qiboteam / rtqem

Impact of error mitigation using a quantum device as a regressor.

RTQEM

Code related to the paper: arXiv:2311.05680.

The rtqem algorithm implements a Real-Time Quantum Error Mitigation procedure to perform multi-dimensional fit in a quantum noisy scenario.

The optimization is gradient-based and makes use of the Adam optimizer.

In a few words, in the RTQEM we use a learning-based quantum error mitigation method (Importance Clifford Sampling) to mitigate both gradients and predictions during the gradient descent execution. In an evolving-noise scenario, it is possible to set a threshold which triggers the re-learning of the noise map when the system has significantly changed since the last learnt map.

A schematic representation of the algorithm follows:

Introduction to the usage

The code is organized as follows:

• The main functions are: training.py and vqregressor.py. The first script is called to train a variational quantum regressor initialized according to the second script.
• The training procedure can be set up as follows:
  • select one of the available targets, which can be found in (src/rtqem/targets/.);
  • once the target is defined, one can fill the target's configuration file with all the desired hyper-parameters. As an example, if the uquark target is selected, the src/rtqem/uquark.conf file can be used to define the settings of the training.

Run an example of RTQEM optimization!

As an example, we provide the instructions to run the uquark fit with and without RTQEM.

The uquark configuration file is already set up to run in RTQEM mode in a strong-noise scenario. In order to run the optimization:

cd src/rtqem/
python training.py uquark

After the execution, some data will be generated:

• ./src/rtqem/liveshow.png is a live-plotting of the fit;
• ./src/rtqem/uquark/ will contain data.npy and labels.npy, corresponding to the training input and output data.
• src/rtqem/cache will contain the best parameters collected during the optimization, the loss function history, the gradients history, a final plot of the predictions on the training sample and a folder in which the parameters are saved epoch by epoch. All the described output will be saved with a label which describes the optimization configuration. More details about this can be found in src/rtqem/savedata_utils.py.

To run the training without RTQEM, edit the file src/uquark/uquark.conf by replacing the mitigation config. line 9 with:

  "mitigation": {"step":false,"final":false,"method":null, "readout":null},

This will result in a noisy training simulation without RTQEM.

How to customize the training experience?

Many hyper-parameters can be used to customize the training. A detailed list follows:

• ndim (int): the dimensionality of the problem. The uquark target, for example, is mono-dimensional.
• nqubits (int): number of qubits used to build the parametric circuit. In our ansatz, this parameter must be equal to ndim.
• nlayers (int): number of layers of the quantum machine learning model.
• function (str): target name.
• normalize_data (bool): if True the target function is normalized between [0,1].
• noise (bool): if True, local Pauli noise is injected into the system.
• noise_update (int): every noise_update epochs the noise changes according to a specific evolution model.
• noise_threshold (float): if this threshold is exceeded and a mitigation method is selected, the noise map is re-learned.
• evolution_model (string): can be one between "random_walk", "diffusion", "heating".
• diffusion_parameter (float): mutation step of the noise into the selected noise evolution model.
• qm (float): readout noise parameter.
• noise_magnitude ([float, float, float]): local Pauli noise parameters (they must be positive and their sum must be lower than or equal to one).
• bp_bound: compute and print the bound imposed by Noise Induced Barren Plateaus according to the selected number of qubits, number of layers and noise magnitude.
• mitigation (dict): mitigation arguments:
  • step (bool): if true, gradients and predictions are mitigated over training;
  • final (bool): if true, the final predictions only are mitigated;
  • method (str): can be mit_obs or CDR. We suggest to set mit_obs, since it corresponds to the last results presented in the paper. Set null if no mitigation is desired.
  • readout (str): can be "calibration_matrix" (link to the paper), "ibu" (link to the paper) or "randomized" (link to the paper). Set null if no readout mitigation is required.
• expectation_from_samples (bool): if False, exact simulation is performed. If True, shot-noise simulation is performed with set number of shots.
• nshots (int) number of shots for each circuit evaluation.
• optimizer (str) can be "Adam" (gradient-based) or "cma" (evolutionary strategy). We suggest to set "Adam", since it corresponds to the last results of the paper.
• epochs (int): number of epochs of the optimization.
• ndata (int): number of training data.
• batchsize (int): size of the batches if mini-batch gradient descent is performed.
• learning_rate (float): Adam's learning rate.
• restart from epoch (int): restart from a specific training epoch using the cached parameters.
• nthreads (int): number of threads on which the user wants to parallelise the code.