def algorithmic_gradKL(nn_state, psi_dict, vis, unitary_dict, bases):
grad_KL = [
torch.zeros(nn_state.rbm_am.num_pars,
dtype=torch.double,
device=nn_state.device),
torch.zeros(nn_state.rbm_ph.num_pars,
dtype=torch.double,
device=nn_state.device)
]
Z = partition(nn_state, vis).to(device=nn_state.device)
for i in range(len(vis)):
grad_KL[0] += (cplx.norm(psi_dict[bases[0]][:, i]) *
nn_state.rbm_am.effective_energy_gradient(vis[i]) /
float(len(bases)))
grad_KL[0] -= (probability(nn_state, vis[i], Z) *
nn_state.rbm_am.effective_energy_gradient(vis[i]) /
float(len(bases)))
for b in range(1, len(bases)):
for i in range(len(vis)):
rotated_grad = nn_state.gradient(bases[b], vis[i])
grad_KL[0] += (cplx.norm(psi_dict[bases[b]][:, i]) *
rotated_grad[0] / float(len(bases)))
grad_KL[1] += (cplx.norm(psi_dict[bases[b]][:, i]) *
rotated_grad[1] / float(len(bases)))
grad_KL[0] -= (probability(nn_state, vis[i], Z) *
nn_state.rbm_am.effective_energy_gradient(vis[i]) /
float(len(bases)))
return grad_KL
def compute_numerical_kl(nn_state, psi_dict, vis, Z, unitary_dict, bases):
N = nn_state.num_visible
psi_r = torch.zeros(2, 1 << N, dtype=torch.double)
KL = 0.0
for i in range(len(vis)):
KL += (cplx.norm(psi_dict[bases[0]][:, i]) *
cplx.norm(psi_dict[bases[0]][:, i]).log() / float(len(bases)))
KL -= (cplx.norm(psi_dict[bases[0]][:, i]) *
probability(nn_state, vis[i], Z).log().item() /
float(len(bases)))
for b in range(1, len(bases)):
psi_r = rotate_psi(nn_state, bases[b], unitary_dict, vis)
for ii in range(len(vis)):
if (cplx.norm(psi_dict[bases[b]][:, ii]) > 0.0):
KL += (cplx.norm(psi_dict[bases[b]][:, ii]) *
cplx.norm(psi_dict[bases[b]][:, ii]).log() /
float(len(bases)))
KL -= (cplx.norm(psi_dict[bases[b]][:, ii]) *
cplx.norm(psi_r[:, ii]).log() / float(len(bases)))
KL += (cplx.norm(psi_dict[bases[b]][:, ii]) * Z.log() /
float(len(bases)))
return KL
def fidelity(nn_state, target_psi, bases=None):
nn_state.compute_normalization()
F = torch.tensor([0., 0.], dtype=torch.double, device=nn_state.device)
target_psi = target_psi.to(nn_state.device)
for i in range(len(nn_state.space)):
psi = nn_state.psi(nn_state.space[i]) / (nn_state.Z).sqrt()
F[0] += target_psi[0, i] * psi[0] + target_psi[1, i] * psi[1]
F[1] += target_psi[0, i] * psi[1] - target_psi[1, i] * psi[0]
return cplx.norm(F)
def compute_numerical_NLL(nn_state, data_samples, data_bases, Z, unitary_dict,
vis):
NLL = 0
batch_size = len(data_samples)
b_flag = 0
for i in range(batch_size):
bitstate = []
for j in range(nn_state.num_visible):
ind = 0
if (data_bases[i][j] != 'Z'):
b_flag = 1
bitstate.append(int(data_samples[i, j].item()))
ind = int("".join(str(i) for i in bitstate), 2)
if (b_flag == 0):
NLL -= ((probability(nn_state, data_samples[i], Z)).log().item() /
batch_size)
else:
psi_r = rotate_psi(nn_state, data_bases[i], unitary_dict, vis)
NLL -= (cplx.norm(psi_r[:, ind]).log() -
Z.log()).item() / batch_size
return NLL
def KL(nn_state, target_psi, bases=None):
psi_r = torch.zeros(2,
1 << nn_state.num_visible,
dtype=torch.double,
device=nn_state.device)
KL = 0.0
unitary_dict = unitaries.create_dict()
target_psi = target_psi.to(nn_state.device)
space = nn_state.generate_Hilbert_space(nn_state.num_visible)
nn_state.compute_normalization()
if bases is None:
num_bases = 1
for i in range(len(space)):
KL += cplx.norm(target_psi[:, i]) * cplx.norm(target_psi[:,
i]).log()
KL -= cplx.norm(target_psi[:, i]) * cplx.norm(
nn_state.psi(space[i])).log()
KL += cplx.norm(target_psi[:, i]) * nn_state.Z.log()
else:
num_bases = len(bases)
for b in range(1, len(bases)):
psi_r = rotate_psi(nn_state, bases[b], unitary_dict)
target_psi_r = rotate_psi(nn_state, bases[b], unitary_dict,
target_psi)
for ii in range(len(space)):
if (cplx.norm(target_psi_r[:, ii]) > 0.0):
KL += cplx.norm(target_psi_r[:, ii]) * cplx.norm(
target_psi_r[:, ii]).log()
KL -= cplx.norm(target_psi_r[:, ii]) * cplx.norm(
psi_r[:, ii]).log().item()
KL += cplx.norm(target_psi_r[:, ii]) * nn_state.Z.log()
return KL / float(num_bases)
def test_norm(self):
scalar = torch.tensor([3, 4], dtype=torch.double)
expect = torch.tensor(5, dtype=torch.double)
self.assertTensorsEqual(cplx.norm(scalar), expect, msg="Norm failed!")