def test_positions(self):
"""Tests that different positions are handled correctly.
"""
desc = SOAP([1, 6, 8], 10.0, 2, 0, periodic=False, crossover=True)
self.assertEqual(desc.get_number_of_features(), desc.create(H2O, positions=[[0, 0, 0]]).shape[1])
self.assertEqual(desc.get_number_of_features(), desc.create(H2O, positions=[0]).shape[1])
self.assertEqual(desc.get_number_of_features(), desc.create(H2O).shape[1])
desc = SOAP([1, 6, 8], 10.0, 2, 0, periodic=True, crossover=True,)
self.assertEqual(desc.get_number_of_features(), desc.create(H2O, positions=[[0, 0, 0]]).shape[1])
self.assertEqual(desc.get_number_of_features(), desc.create(H2O, positions=[0]).shape[1])
self.assertEqual(desc.get_number_of_features(), desc.create(H2O).shape[1])
desc = SOAP([1, 6, 8], 10.0, 2, 0, periodic=True, crossover=False,)
self.assertEqual(desc.get_number_of_features(), desc.create(H2O, positions=[[0, 0, 0]]).shape[1])
self.assertEqual(desc.get_number_of_features(), desc.create(H2O, positions=[0]).shape[1])
self.assertEqual(desc.get_number_of_features(), desc.create(H2O).shape[1])
desc = SOAP([1, 6, 8], 10.0, 2, 0, periodic=False, crossover=False,)
self.assertEqual(desc.get_number_of_features(), desc.create(H2O, positions=[[0, 0, 0]]).shape[1])
self.assertEqual(desc.get_number_of_features(), desc.create(H2O, positions=[0]).shape[1])
self.assertEqual(desc.get_number_of_features(), desc.create(H2O).shape[1])
with self.assertRaises(ValueError):
desc.create(H2O, positions=['a'])
def createDescriptorsAllSOAP(data,
species,
sigma_SOAP,
cutoff_SOAP,
nmax_SOAP,
lmax_SOAP,
periodic,
sparse_SOAP=default_sparse_SOAP):
# Initialize SOAP
soap = SOAP(species=species,
sigma=sigma_SOAP,
periodic=periodic,
rcut=cutoff_SOAP,
nmax=nmax_SOAP,
lmax=lmax_SOAP,
sparse=sparse_SOAP)
# Compute number of features
n_features = soap.get_number_of_features()
n_atoms = np.shape(data[0])[0]
n_steps = len(data)
# Computing descriptors
descriptors = np.empty((n_atoms, n_steps, n_features), dtype=object)
for index_structure in tqdm.tqdm(range(n_steps)):
descriptors[:, index_structure, :] = soap.create(data[index_structure])
descriptors_ = []
for atom in range(n_atoms):
descriptors_.append(descriptors[atom, :, :])
return descriptors_
def test_number_of_features(self):
"""Tests that the reported number of features is correct."""
lmax = 5
nmax = 5
n_elems = 2
desc = SOAP(species=[1, 8], rcut=3, nmax=nmax, lmax=lmax, periodic=True)
# Test that the reported number of features matches the expected
n_features = desc.get_number_of_features()
expected = int((lmax + 1) * (nmax * n_elems) * (nmax * n_elems + 1) / 2)
self.assertEqual(n_features, expected)
# Test that the outputted number of features matches the reported
n_features = desc.get_number_of_features()
vec = desc.create(H2O)
self.assertEqual(n_features, vec.shape[1])
def test_properties(self):
"""Used to test that changing the setup through properties works as
intended.
"""
# Test changing species
a = SOAP(
species=[1, 8],
rcut=3,
nmax=3,
lmax=3,
sparse=False,
)
nfeat1 = a.get_number_of_features()
vec1 = a.create(H2O)
a.species = ["C", "H", "O"]
nfeat2 = a.get_number_of_features()
vec2 = a.create(molecule("CH3OH"))
self.assertTrue(nfeat1 != nfeat2)
self.assertTrue(vec1.shape[1] != vec2.shape[1])
def test_positions(self):
"""Tests that different positions are handled correctly.
