Quantum Assisted Machine Learning

QAML is a library for generative learning models built as an extension of PyTorch. This library provides a set of custom Pytorch classes based on Pytorch Module, Function, and Dataset to allow easy integration with Pytorch models.

This compatibility and modularity is what makes possible the use of Quantum Processing Units to accelerate the sampling process required in generative models.

QAML Configurations:

• Classic Restricted Boltzmann Machine (RBM)

• Quantum-Assisted Restricted Boltzmann Machine (QARBM)

• Classic Boltzmann Machine (BM)

• Quantum-Assisted Boltzmann Machine (QABM)

Code Example

import qaml
import torch

# Model
bm = qaml.nn.BoltzmannMachine(36,16)

# Optimizer
opt = torch.optim.SGD(bm.parameters(), ...) # Hyperparameters

# Positive-phase Sampler
SAMPLER_NAME = "Advantage_system4.1" # Example of D-Wave Sampler
pos = qaml.sampler.BatchQASampler(bm,SAMPLER_NAME,vmask=True)
POS_BATCH = len(pos_sampler.batch_embeddings)

# Negative-phase Sampler
neg = qaml.sampler.BatchQASampler(bm,SAMPLER_NAME)
NEG_BATCH = len(neg_sampler.batch_embeddings)

# Automatic Differentiation
ML = qaml.autograd.MaximumLikelihood

# Dataset
data = qaml.datasets.BAS(6,6)
loader = torch.utils.data.DataLoader(data,POS_BATCH)

# Training loop
err_log = []
for epoch in range(N):
    epoch_error = torch.Tensor([0.])
    for img_batch in loader:

        # Positive Phase
        v0, h0 = pos(img_batch.detach(),num_reads=100)
        
        # Negative Phase
        vq, hq = neg(num_reads=1000)

        # Reconstruction error from Contrastive Divergence
        err = ML.apply((v0,h0),(vq,hq), *bm.parameters())

        # Do not accumulate gradients
        optimizer.zero_grad()

        # Compute gradients
        err.backward()

        # Update parameters
        optimizer.step()

        #Accumulate error for this epoch
        epoch_error  += err
    # Error Log
    err_log.append(epoch_error.item())
    print(f"Epoch {t} Recon Error = {epoch_error.item()}")

References:

I want to acknowledge the projects I used as inspiration for QAML.

Gabriel Bianconi's pytorch-rbm: Perfect example of a baseline RBM implementation along with MNIST example.

Jan Melchior's PyDeep: Most thorough Boltzmann Machine learning library out there with impressive modularity.

PIQUIL's QuCumber: Practical application with a PyTorch RBM backend.