the meantime it was confirmed that the call:

kernel_matrix_testing = result['kernel_matrix_testing']

returns [test x n_support_vectors], not [test x train], and in debugger it could be seen too, so our assumption of being able to extract both train and test kernel matrices from prior version of QSVM would not work.

In the new version it seems that if one wants both train and test kernel matrices, then those need to be calculated individually and cannot be extracted from the QSVC:

adhoc_matrix_train = adhoc_kernel.evaluate(x_vec=train_features) adhoc_matrix_test = adhoc_kernel.evaluate(x_vec=test_features, y_vec=train_features)

[woodsp-ibm]

For the QSVM you can see the code and what it produces when it runs here - https://github.com/Qiskit/qiskit-aqua/blob/main/qiskit/aqua/algorithms/classifiers/qsvm/_qsvm_binary.py#L133

The QSVC is just really just a very simple sub-class of sklearn SVC with the QuantumKernel. Using the QSVC is the same therefore as using the sklearn SVC. And as that tutorial I linked above shows you can use the QuantumKernel directly with sklearn SVC and you do not even need to use the QSVC if you don't want to.