Deep Neural Decoders for Fault-Tolerant Quantum Error Correction

This repository contains source codes and results reported in

C. Chamberland, P. Ronagh, Deep neural decoders for near term fault-tolerant experiments, arXiv 2018

The package provides a general framework for training and using deep learning for fault tolerant protocols that use CSS codes. The methods implemented here can have practical applications for near term fault-tolerant experiments.

Prerequisites

The implementation has been tested with Mac OS 10.12, Mac OS 10.13, and on Linux Debian 9 (4.9.65-3+deb9u1 (2017-12-23) x86_64 and 4.9.51-1 (2017-09-28) x86_64) but should work with most Mac OSX and Linux systems. To run the codes you need to have the following installed.

Python 2.7
BayesOpt - A Bayesian optimization library
Tensorflow Tensorflow 1.4

To run the circuit simulations Matlab is required.

Reports

The following neural networks were tested and the results are available in the Reports folder. The number of hidden layers for feedforward networks was incremented until no further improvement was observed in the performance.

FTEC	T	D	RNN	FF0	FF1	FF2	FF3	FF4	CNN	B Range	Tune
Steane	PU	3	###		###	###	###			1e-4 6e-4	4e-4
Steane	LU	3	###	###	###	###	###	###		1e-4 6e-4	4e-4
Knill	PU	3	###			###	###			1e-4 6e-4	4e-4
Knill	LU	3	###	###	###	###	###	###		1e-4 6e-4	4e-4
Surface	PU	3	###	###	###	###				1e-4 6e-4	4e-4
Surface	LU	3	###	###	###	###				1e-4 6e-4	4e-4
Steane	PU	5	###			###	###			6e-4 2e-3	1e-3
Steane	LU	5	###			###	###			6e-4 2e-3	1e-3
Knill	PU	5				###				6e-4 2e-3	4e-4
Knill	LU	5				###				6e-4 2e-3	4e-4
Surface	PU	5	###			###			###	3e-4 8e-4	6e-4
Surface	LU	5	###			###			###	3e-4 8e-4	6e-4

PU. Pure-Error quantum error correction as base decoder
LU. Lookup table qunatum error correction as base decoder
Steane. Steane EC protocol for fault-tolerant error correction. In this experiment the exRec CNOT was simulated.
Knill. Knill EC protocol for fault-tolerant error correction. In this experiment the exRec CNOT was simulated.
Surface. Rotated surface code single EC in fault-tolerant protocols (3 rounds of EC for surface code of distance 3, and 6 rounds of EC for distance 5).
B Range. Range of physical error rates used for depolarizing noise channel.
Tune. Physical error rate at which hyperparameter tuning was performed using BayesOpt.

Reproducing the results

Here I explain how to reproduce our results and/or build upon them. Feel free to contact me for further details, discussion, issues and feedback.

Pickle your dataset

The training mode uses pickle files. To generate pickles from your dataset use

python -u Run.py gen param_file

where param_file is the addresss of a json file like this.

Containing the following keys in this example:

    "FT scheme": "ExRecCNOT", 
    "EC scheme": "ColorD5", 
    "raw folder": "../../Data/Compact/Steane_CNOT_D5/",
    "pickle folder": "../../Data/Pkl/LookUp/Steane_CNOT_D5/",
    "report folder": "../../Reports/LookUp/Steane_CNOT_D5/",
    "param folder": "../../Param/LookUp/Steane_CNOT_D5/",
    "look up": true

raw folder is a text file with zeros and ones in the format explained below. The results will be saved in pickle folder. If the training set is for the PE-based method the look up should be set to false.

Hyperparameter tuning

The choice of neural network to use and the hyperparameters used by it can be set in a json file like this.

You can alternatively use hyperparameter tuning feature I implemented:

python -u Run.py tune param_file hyperparam_range_file

param_file is again a param like, e.g. this. hyperparam_range_file is a file putting box-constrains on the range of each hyperparam in the domain explored in Bayesian optimization, E.g. this.

Train using your pickle

Once you have set all your parameters in a json file like this then run

python -u Run.py bench training_param_file [index_0] [index_1]

index_0 and index_1 lets you do training and benchmarking on a small range of files available in pickle folder.

Add your favorite neural network

There are many neural networks implemented in TensorFlow in Networks.py. Many of them proved useless. I will let them be there as examples of what I have tried. In the paper, only the following neural networks are used:

surface_conv3d_cost
ff_cost
rnn_cost

You can add your model and a line to cost_function in Model.py and you are good to go.

Add your favorite CSS code

Examples _*.py shows how to add a new CSS code. Choose a name and use it in env parameter of the param file. The new _*.py file has to be imported as lookup like in this line.

License

This project is licensed under the MIT License - see the LICENSE.md file for details.

trxw / DeepNeuralDecoder