GTorlai
commented
2 years ago Hi @JoeGibbs88, when using AD functionalities through Zygote sometimes a specific syntax is required. In your example there are a few things that needs to be changed to make it work. First, when using vcat to build the circuit you should also include the square brakets around each new gate:
function variationalcircuit(params)
circuit = Tuple[]
circuit = vcat(circuit, [("X", 1)])
circuit = vcat(circuit, [("G", (1, 2), (θ = params[1],))])
circuit = vcat(circuit, [("Phase", 1, (ϕ = params[2],))])
circuit = vcat(circuit, [("Phase", 2, (ϕ = params[3],))])
circuit = vcat(circuit, [("G", (1, 2), (θ = -params[1],))])
return circuit
endNext, the loss function needs also couple of changes. First, differentiation through an MPS constructor is not yet defined in ITensors (see https://github.com/ITensor/ITensors.jl/issues/776). So for now you need to create the initial state outside the loss function. The other change is in the fidelity call, which is also currently not differentiable. For the time being, you could instead use the inner function in ITensor, which in this case computes the inner product of two MPSs, and build the fidelity from it. The reason fidelity is non-differentiable is because we wrote it to be numerically stable for large MPS, which requires some additional operators on top of inner whose derivatives are not yet taken care of (though we shall make a fix for this soon). So the loss function becomes:
psi0 = productstate(hilbert)
function loss(params)
circuit = variationalcircuit(params)
psi = runcircuit(psi0, circuit)
return 1 - abs2(inner(psi, target))
endWith the above modification the code runs and outputs:
0.05545295586765009
[ Info: LBFGS: initializing with f = 0.055452955868, ‖∇f‖ = 2.8750e-01
[ Info: LBFGS: iter 1: f = 0.001045920807, ‖∇f‖ = 3.7467e-02, α = 4.30e-01, m = 0, nfg = 2
[ Info: LBFGS: iter 2: f = 0.000157296908, ‖∇f‖ = 1.4674e-03, α = 1.00e+00, m = 1, nfg = 1
[ Info: LBFGS: iter 3: f = 0.000155638615, ‖∇f‖ = 7.1383e-04, α = 1.00e+00, m = 2, nfg = 1
[ Info: LBFGS: iter 4: f = 0.000154660387, ‖∇f‖ = 8.0494e-04, α = 1.00e+00, m = 3, nfg = 1
[ Info: LBFGS: iter 5: f = 0.000148931986, ‖∇f‖ = 2.0502e-03, α = 1.00e+00, m = 4, nfg = 1
[ Info: LBFGS: iter 6: f = 0.000137419061, ‖∇f‖ = 3.7131e-03, α = 1.00e+00, m = 5, nfg = 1
[ Info: LBFGS: iter 7: f = 0.000099576772, ‖∇f‖ = 5.6274e-03, α = 1.00e+00, m = 6, nfg = 1
[ Info: LBFGS: iter 8: f = 0.000014151556, ‖∇f‖ = 4.3621e-03, α = 5.72e-01, m = 7, nfg = 3
[ Info: LBFGS: iter 9: f = 0.000010465556, ‖∇f‖ = 1.5371e-03, α = 4.85e-01, m = 8, nfg = 2
[ Info: LBFGS: iter 10: f = 0.000008838465, ‖∇f‖ = 1.6771e-03, α = 1.00e+00, m = 8, nfg = 1
[ Info: LBFGS: iter 11: f = 0.000002607997, ‖∇f‖ = 1.8504e-03, α = 1.00e+00, m = 8, nfg = 1
[ Info: LBFGS: iter 12: f = 0.000000235906, ‖∇f‖ = 2.6223e-05, α = 1.00e+00, m = 8, nfg = 1
[ Info: LBFGS: iter 13: f = 0.000000007655, ‖∇f‖ = 7.5840e-05, α = 1.00e+00, m = 8, nfg = 1
[ Info: LBFGS: iter 14: f = 0.000000000038, ‖∇f‖ = 7.0155e-06, α = 1.00e+00, m = 8, nfg = 1
[ Info: LBFGS: iter 15: f = 0.000000000000, ‖∇f‖ = 7.6904e-08, α = 1.00e+00, m = 8, nfg = 1
[ Info: LBFGS: converged after 16 iterations: f = 0.000000000000, ‖∇f‖ = 7.0827e-09Hope this helps, and let us know if you still have issues with this!
ghost
mtfishman
I am trying to maximise the fidelity of two states, and when I try to use the optimiser as used in your VQE example, I get an error seemingly originating from Zygote via the loss' call. Any suggestions for why this is happening?
Code:
Full Error: