/qhack-1

Primary LanguagePython

Training Quantum Embedding Kernels on Near-Term Quantum Computers

Here we share code to run the numeric experiments and generate the figures for our paper on trainable quantum embedding kernels (QEKs).

All following descriptions refer to the paper/ subdirectory. In the subdirectory qhack2021/ you can find our hackathon submission to qhack 2021.

How to use this repository

Install the requirements.txt using pip in order to be able to execute all notebooks (assuming you are in the paper/ subdirectory):

pip install -r src/requirements.txt --upgrade

This might take a moment if you start from a fresh virtual environment as packages like sklearn and tensorflow are required.

The .py files in the home directory are jupyter notebooks, stored with Jupytext for convenient synchronization. The requirements contain this package, so that after installation you can open any of the python files in jupyter-notebook, hit save and will obtain a .ipynb copy automatically, in which we then recommend to run the computations.

Some of the notebooks take considerable time to run the full simulations or even require a AWS bucket and credits to run (for the QPU experiment). In particular for the kernel training notebooks (noiseless_*), the recomputation will require hours of runtime. However, we fixed the pseudo-randomness seeds where possible and stored the corresponding output for the mitigation numerics so you can reuse those results for visualization and tinkering. For the hardware experiment we included a surrogate simulation of the experiment, though without our noise model. The mitigation of the original hardware kernel matrix is stored and seed-fixed as well but can be recomputed within about 15 minutes.

Structure

The first numeric experiment is about noiseless simulations of the QEKs and training them via kernel target alignment. These experiments can be found in the files

  • noiseless_checkerboard_and_donuts.py
  • noiseless_MNIST.py and the decision boundary plots are generated in
  • noiseless_plot.py

The second experiment is about device noise mitigation and regularization of the kernel matrix. The data is generated and processed in

  • noisy_generating.py
  • noisy_processing.py respectively. A small demo on how to use the mitigation and regularization methods introduced in our qml.kernels patch to PennyLane and on the combinations of these methods we looked at can be found in
  • post_processing_demo.py

The third experiment is about evaluation and mitigation of a QEK on a real QPU. The data generation and processing again is split up into

  • qpu_generating.py
  • qpu_processing.py

Finally there are directories

  • data/ for data (no, really!)
  • images/ for the images that are produced by the notebooks in the home directory.
  • src/ for some source code like helper functions, dataset generation etc.
  • style/ for reference files to be used in calls to the python package RSMF

In addition, the code and experimental outcomes of the updated version of the paper can be found in the folder titled revision