Speedups For Kernel Matrix Computation?

PennyLane Help
1

Hello, I am wondering if there is a way to speed up my kernel computation via GPU or Jax optimizations.

I am currently trying to run quantum support vector machine on cybersecurity samples to train for classification. The training of the kernel matrix takes many hours depending on how many samples I use (~1000) and I think I am using the GPU accelerations available but it does not speed it up much at all vs just using regular cpu.

Is there something I am doing wrong or missing? Thank you.

# Block 1: Defining My Kernel
import pennylane as qml
from pennylane import numpy as pnp
import jax
import jax.numpy as jnp # Use JAX numpy

n_qubits = n_components # This is currently 8

dev = qml.device("lightning.gpu", wires=n_qubits)

FEATURE_MAP_REPS = 2

def angle_encoding_feature_map(x):
    qml.AngleEmbedding(features=x, wires=range(n_qubits), rotation='Z')

feature_map_to_use = angle_encoding_feature_map

@qml.qnode(dev, interface='jax')
def kernel_circuit(x1, x2):
    feature_map_to_use(x1)
    qml.adjoint(feature_map_to_use)(x2)
    return qml.expval(qml.Projector([0] * n_qubits, wires=range(n_qubits)))

@jax.jit
def kernel_jit(x1, x2):
    return kernel_circuit(x1, x2)

# Block 2: Computing the Kernel Matrix for train and test

X_train_np = jnp.array(X_train_balanced.values)
X_test_np = jnp.array(X_test.values)

print("Starting to compute training kernel matrix")
## Compute the training kernel matrix
kernel_train = qml.kernels.square_kernel_matrix(
    X_train_np,
    kernel=kernel_jit,
)

print("Starting to compute testing kernel matrix")
## Compute the testing kernel matrix
kernel_test = qml.kernels.kernel_matrix(
    X_test_np,
    X_train_np,
    kernel=kernel_jit 
)
2

Hi @sojak ,

For 8 qubits it’s unlikely that you’ll get any speedups from using GPUs (the overheads are too high) so it’s best to stick to CPU, for example lightning.qubit.

I think the issue is that you probably have a very big dataset. Encoding data into quantum computers is a big bottleneck.

Without having access to X_train_balanced and X_test I can’t run your code but my suggestion would be to reduce your dataset as much as possible.

Let me know if this answers your question.

I hope this helps!

1 Like