"""
desc = SOAP(species=[1, 6, 8], rcut=10.0, nmax=2, lmax=0, periodic=False, crossover=True)
n_feat = desc.get_number_of_features()
self.assertEqual((1, n_feat), desc.create(H2O, positions=np.array([[0, 0, 0]])).shape)
self.assertEqual((1, n_feat), desc.create(H2O, positions=[[0, 0, 0]]).shape)
self.assertEqual((3, n_feat), desc.create(H2O, positions=[0, 1, 2]).shape)
self.assertEqual((3, n_feat), desc.create(H2O, positions=np.array([0, 1, 2])).shape)
self.assertEqual((3, n_feat), desc.create(H2O).shape)
desc = SOAP(species=[1, 6, 8], rcut=10.0, nmax=2, lmax=0, periodic=True, crossover=True,)
n_feat = desc.get_number_of_features()
self.assertEqual((1, n_feat), desc.create(H2O, positions=np.array([[0, 0, 0]])).shape)
self.assertEqual((1, n_feat), desc.create(H2O, positions=[[0, 0, 0]]).shape)
self.assertEqual((3, n_feat), desc.create(H2O, positions=[0, 1, 2]).shape)
self.assertEqual((3, n_feat), desc.create(H2O, positions=np.array([0, 1, 2])).shape)
self.assertEqual((3, n_feat), desc.create(H2O).shape)
desc = SOAP(species=[1, 6, 8], rcut=10.0, nmax=2, lmax=0, periodic=True, crossover=False,)
n_feat = desc.get_number_of_features()
self.assertEqual((1, n_feat), desc.create(H2O, positions=np.array([[0, 0, 0]])).shape)
self.assertEqual((1, n_feat), desc.create(H2O, positions=[[0, 0, 0]]).shape)
self.assertEqual((3, n_feat), desc.create(H2O, positions=[0, 1, 2]).shape)
self.assertEqual((3, n_feat), desc.create(H2O, positions=np.array([0, 1, 2])).shape)
self.assertEqual((3, n_feat), desc.create(H2O).shape)
desc = SOAP(species=[1, 6, 8], rcut=10.0, nmax=2, lmax=0, periodic=False, crossover=False,)
n_feat = desc.get_number_of_features()
self.assertEqual((1, n_feat), desc.create(H2O, positions=np.array([[0, 0, 0]])).shape)
self.assertEqual((1, n_feat), desc.create(H2O, positions=[[0, 0, 0]]).shape)
self.assertEqual((3, n_feat), desc.create(H2O, positions=[0, 1, 2]).shape)
self.assertEqual((3, n_feat), desc.create(H2O, positions=np.array([0, 1, 2])).shape)
self.assertEqual((3, n_feat), desc.create(H2O).shape)
with self.assertRaises(ValueError):
desc.create(H2O, positions=['a'])
def backend(sn, soap_mask, tracer_atomic_number, environment_list):
soap = SOAP(
species = environment_list,
crossover = soap_opts['crossover'],
rcut = soap_opts['cutoff'],
nmax = soap_opts['n_max'],
lmax = soap_opts['l_max'],
rbf = soap_opts['rbf'],
sigma = soap_opts['atom_sigma'],
periodic = soap_opts['periodic'],
sparse = False
)
def dscribe_soap(structure, positions):
out = soap.create(structure, positions = positions).astype(np.float)
return out
dscribe_soap.n_dim = soap.get_number_of_features()
return dscribe_soap
class SOAPComputer(torch.nn.Module):
def __init__(self, **config):
super(SOAPComputer, self).__init__()
self.soap = SOAP(**config)
def __len__(self):
return self.soap.get_number_of_features()
def forward(self, species_coordinates):
species, coordinates = species_coordinates
if len(coordinates.shape) == 2:
coordinates = coordinates[None, :]
mol = Atoms(species, coordinates[0])
#mask = [n for n,s in enumerate(mol.get_chemical_symbols()) if s==self.target]
descriptors = []
for coords in coordinates:
mol.set_positions(coords)
descriptor = self.soap.create(mol) #, positions=mask)
descriptors.append(descriptor)
return torch.tensor(descriptors, dtype=torch.float64)
def test_parallel_sparse(self):
"""Tests creating sparse output parallelly.
