DyPP: Dynamic Parameter Prediction to Accelerate Convergence of Variational Quantum Algorithms

Description

This repository contains the implementation of DyPP and its application to accelerate QNNs convergence on MNIST classification problem as in:

• Paper : DyPP: Dynamic Parameter Prediction to Accelerate Convergence of Variational Quantum Algorithms
• Authors : Satwik Kundu, Debarshi Kundu and Swaroop Ghosh

Setup

To install, execute the following commands:

$ git clone https://github.com/satwik-kundu/dynamic-parameter-prediction.git
$ cd dynamic-parameter-prediction
$ pip install -r requirements.txt

Dataset Generation

In this work, we used a convolutional autoencoder (modified) to reduce feature dimension from original $28*28$ to tunable dimension $d$ specified in autoencoder.py. To generate reduced MNIST feature set of dimension 32, run:

$ python autoencoder.py --dimension 32

For d = 32, running the above command would generate the reduced dataset stored in a csv file named latent_mnist_test_d32.csv.

Build QNN, Prediction and Training

The QNN circuit builder has been modified from our previous work and individual circuit blocks are implemented in app. The pytorch implementation for the QNN model is defined in qnn_torch.py, where:

Parameter	Description	Value Range
qubit	Number of qubits	Any
enc	Features encoded per qubit	$[1, 6]$
pqc	Parameterized quantum circuit architectures	$[1, 19]$
layers	Number of pqc layers	Any
meas	Measurement operations	$[1, 5]$

The hyperparameters for our parameter prediction technique are also defined in qnn_torch.py, where:

Parameter	Description	Value Range
p	Prediction interval	$[4, -]$
d	Prediction distance	$[4, 8]$
k	Proportionality constant	0.0001 (constant)
dr	Decay rate	0.95 (constant)

After setting the hyperparameters, to train a QNN model on dataset latent_mnist_test_d32.csv, choose encoding and qubit count accordingly and run:

python qnn_torch.py latent_mnist_test_d32.csv

Citation

If this project contributes to your research, kindly cite our paper:

@article{kundu2023dypp,
  title={DyPP: Dynamic Parameter Prediction to Accelerate Convergence of Variational Quantum Algorithms},
  author={Kundu, Satwik and Kundu, Debarshi and Ghosh, Swaroop},
  journal={arXiv preprint arXiv:2307.12449},
  year={2023}
}