def get_extra_configs(self, configs):
""" Returns an nstep length array of configurations
starting from self._extra_config """
nconf = configs.configs.shape[0]
aux_configs_a = []
aux_configs_b = []
for step in range(self._nsweeps):
aux_configs_a.append(np.copy(self._aux_configs_a))
accept_a, self._aux_configs_a = sample_onebody(
self._mol,
self._orb_coeff[self._spin_sector[0]],
self._aux_configs_a,
tstep=self._tstep,
)
aux_configs_b.append(np.copy(self._aux_configs_b))
accept_b, self._aux_configs_b = sample_onebody(
self._mol,
self._orb_coeff[self._spin_sector[1]],
self._aux_configs_b,
tstep=self._tstep,
)
aux_configs_a = np.array(aux_configs_a)
aux_configs_b = np.array(aux_configs_b)
# Generates random choice of aux_config_a and aux_config_b for moving electron_a and electron_b
naux_a = self._aux_configs_a.shape[0]
naux_b = self._aux_configs_b.shape[0]
auxassignments_a = np.random.randint(0, naux_a, size=(self._nsweeps, nconf))
auxassignments_b = np.random.randint(0, naux_b, size=(self._nsweeps, nconf))
return [aux_configs_a, aux_configs_b], [auxassignments_a, auxassignments_b]
def __init__(
self,
mol,
orb_coeff,
spin,
nsweeps=4,
tstep=0.50,
warmup=200,
naux=500,
ijkl=None,
):
assert (
len(orb_coeff.shape) == 3
), "orb_coeff should be a list of orbital coefficients with size (2,num_mobasis,num_orb)."
self._mol = mol
self._orb_coeff = orb_coeff
self._tstep = tstep
self._nsweeps = nsweeps
self._spin = spin
self._spin_sector = spin
self._electrons_a = np.arange(spin[0] * mol.nelec[0],
mol.nelec[0] + spin[0] * mol.nelec[1])
self._electrons_b = np.arange(spin[1] * mol.nelec[0],
mol.nelec[0] + spin[1] * mol.nelec[1])
self._pairs = np.array(
np.meshgrid(self._electrons_a,
self._electrons_b)).T.reshape(-1, 2)
self._pairs = self._pairs[
self._pairs[:, 0] !=
self._pairs[:, 1]] # Removes repeated electron pairs
# Initialization and warmup of aux_configs_a
self._aux_configs_a = initial_guess(
mol, int(naux / sum(self._mol.nelec))).configs.reshape(-1, 3)
for i in range(warmup):
accept_a, self._aux_configs_a = sample_onebody(
mol, orb_coeff[self._spin_sector[0]], self._aux_configs_a,
tstep)
# Initialization and warmup of aux_configs_b
self._aux_configs_b = initial_guess(
mol, int(naux / sum(self._mol.nelec))).configs.reshape(-1, 3)
for i in range(warmup):
accept_b, self._aux_configs_b = sample_onebody(
mol, orb_coeff[self._spin_sector[1]], self._aux_configs_b,
tstep)
# Default to full 2rdm if ijkl not specified
if ijkl is None:
norb_up = orb_coeff[0].shape[1]
norb_down = orb_coeff[1].shape[1]
ijkl = [[i, j, k, l] for i in range(norb_up)
for j in range(norb_up) for k in range(norb_down)
for l in range(norb_down)]
self._ijkl = np.array(ijkl).T
def warm_up(self, naux):
# Initialization and warmup of configurations
nwalkers = int(naux / sum(self._mol.nelec)) + 1
self._aux_configs = []
for spin in [0, 1]:
self._aux_configs.append(mc.initial_guess(self._mol, nwalkers))
self._aux_configs[spin].reshape((-1, 1, 3))
self._aux_configs[spin].resample(range(naux))
_, self._aux_configs[spin], _ = obdm.sample_onebody(
self._aux_configs[spin], self.orbitals, 0, nsamples=self._warmup
)
self._aux_configs[spin] = self._aux_configs[spin][-1]
def __init__(
self,
mol,
orb_coeff,
spin,
nsweeps=4,
tstep=0.50,
warmup=200,
naux=500,
ijkl=None,
kpts=None,
):
assert (
len(orb_coeff.shape) == 3
), "orb_coeff should be a list of orbital coefficients with size (2,num_mobasis,num_orb)."