"""
# Test indices
samples = [molecule("CO"), molecule("N2O")]
desc = SOAP(
species=[6, 7, 8],
rcut=5,
nmax=3,
lmax=3,
sigma=1,
periodic=False,
crossover=True,
average=False,
sparse=True,
)
n_features = desc.get_number_of_features()
# Multiple systems, serial job
output = desc.create(
system=samples,
positions=[[0], [0, 1]],
n_jobs=1,
).toarray()
assumed = np.empty((3, n_features))
assumed[0, :] = desc.create(samples[0], [0]).toarray()
assumed[1, :] = desc.create(samples[1], [0]).toarray()
assumed[2, :] = desc.create(samples[1], [1]).toarray()
self.assertTrue(np.allclose(output, assumed))
# Test when position given as indices
output = desc.create(
system=samples,
positions=[[0], [0, 1]],
n_jobs=2,
).toarray()
assumed = np.empty((3, n_features))
assumed[0, :] = desc.create(samples[0], [0]).toarray()
assumed[1, :] = desc.create(samples[1], [0]).toarray()
assumed[2, :] = desc.create(samples[1], [1]).toarray()
self.assertTrue(np.allclose(output, assumed))
# Test with no positions specified
output = desc.create(
system=samples,
positions=[None, None],
n_jobs=2,
).toarray()
assumed = np.empty((2+3, n_features))
assumed[0, :] = desc.create(samples[0], [0]).toarray()
assumed[1, :] = desc.create(samples[0], [1]).toarray()
assumed[2, :] = desc.create(samples[1], [0]).toarray()
assumed[3, :] = desc.create(samples[1], [1]).toarray()
assumed[4, :] = desc.create(samples[1], [2]).toarray()
self.assertTrue(np.allclose(output, assumed))
# Test with cartesian positions
output = desc.create(
system=samples,
positions=[[[0, 0, 0], [1, 2, 0]], [[1, 2, 0]]],
n_jobs=2,
).toarray()
assumed = np.empty((2+1, n_features))
assumed[0, :] = desc.create(samples[0], [[0, 0, 0]]).toarray()
assumed[1, :] = desc.create(samples[0], [[1, 2, 0]]).toarray()
assumed[2, :] = desc.create(samples[1], [[1, 2, 0]]).toarray()
self.assertTrue(np.allclose(output, assumed))
# Test averaged output
desc._average = True
output = desc.create(
system=samples,
positions=[[0], [0, 1]],
n_jobs=2,
).toarray()
assumed = np.empty((2, n_features))
assumed[0, :] = desc.create(samples[0], [0]).toarray()
assumed[1, :] = 1/2*(desc.create(samples[1], [0]).toarray() + desc.create(samples[1], [1]).toarray())
self.assertTrue(np.allclose(output, assumed))
def create_data_SOAP(data, metadata):
particles, scaler, test_size, rcut, nmax, lmax, N_PCA, sigma_SOAP = [
metadata[x] for x in [
'particles', 'scaler', 'test_size', 'rcut', 'nmax', 'lmax',
'N_PCA', 'sigma_SOAP'
]
]
soap = SOAP(
species=np.unique(particles),
sigma=sigma_SOAP,
periodic=False,
rcut=rcut,
nmax=nmax,
lmax=lmax,
sparse=False,
#rbf='polynomial'
)
nb_features = soap.get_number_of_features()
descriptors = pd.np.empty(
(data.index.max() + 1, len(particles), nb_features))
for i_time in tqdm.tqdm(range(data.index.max() + 1)):
descriptors[i_time] = soap.create(data['molec'][i_time],
positions=np.arange(len(particles)))
#create training set
