How does Strawberry integrate with well-known machine learning libraries like Pytorch?

Photonic Software Strawberry Fields
1

Hi, I found a headache in that the photonic quantum circuits written in Strawberry cannot be converted to torchlayer, so it is not as convenient as pennylane. Also, running Strawberry took me 20 days to run through the programme once, which is extremely slow. I tried to find other platforms for photonic programming and failed.How does Strawberry integrate with famous machine learning libraries like Pytorch? Preferably GPU accelerated.

2

Hi @RX1 ,

Strawberry Fields was designed specifically to run experiments on our old Gaussian Boson Sampling devices. It’s not being actively developed at the moment, so it’s normal that you’re seeing better results with PennyLane. If you’re interested in high-level simulations, PennyLane will be the best choice.

If you’re interested in specific continuous-variable simulations, then you can try using MrMustard. Check the API reference to learn how to use it. However, it may be tricky (or impossible) to integrate it with certain classical machine learning libraries such as PyTorch.

So unfortunately it doesn’t look like any one library will have the perfect solution and you’ll have to pick the one that best suits your specific project.

1 Like
3

Hello, I’m trying to get a bit of code to be fully implemented in Mr Mustard, but I can’t find a rich case, and several attempts have failed, mainly because I can’t find a corresponding replacement for the sf’s engine there. Can you please give me the complete equivalent replacement case?

def quantum(u1, u2, is_u1_input: bool = True, shots: int = 16):
    prog = sf.Program(2)
    with prog.context as q:
        sf.ops.Vac() | q[1]
        if is_u1_input:
            sf.ops.Coherent(u1) | q[0]
        else:
            sf.ops.Coherent(u2) | q[0]
        sf.ops.BSgate() | q
        sf.ops.Dgate(-u1/np.sqrt(2)) | q[1]
        sf.ops.Dgate(-u2/np.sqrt(2)) | q[0]
        sf.ops.MeasureFock() | (q[0], q[1])
        eng = sf.Engine('gaussian')
        if eng.backend.circuit is not None:
            eng.reset()
        with warnings.catch_warnings():
            warnings.filterwarnings("ignore", category=UserWarning)
            run = eng.run(prog, shots=shots)
        bool_arr = np.all(run.samples == 0, axis=1)
        results = bool_arr.astype(int)
        return results.mean()
4

Hi @RX1 ,

There’s no exact replacement. You’ll have to build your program from scratch in MrMustard should you choose to use it. There’s no “Engine” in MrMustard, there are simply different representations.