Quantum Machine Learning Undergraduate Project Repo

This work is based off the Xanadu AI tutorial [1] for quantum transfer learning and aims to provide a framework for quantum transfer learning applications. The Xanadu notebook as been refactored into a more traditional machine learning library, consisting of a main testing/training script, a classical neural network module, and various quantum modules. The purpose is to give researchers an library for developing and testing quantum machine learning algorithms and architectures.

After cloning and installing all the requirements from qml/requirements.txt, the main training/testing function is called via

~/QML/qml$ python main.py 

To see a list of optional parameters for switching modes between quantum models and classical models and tuning hyperparamteres:

~/QML/qml$ python main.py -h

Example training usage might be:

~/QML/qml$ python main.py --model quantum --num_epoch 5 --depth 5 --qubits 4 --mode train

(note that n_qubits refers to the number of nodes in each layer of a classical analogue. Interesting for comparing quantum to classical for the same architecture, however on our hymenoptera dataset the default depth of 6 nodes results in overfiting)

To create and test custom quantum variational circuits, see citcuits.py. There are already some custom circuits which have been implemented from [2]. A custom circuit class must be derived from the class QuantumCircuit, which defines the circuits device and forward pass. The only thing required from a custom circuit class is the torch parameter definitions, which can be seen in the __init__ function below. Then, define each layer of the circuit in the layer method. rx_layer and ry_layer are from the layers.py module, which can be expanded for ease of use. The inputs into layer are qubits, which defines the number of qubits used for the circuit, and w, which are the weights for any tunable gates.

class Circuit2(QuantumCircuit):
    def __init__(self, qubits, depth, delta, dev):
        super().__init__(qubits, depth, delta, dev)
        params = 2 * qubits
        self.params_per_layer = torch.Tensor([params]).type(torch.int32)
        self.q_params = nn.Parameter(delta * torch.randn(self.params_per_layer * depth))

    @staticmethod
    def layer(qubits, w):
        rx_layer(w[:qubits])
        rz_layer(w[qubits:])
        for i in range(qubits):
            qml.CNOT(wires=[i, i + 1])

Calling a custom circuit class can be done with the --circuit command line argument, where the following string must match the class name from circuits.py (case sensitive). For example:

~/QML/qml$ python main.py --num_epoch 10 --depth 2 --qubits 4 --circuit Circuit2

A simple command line argument has been included to run any training (not recommended) or testing (more recommended) on real quantum hardware. Simple include the --backend flag with a proper ibmq backend name such as:

~/QML/qml$ python main.py --num_epoch 10 --depth 2 --qubits 4 --backend ibmq_london

This might require prior account credentials loading on your machine.

[1] A. Mari, T. R. Bromley, J. Izaac, M. Schuld, and N. Killo-ran, “Transfer learning in hybrid classical-quantum neu-ral networks,”arXiv:1912.08278 [quant-ph, stat], Dec.2019.

[2] S. Sim, et al. “Expressibility and Entangling Capability of Parameterized Quantum Circuits for Hybrid Quantum‐Classical Algorithms.” (2019).