try:
data['is_train']
except KeyError:
data['is_train'] = create_is_train(data.index.max() + 1)
else:
pass
#selecting best params
if N_PCA:
try:
metadata['PCAs']
except KeyError:
PCAs = select_best_params(descriptors[data['is_train'].values],
nb_features, N_PCA)
new_descriptors = pd.np.empty(
(data.index.max() + 1, len(particles), N_PCA))
new_descriptors[:, :2, :] = PCAs[0].transform(
descriptors[:, :2, :].reshape(
descriptors[:, :2, :].shape[0] * 2,
nb_features)).reshape(descriptors.shape[0], 2, N_PCA)
new_descriptors[:, 2:, :] = PCAs[1].transform(
descriptors[:, 2:, :].reshape(
descriptors[:, 2:, :].shape[0] * 5,
nb_features)).reshape(descriptors.shape[0], 5, N_PCA)
descriptors = new_descriptors
metadata['old_N_feature'] = nb_features
nb_features = N_PCA
metadata['PCAs'] = PCAs
else:
PCAs = metadata['PCAs']
new_descriptors = pd.np.empty(
(data.index.max() + 1, len(particles), N_PCA))
new_descriptors[:, :2, :] = PCAs[0].transform(
descriptors[:, :2, :].reshape(
descriptors[:, :2, :].shape[0] * 2,
nb_features)).reshape(descriptors.shape[0], 2, N_PCA)
new_descriptors[:, 2:, :] = PCAs[1].transform(
descriptors[:, 2:, :].reshape(
descriptors[:, 2:, :].shape[0] * 5,
nb_features)).reshape(descriptors.shape[0], 5, N_PCA)
descriptors = new_descriptors
nb_features = N_PCA
else:
pass
#scaling
if scaler == False:
pass
elif type(scaler) == type(None):
descriptors, scaler = scale_descriptors(data, descriptors)
else:
descriptors[:, 0:2, :] = scaler[0].transform(
descriptors[:, 0:2, :].reshape(descriptors[:, 0:2, :].shape[0] * 2,
nb_features)).reshape(
descriptors.shape[0], 2,
nb_features)
descriptors[:,
2:, :] = scaler[1].transform(descriptors[:, 2:, :].reshape(
descriptors[:, 2:, :].shape[0] * 5,
nb_features)).reshape(descriptors.shape[0], 5,
nb_features)
metadata['scaler'] = scaler
return data.join(pd.DataFrame({'descriptor': list(descriptors)})), metadata
# Create SOAP output for the system
soap_water = soap.create(water, positions=[0])
print(soap_water)
print(soap_water.shape)
# Create output for multiple system
samples = [molecule("H2O"), molecule("NO2"), molecule("CO2")]
positions = [[0], [1, 2], [1, 2]]
coulomb_matrices = soap.create(samples, positions) # Serial
coulomb_matrices = soap.create(samples, positions, n_jobs=2) # Parallel
# Lets change the SOAP setup and see how the number of features changes
small_soap = SOAP(species=species, rcut=rcut, nmax=2, lmax=0)
big_soap = SOAP(species=species, rcut=rcut, nmax=9, lmax=9)
n_feat1 = small_soap.get_number_of_features()
n_feat2 = big_soap.get_number_of_features()
print(n_feat1, n_feat2)
# Periodic systems
from ase.build import bulk
copper = bulk('Cu', 'fcc', a=3.6, cubic=True)
print(copper.get_pbc())
periodic_soap = SOAP(species=[29],
rcut=rcut,
nmax=nmax,
lmax=nmax,
periodic=True,
sparse=False)
def test_parallel_sparse(self):
"""Tests creating sparse output parallelly."""
