Typical Machine Learning Datasets as Low-Depth Quantum Circuits

This project provides a comprehensive toolset for representing image datasets as quantum circuits, optimizing them, and evaluating the impact of this optimization on classification tasks. The workflow is divided into three main stages: data preparation, quantum circuit optimization, and classification. The plots directory contains scripts and notebooks to visualize the results of these stages.

Getting Started

Prerequisites

You can install the required packages using the requirements.txt file after creating a environment with python==3.12.0:

conda create -n "quantum_datasets" python==3.12.0
conda activate quantum_datasets
pip install -r requirements.txt

Installation

1. Clone the repository:

   git clone https://github.com/fkiwit/low-depth-image-circuits
   cd low-depth-image-circuits

2. The project uses custom scripts and does not require a formal installation. Ensure all dependencies from requirements.txt are installed.

Workflow

The project is structured around a three-step pipeline:

1. Data Preparation

The first step is to download and process the image datasets. The prepare_data.py script handles this.

• Supported Datasets: MNIST, Fashion-MNIST, CIFAR-10, and Imagenette.
• Processing: Images are resized, cropped, and converted to either RGB or grayscale.
• Quantum Encoding: The processed images are encoded into quantum states using the Flexible Representation of Quantum Images (FRQI) method.
• Output: The script saves the processed images, labels, and their corresponding quantum states as .npy files in the data/ directory.

Usage:

python prepare_data.py --dataset_name <dataset> [--n_patches <num_patches>]

Example:

python prepare_data.py --dataset_name mnist --n_patches 1

2. Quantum Circuit Optimization

The circuit_optimization/ directory contains scripts to optimize the quantum state representations of the images into shallow quantum circuits.

• Core Logic: Scripts like run_sweeping.py, run_bfgs.py, and run_sweeping_random.py load quantum states and use various optimization algorithms (sweeping, BFGS, randomized) to find a low-depth quantum circuit that approximates each state.
• Data Collection: The data_collector_*.py scripts are used to gather data from the optimization runs.
• Parallelization: The optimization process can be parallelized using ray for efficiency.
• Circuit Types: Different types of circuits can be used for optimization, such as unitary, orthogonal, or sparse, which correspond to different gate sets and circuit structures.
• Output: The scripts save the parameters of the optimized circuits, which can be used for classification.

Usage:

The run_sweeping_batch.sh, run_sweeping_local.sh, and run_bfgs_local.sh scripts can be used to run the optimizations.

Example of running run_sweeping.py directly:

cd circuit_optimization
python run_sweeping.py --dataset mnist --circuit orthogonal --layers 4

3. Classification

The classifier/ directory is dedicated to training various classifiers on the (potentially optimized) quantum data to evaluate the performance of the optimization.

• Goal: To assess how different levels of optimization affect classification accuracy.
• Supported Classifiers:
  • Support Vector Machines (SVM): With a custom quantum kernel (utils/svm_training.py).
  • Tensor Network Classifiers: Matrix Product State (MPS) and Matrix Product Operator (MPO) based classifiers (utils/tensor_network_training.py).
  • Variational Quantum Classifiers (VQC): (utils/vqc_training.py).
  • Classical Convolutional Neural Networks (CNN): As a baseline (utils/cnn_training.py).
• Training Orchestration: The launch_training_NN.py script (for neural networks, VQCs, and tensor networks) and launch_training_SVM.sh (for SVMs) manage the training and hyperparameter tuning, using ray.tune and GNU parallel.

Usage:

To train the neural network based classifiers:

cd classifier
python launch_training_NN.py

To train SVM classifiers:

cd classifier
./launch_training_SVM.sh

Project Structure

.
├── circuit_optimization/       # Scripts for quantum circuit optimization
│   ├── utils/                  # Utility functions for optimization
│   ├── run_sweeping.py         # Main script for sweeping optimization
│   ├── run_bfgs.py             # Main script for BFGS optimization
│   └── ...
├── classifier/                 # Scripts for training classifiers
│   ├── utils/                  # Implementations of different classifiers
│   ├── launch_training_NN.py   # Script to launch NN-based training
│   ├── launch_training_SVM.sh  # Script to launch SVM training
│   └── ...
├── data/                       # Default directory for datasets
├── plots/                      # Contains scripts and notebooks for generating plots and visualizations of the results from the optimization and classification tasks.
├── prepare_data.py             # Script for data preparation and quantum encoding
├── requirements.txt            # Python dependencies
└── README.md                   # This file

Pennylane Datasets

The datasets are also available on Pennylane Datasets: https://pennylane.ai/datasets/collection/low-depth-image-circuits

Citation

If you use this project in your research, please cite the following paper:

@article{10.1088/2058-9565/ae0123,
	author={Kiwit, Florian J and Jobst, Bernhard and Luckow, Andre and Pollmann, Frank and Riofrío, Carlos A.},
	title={Typical Machine Learning Datasets as Low-Depth Quantum Circuits},
	journal={Quantum Science and Technology},
	url={http://iopscience.iop.org/article/10.1088/2058-9565/ae0123},
	year={2025}
}

The paper is available at: https://iopscience.iop.org/article/10.1088/2058-9565/ae0123