def parse_composition(structure_type, s, ctype):
toks = s.strip().split()
if len(toks) == 1:
c = Composition({toks[0].split(":")[0]: 1})
else:
c = Composition(
{t.split(":")[0]: float(t.split(":")[1])
for t in toks})
c = Composition({k2: v2 / sum(c.values()) for k2, v2 in c.items()})
if len(c) != 2:
raise ValueError("Bad composition on %s." % ctype)
frac = [c.get_atomic_fraction(k) for k in c.keys()]
if structure_type == 'garnet':
if ctype == "A":
if abs(frac[0] - 0.5) > 0.01:
raise ValueError("Bad composition on %s. "
"Only 1:1 mixing allowed!" % ctype)
elif ctype in ["C", "D"]:
if not (abs(frac[0] - 1.0 / 3) < 0.01
or abs(frac[1] - 1.0 / 3) < 0.01):
raise ValueError("Bad composition on %s. "
"Only 2:1 mixing allowed!" % ctype)
elif structure_type == 'perovskite':
if abs(frac[0] - 0.5) > 0.01:
raise ValueError("Bad composition on %s. "
"Only 1:1 mixing allowed!" % ctype)
try:
for k in c.keys():
k.oxi_state
if k not in ELS[structure_type][ctype]:
raise ValueError("%s is not a valid species for %s site." %
(k, ctype))
except AttributeError:
raise ValueError("Oxidation states must be specified for all species!")
return c
def __getitem__(self, idx):
"""
Returns
-------
atom_weights: torch.Tensor shape (M, 1)
weights of atoms in the material
atom_fea: torch.Tensor shape (M, n_fea)
features of atoms in the material
self_fea_idx: torch.Tensor shape (M*M, 1)
list of self indices
nbr_fea_idx: torch.Tensor shape (M*M, 1)
list of neighbor indices
target: torch.Tensor shape (1,)
target value for material
cry_id: torch.Tensor shape (1,)
input id for the material
"""
cry_id, composition, target = self.df.iloc[idx]
comp_dict = Composition(composition).get_el_amt_dict()
elements = list(comp_dict.keys())
weights = list(comp_dict.values())
weights = np.atleast_2d(weights).T / np.sum(weights)
assert len(elements) != 1, f"cry-id {cry_id} [{composition}] is a pure system"
try:
atom_fea = np.vstack(
[self.elem_features.get_fea(element) for element in elements]
)
except AssertionError:
raise AssertionError(
f"cry-id {cry_id} [{composition}] contains element types not in embedding"
)
except ValueError:
raise ValueError(
f"cry-id {cry_id} [{composition}] composition cannot be parsed into elements"
)
env_idx = list(range(len(elements)))
self_fea_idx = []
nbr_fea_idx = []
nbrs = len(elements) - 1
for i, _ in enumerate(elements):
self_fea_idx += [i] * nbrs
nbr_fea_idx += env_idx[:i] + env_idx[i + 1 :]
# convert all data to tensors
atom_weights = torch.Tensor(weights)
atom_fea = torch.Tensor(atom_fea)
self_fea_idx = torch.LongTensor(self_fea_idx)
nbr_fea_idx = torch.LongTensor(nbr_fea_idx)
if self.task == "regression":
targets = torch.Tensor([float(target)])
elif self.task == "classification":
targets = torch.LongTensor([target])
return (
(atom_weights, atom_fea, self_fea_idx, nbr_fea_idx),
targets,
composition,
cry_id,
)
