---

Quantum-Assisted Restricted Boltzmann Machine Training Framework

This package provides a framework for training Restricted Boltzmann Machines (RBMs) via classical algorithms (Contrastive Divergence and Persistent Contrastive Divergence) and quantum-assisted methods, which involve the use of Quantum Samplimg.

Using the QUBO.jl package, this package allows training RBMs using different quantum computers and simulators and converting continuous visible nodes to binary visible nodes seamlessly.

Installation

] add www.github.com/NITeQ/QARBoM.jl

Getting started

Defining your dataset and RBM

using QARBoM

# Get a dataset. Should be a Vector{Vector{<:Number}} where each inner vector is a sample.
train_data = MY_DATA_TRAIN
test_data = MY_DATA_TEST


# Create a new RBM with:
# - 10 visible nodes (the size of each sample)
# - 5 hidden nodes (the number of hidden nodes that you can choose)

rbm = RBM(10, 5)

Training RBM using Contrastive Divergence

# Train the RBM using Persistent Contrastive Divergence
N_EPOCHS = 100
BATCH_SIZE = 10

QARBoM.train!(
    rbm, 
    train_data,
    CD; 
    n_epochs = N_EPOCHS,  
    cd_steps = 3, # number of gibbs sampling steps
    learning_rate = [0.0001/(j^0.8) for j in 1:N_EPOCHS], 
    metrics = [MeanSquaredError], # the metrics you want to track
    x_test_dataset = test_data,
    early_stopping = true,
    file_path = "my_cd_metrics.csv",
)

Training RBM using Persistent Contrastive Divergence

# Train the RBM using Persistent Contrastive Divergence
N_EPOCHS = 100
BATCH_SIZE = 10

QARBoM.train!(
    rbm, 
    train_data,
    PCD; 
    n_epochs = N_EPOCHS, 
    batch_size = BATCH_SIZE, 
    learning_rate = [0.0001/(j^0.8) for j in 1:N_EPOCHS], 
    metrics = [MeanSquaredError], # the metrics you want to track
    x_test_dataset = test_data,
    early_stopping = true,
    file_path = "my_pcd_metrics.csv",
)

Train RBM using Quantum Sampling

using DWave # or any other quantum sampler


# Define a setup for you quantum sampler
MOI = QARBoM.ToQUBO.MOI
MOI.supports(::DWave.Neal.Optimizer, ::MOI.ObjectiveSense) = true

function setup_dwave(model, sampler)
  MOI.set(model, MOI.RawOptimizerAttribute("num_reads"), 25)
  MOI.set(model, MOI.RawOptimizerAttribute("num_sweeps"), 100)
end


QARBoM.train!(
    rbm, 
    train_data,
    QSampling; 
    n_epochs = N_EPOCHS, 
    batch_size = 5, 
    learning_rate = [0.0001/(j^0.8) for j in 1:N_EPOCHS], 
    x_test_dataset = test_data,
    early_stopping = true,
    file_path = "qubo_train.csv",
    model_setup=setup_dwave,
    sampler=DWave.Neal.Optimizer,
)

RBM for classification

You can use the RBMClassifier to train an RBM for classification. It's architecture was based on a paper by Larochelle et al. (see Figure 1 from this paper).

using QARBoM

rbm = RBMClassifier(
    10, # number of visible nodes
    5, # number of hidden nodes
    2, # number of nodes for the label
)

QARBoM.train!(
    rbm, 
    train_data,
    y_train,
    PCD; 
    n_epochs = N_EPOCHS, 
    batch_size = BATCH_SIZE, 
    learning_rate = [0.0001/(j^0.8) for j in 1:N_EPOCHS], 
    label_learning_rate = [0.001/(j^0.6) for j in 1:N_EPOCHS], 
    metrics = [Accuracy],
    x_test_dataset = test_data,
    y_test_dataset = y_test,
    early_stopping = true,
    file_path = "my_pcd_metrics_classification.csv",
)

Non-binary visible nodes

You can work with continuous visible nodes with QARBoM.

According to Hinton's article, you need to normalize your dataset to have zero mean and unit variance.

In order to use continuous visible nodes for quantum sampling, there is an extra step. We leverage QUBO.jl to convert continuous visible nodes to binary visible nodes. To do so, you need to define the maximum and minimum values for each visible node. See the example below:

using QARBoM, DWave

train_data = MY_DATA_TRAIN
test_data = MY_DATA_TEST

label_train_data = MY_DATA_TRAIN
label_test_data = MY_DATA_TEST

rbm = RBMClassifier(
    10, # number of visible nodes
    5, # number of hidden nodes
    2, # number of nodes for the label
)

QARBoM.train!(
    rbm, 
    train_data,
    label_train_data,
    QSampling; 
    n_epochs = N_EPOCHS, 
    batch_size = 5, 
    learning_rate = [0.0001/(j^0.8) for j in 1:N_EPOCHS], 
    label_learning_rate = [0.0001/(j^0.8) for j in 1:N_EPOCHS], 
    metrics = [Accuracy],
    x_test_dataset = test_data,
    y_test_dataset = label_test_data,
    early_stopping = true,
    file_path = "qubo_train.csv",
    model_setup=setup_dwave,
    sampler=DWave.Neal.Optimizer,
    max_visible = max_visible, # Vector{Float64}
    min_visible = min_visible # Vector{Float64}
)