self._mol = mol
self._tstep = tstep
self._nsweeps = nsweeps
self._spin = spin
if kpts is None:
self.orbitals = MoleculeOrbitalEvaluator(mol, orb_coeff)
else:
if not hasattr(mol, "original_cell"):
mol = supercell.get_supercell(mol, np.eye(3))
self.orbitals = PBCOrbitalEvaluatorKpoints(mol, orb_coeff, kpts)
self._spin_sector = spin
self._electrons = [
np.arange(spin[s] * mol.nelec[0],
mol.nelec[0] + spin[s] * mol.nelec[1]) for s in [0, 1]
]
# Initialization and warmup of configurations
nwalkers = int(naux / sum(self._mol.nelec))
self._aux_configs = []
for spin in [0, 1]:
self._aux_configs.append(initial_guess(mol, nwalkers))
self._aux_configs[spin].reshape((-1, 1, 3))
_, self._aux_configs[spin], _ = sample_onebody(
self._aux_configs[spin], self.orbitals, 0, nsamples=warmup)
self._aux_configs[spin] = self._aux_configs[spin][-1]
# Default to full 2rdm if ijkl not specified
if ijkl is None:
norb_up = orb_coeff[0].shape[1]
norb_down = orb_coeff[1].shape[1]
ijkl = [[i, j, k, l] for i in range(norb_up)
for j in range(norb_up) for k in range(norb_down)
for l in range(norb_down)]
self._ijkl = np.array(ijkl).T
def get_configurations(self, nconf):
"""
Obtain a sequence of auxilliary configurations. This function returns one auxilliary configuration
for each nconf.
Changes internal state: self._aux_configs is updated to the last sampled location.
This will resample the auxilliary configurations to match the number of walkers.
returns a dictionary with the following elements, separated by spin:
assignments: [nsweeps, nconf]: assignment of configurations for each sweep to an auxilliary walker.
orbs: [nsweeps, conf, norb]: orbital values
configs: [nsweeps] Configuration object with nconf configurations of 1 electron
acceptance: [nsweeps, naux] acceptance probability for each auxilliary walker
TODO: Should we just resize the configurations to nconf instead of taking naux as an input?
"""
configs = []
assignments = []
orbs = []
acceptance = []
for spin in [0, 1]:
naux = self._aux_configs[spin].configs.shape[0]
accept, tmp_config, tmp_orbs = sample_onebody(
self._aux_configs[spin],
self.orbitals,
spin,
self._nsweeps,
tstep=self._tstep,
)
assignments.append(
np.random.randint(0, naux, size=(self._nsweeps, nconf)))
self._aux_configs[spin] = tmp_config[-1].copy()
acceptance.append(accept)
for conf, assign in zip(tmp_config, assignments[-1]):
conf.resample(assign)
configs.append(tmp_config)
orbs.append([
orb[assign, ...]
for orb, assign in zip(tmp_orbs, assignments[-1])
])
return {
"acceptance": acceptance,
"orbs": orbs,
"configs": configs,
"assignments": assignments,
}
def __call__(self, configs, wf, extra_configs=None, auxassignments=None):
"""Gathers quantities from equation (10) of DOI:10.1063/1.4793531."""
# Constructs results dictionary
nconf = configs.configs.shape[0]
results = {}
orb_a_size = self._orb_coeff[self._spin_sector[0]].shape[1]
orb_b_size = self._orb_coeff[self._spin_sector[1]].shape[1]
results["value"] = np.zeros((nconf, self._ijkl.shape[1]))
for i, e in enumerate(["a", "b"]):
results["norm_%s" % e] = np.zeros(
(nconf, self._orb_coeff[self._spin_sector[i]].shape[1])
)
results["acceptance_%s" % e] = np.zeros(nconf)
# Returns empty arrays if no electron pairs
if len(self._pairs) == 0:
return results
if extra_configs is None:
# Generates aux_configs_a and aux_configs_b
aux_configs_a = []
aux_configs_b = []
for step in range(self._nsweeps):
aux_configs_a.append(np.copy(self._aux_configs_a))
accept_a, self._aux_configs_a = sample_onebody(
self._mol,
self._orb_coeff[self._spin_sector[0]],
self._aux_configs_a,
tstep=self._tstep,
)
aux_configs_b.append(np.copy(self._aux_configs_b))
