def test_dmrg_shared_exec_iterative_solve_one_sweep():
max_mps_rank = 5
num = 7
mpo_rank = 5
size = 5
num_iter = 10
T.set_backend("numpy")
h = qtn.MPO_rand_herm(num, mpo_rank, size)
dmrg_quimb = qtn.DMRG2(h, bond_dims=[max_mps_rank])
h_tensors = load_quimb_tensors(h)
mps_tensors = load_quimb_tensors(dmrg_quimb.state)
# dmrg based on ad
mps_tensors, energy = dmrg_shared_exec_iterative_solve(
h_tensors, mps_tensors, num_iter=num_iter, max_mps_rank=max_mps_rank)
# dmrg based on quimb
opts = {'max_bond': max_mps_rank}
for _ in range(num_iter):
quimb_energy = dmrg_quimb.sweep_right(canonize=True,
verbosity=0,
**opts)
# We only test on energy (lowest eigenvalue of h), rather than the output
# mps (eigenvector), because the eigenvectors can vary a lot while keeping the
# eigenvalue unchanged.
assert (abs(energy - quimb_energy) < 1e-5)
def rand_tn1d_sect(n, bd, dtype=complex):
mps = qtn.MPS_rand_state(n + 2, bd, dtype=dtype)
mpo = qtn.MPO_rand_herm(n + 2, 5, dtype=dtype)
norm = qtn.TensorNetwork(qtn.align_TN_1D(mps.H, mpo, mps))
lix = qtn.bonds(norm[0], norm[1])
rix = qtn.bonds(norm[n], norm[n + 1])
to = norm[1:n + 1]
return qtn.TNLinearOperator1D(to, lix, rix, 1, n + 1)
def rand_tn1d_sect(n, bd, dtype=complex):
mps = qtn.MPS_rand_state(n + 2, bd, dtype=dtype)
mpo = qtn.MPO_rand_herm(n + 2, 5, dtype=dtype)
norm = qtn.TensorNetwork(qtn.align_TN_1D(mps.H, mpo, mps))
# greedy not good and contracting with large bsz
norm.structure_bsz = 2
lix = qtn.bonds(norm[0], norm[1])
rix = qtn.bonds(norm[n], norm[n + 1])
to = norm[1:n + 1]
return qtn.TNLinearOperator1D(to, lix, rix, 1, n + 1)