Why running a hybrid neural network is prohibitively slow?

PennyLane Plugins PennyLane Lightning
1

Hi,

I am trying to train a hybrid neural network with using the Torch interface. when I try to run it using the lightning devices (tried lightning.gpu, lightning.kokkos and compare with lightning.qubit on the CPU which was also significantly slower than default.qubit), it gets prohibitively slow, as it doesn’t finish a single epoch in over an hour…
Is there anything that can be done about it? I’m running pretty big torch models with a small quantum circuit in it as a single layer and this way it has to run on CPUs instead of GPUs.

n_qubits = 5
n_layers = 1

the Qlayer

Qlayer's qnode 
@qml.qnode(qdev, interface="torch")
def qnode(inputs, weights):
    qml.AngleEmbedding(inputs, wires=range(self.n_qubits))
    qml.StronglyEntanglingLayers(weights, wires=range(self.n_qubits))
    measurement = [qml.expval(qml.PauliZ(wires=i)) for i in range(self.n_qubits)]
    return measurement
network architecture 
self.model = nn.Sequential(
                                nn.Linear(2, 64),
                                nn.Tanh(),
                                nn.Linear(64, 64), 
                                nn.Tanh(),
                                nn.Linear(64, 64),   
                                nn.Tanh(),
                                nn.Linear(64, n_qubits),   
                                nn.Tanh(),
                                QLayer(n_qubits, n_layers),  # Quantum layer
                                nn.Linear(n_qubits, 2)

Thanks in advance!

2

Hi @ZivChen , unfortunately I don’t think there’s anything you can do.

I would be curious to know though, how much slower is it to run this model, compared to running the purely classical model on the same CPU?

3

Hi Catalina,

Currently it is infinitely slower, because I didn’t see it finish even one epoch… I’ll running it tonight and update you in the morning.

Best,
Ziv

4

@ZivChen if it’s infinitely slow then there might be something else going on. Are you able to share your full code to see if I can identify anything happening there?

1 Like
5

Hi @CatalinaAlbornoz,

Thanks for the reply and sorry for taking so long to reply.
I can’t share the full code as it is sensitive…
I’m currently trying other methods to speed up the hybrid network training on GPU.

Thanks,
Ziv

6

I understand @ZivChen.

In that case I’d recommend simplifying your QLayer as much as possible to try to identify the cause for the issue. Try using a single layer and seeing if this makes things faster for example. You can also try using qml.specs() or other PennyLane tools for inspecting and debugging circuits.

If you can share a code example that reproduces this behaviour then we can try to replicate it and look into it.

Let us know if any of our inspecting and debugging tools help you!

1 Like