accept_b, self._aux_configs_b = sample_onebody(
self._mol,
self._orb_coeff[self._spin_sector[1]],
self._aux_configs_b,
tstep=self._tstep,
)
results["acceptance_a"] += np.mean(accept_a)
results["acceptance_b"] += np.mean(accept_b)
results["acceptance_a"] /= self._nsweeps
results["acceptance_b"] /= self._nsweeps
aux_configs_a = np.array(aux_configs_a)
aux_configs_b = np.array(aux_configs_b)
# Generates random choice of aux_config_a and aux_config_b for moving electron_a and electron_b
naux_a = self._aux_configs_a.shape[0]
naux_b = self._aux_configs_b.shape[0]
auxassignments_a = np.random.randint(0, naux_a, size=(self._nsweeps, nconf))
auxassignments_b = np.random.randint(0, naux_b, size=(self._nsweeps, nconf))
else:
assert auxassignments is not None
aux_configs_a = extra_configs[0]
aux_configs_b = extra_configs[1]
naux_a = self._aux_configs_a.shape[0]
naux_b = self._aux_configs_b.shape[0]
auxassignments_a = auxassignments[0]
auxassignments_b = auxassignments[1]
# Evaluate VMC configurations
coords = configs.configs.reshape(
(configs.configs.shape[0] * configs.configs.shape[1], -1)
)
ao_configs = self._mol.eval_gto("GTOval_sph", coords)
orb_a_configs = ao_configs.dot(self._orb_coeff[self._spin_sector[0]]).reshape(
(configs.configs.shape[0], configs.configs.shape[1], -1)
)
orb_b_configs = ao_configs.dot(self._orb_coeff[self._spin_sector[1]]).reshape(
(configs.configs.shape[0], configs.configs.shape[1], -1)
)
orb_a_configs = orb_a_configs[:, self._pairs[:, 0], :]
orb_b_configs = orb_b_configs[:, self._pairs[:, 1], :]
# Sweeps over electron pairs
for sweep in range(self._nsweeps):
ao_a_aux = self._mol.eval_gto("GTOval_sph", aux_configs_a[sweep])
ao_b_aux = self._mol.eval_gto("GTOval_sph", aux_configs_b[sweep])
orb_a_aux = ao_a_aux.dot(self._orb_coeff[self._spin_sector[0]])
orb_b_aux = ao_b_aux.dot(self._orb_coeff[self._spin_sector[1]])
fsum_a = np.sum(orb_a_aux * orb_a_aux, axis=1)
fsum_b = np.sum(orb_b_aux * orb_b_aux, axis=1)
norm_a = orb_a_aux * orb_a_aux / fsum_a[:, np.newaxis]
norm_b = orb_b_aux * orb_b_aux / fsum_b[:, np.newaxis]
# We use pySCF's index convention (while Eq. 10 in DOI:10.1063/1.4793531 uses QWalk's)
# QWalk -> tbdm[s1,s2,i,j,k,l] = < c^+_{s1,i} c^+_{s2,j} c_{s2,l} c_{s1,k} > = \phi*_{s1,k} \phi*_{s2,l} \phi_{s2,j} \phi_{s1,i}
# pySCF -> tbdm[s1,s2,i,j,k,l] = < c^+_{s1,i} c^+_{s2,k} c_{s2,l} c_{s1,j} > = \phi*_{s1,j} \phi*_{s2,l} \phi_{s2,k} \phi_{s1,i}
orbratio = (
(
orb_a_aux[auxassignments_a[sweep]][:, self._ijkl[1]]
/ fsum_a[auxassignments_a[sweep], np.newaxis]
)[:, np.newaxis, :]
* (
orb_b_aux[auxassignments_b[sweep]][:, self._ijkl[3]]
/ fsum_b[auxassignments_b[sweep], np.newaxis]
)[:, np.newaxis, :]
* orb_a_configs[..., self._ijkl[0]]
* orb_b_configs[..., self._ijkl[2]]
)
# Calculation of wf ratio (no McMillan trick yet)
epos_a = configs.make_irreducible(
-1, aux_configs_a[sweep][auxassignments_a[sweep]]
)
epos_b = configs.make_irreducible(
-1, aux_configs_b[sweep][auxassignments_b[sweep]]
)
wfratio = []
for ea in self._electrons_a:
electrons_b = self._electrons_b[self._electrons_b != ea]
wfratio_a = wf.testvalue(ea, epos_a)
wf.updateinternals(ea, epos_a)
wfratio_b = wf.testvalue_many(electrons_b, epos_b)
wf.updateinternals(ea, configs.electron(ea))
wfratio.append(wfratio_a[:, np.newaxis] * wfratio_b)
wfratio = np.concatenate(wfratio, axis=1)
# Adding to results
results["value"] += np.einsum("in,inj->ij", wfratio, orbratio)
results["norm_a"] += norm_a[auxassignments_a[sweep]]
results["norm_b"] += norm_b[auxassignments_b[sweep]]
# Average over sweeps and pairs
results["value"] /= self._nsweeps
for e in ["a", "b"]:
results["norm_%s" % e] /= self._nsweeps
return results
