something is wrong while using the QSR

[CyanTurquoise]

Hi, when I am trying to find the ground state energy of a Hamiltonian which contains complex numbers, sometimes when I run the QSR it will give a warning report says that the imaginary part is been discarded, and the results goes to wrong. Exactly, the QSR I run is the following

model = nk.models.RBM(
    alpha=2,  # Hidden units = alpha * N (e.g., 32 for N=16)
    use_hidden_bias=True,  # Include hidden bias (default)
    activation=jnp.tanh,  # Nonlinear activation (default tanh)
    hidden_bias_init=nn.zeros,  # Hidden bias initialization
)
sa = nk.sampler.MetropolisLocal(hilbert, n_chains_per_rank=10) 
op = nk.optimizer.RmsProp(learning_rate=lr,beta=0.9,epscut = 1.0e-6)
vstate = nk.vqs.MCState(
    sampler=sa,
    model=model,
    n_samples=2000,  # Samples for estimates
    n_discard_per_chain=100,  # Burn-in discards
    seed=1234,# Reproducibility
)
qsr = nkx.QSR(
    training_data=(train_samples, rotations),
    training_batch_size=2000,
    optimizer=op,
    variational_state=vstate,
)

The QSR gives the result to be -2.78912663, while the direct compute result is -19.0387950 since the graph is a six atoms line.
It's based on the former work of PRXQuantum.2.040201, while in their code, they set the QSR as (the nk 2.0 or older) :

# RBM wavefunction
ma = nk.machine.RbmSpin(hilbert=hilbert, alpha=1)        # Complex-weights RBM, alpha=num_hidden/N
ma.init_random_parameters(seed=12345, sigma=0.01)
# Sampler
sa = nk.sampler.MetropolisLocal(machine=ma)

# Optimizer + Stochastic Reconfiguration
op = nk.optimizer.RmsProp(learning_rate=lr,beta=0.9,epscut = 1.0e-6)
sr = nk.optimizer.SR(diag_shift=0.1)
# Quantum state tomography object
qst = nk.Qsr(
    sampler=sa,
    optimizer=op,
    samples=training_samples,
    rotations=rotations,
    bases=training_bases,
    n_samples=M,
    n_samples_data=bs,
    sr=sr,
)

it seems that nk.RbmSpin and sr works in this Qsr, but I can't find RbmSpin in documents and QSR also dosen't have parameter sr in version 3.19. Is it any other method can address this problem. Hopes for help

[PhilipVinc]

The tutorial was done for an outdated version of netket at the time.
I seem to recall that in netket 2 rbmspin is similar to the current RBM(..., param_dtype=complex).

Sr can be used by specifying qsr.run(preconditioner=nk.optimizer.SR(diag_shift=...))