# Test indices
samples = [molecule("CO"), molecule("NO")]
desc = SOAP(
species=[6, 7, 8],
rcut=5,
nmax=3,
lmax=3,
sigma=1,
periodic=False,
crossover=True,
average="off",
sparse=True,
)
n_features = desc.get_number_of_features()
# Multiple systems, serial job, fixed size
output = desc.create(
system=samples,
positions=[[0, 1], [0, 1]],
n_jobs=1,
).todense()
assumed = np.empty((2, 2, n_features))
assumed[0, 0] = desc.create(samples[0], [0]).todense()
assumed[0, 1] = desc.create(samples[0], [1]).todense()
assumed[1, 0] = desc.create(samples[1], [0]).todense()
assumed[1, 1] = desc.create(samples[1], [1]).todense()
self.assertTrue(np.allclose(output, assumed))
# Multiple systems, parallel job, fixed size
output = desc.create(
system=samples,
positions=[[0, 1], [0, 1]],
n_jobs=2,
).todense()
assumed = np.empty((2, 2, n_features))
assumed[0, 0] = desc.create(samples[0], [0]).todense()
assumed[0, 1] = desc.create(samples[0], [1]).todense()
assumed[1, 0] = desc.create(samples[1], [0]).todense()
assumed[1, 1] = desc.create(samples[1], [1]).todense()
self.assertTrue(np.allclose(output, assumed))
# Multiple systems, parallel job, all atoms, fixed size
output = desc.create(
system=samples,
positions=[None, None],
n_jobs=2,
).todense()
assumed = np.empty((2, 2, n_features))
assumed[0, 0] = desc.create(samples[0], [0]).todense()
assumed[0, 1] = desc.create(samples[0], [1]).todense()
assumed[1, 0] = desc.create(samples[1], [0]).todense()
assumed[1, 1] = desc.create(samples[1], [1]).todense()
self.assertTrue(np.allclose(output, assumed))
# Multiple systems, parallel job, cartesian positions, fixed size
output = desc.create(
system=samples,
positions=[[[0, 0, 0]], [[1, 2, 0]]],
n_jobs=2,
).todense()
assumed = np.empty((2, 1, n_features))
assumed[0, 0] = desc.create(samples[0], [[0, 0, 0]]).todense()
assumed[1, 0] = desc.create(samples[1], [[1, 2, 0]]).todense()
self.assertTrue(np.allclose(output, assumed))
# Multiple systems, parallel job, indices, variable size
output = desc.create(
system=samples,
positions=[[0], [0, 1]],
n_jobs=2,
)
self.assertTrue(
np.allclose(output[0][0].todense(), desc.create(samples[0], [0]).todense())
)
self.assertTrue(
np.allclose(output[1][0].todense(), desc.create(samples[1], [0]).todense())
)
self.assertTrue(
np.allclose(output[1][1].todense(), desc.create(samples[1], [1]).todense())
)
# Test averaged output
desc.average = "outer"
output = desc.create(
system=samples,
positions=[[0], [0, 1]],
n_jobs=2,
).todense()
assumed = np.empty((2, n_features))
assumed[0] = desc.create(samples[0], [0]).todense()
assumed[1] = (
1
/ 2
* (
desc.create(samples[1], [0]).todense()
+ desc.create(samples[1], [1]).todense()
)
)
self.assertTrue(np.allclose(output, assumed))
class SOAP(BaseFeaturizer):
"""
Smooth overlap of atomic positions (interface via DScribe).
Class for generating a partial power spectrum from Smooth Overlap of Atomic
Orbitals (SOAP). This implementation uses real (tesseral) spherical
harmonics as the angular basis set and provides two orthonormalized
alternatives for the radial basis functions: spherical primitive gaussian
type orbitals ("gto") or the polynomial basis set ("polynomial"). By
default the faster gto-basis is used. Please see the DScribe SOAP
documentation for more details.
Note that SOAP is only featurized for elements identified by "fit" (see
following), thus "fit" must be called before "featurize", or else an error
will be raised.
Based originally on the following publications:
"On representing chemical environments, Albert P. Bartók, Risi
Kondor, and Gábor Csányi, Phys. Rev. B 87, 184115, (2013),
https://doi.org/10.1103/PhysRevB.87.184115
"Comparing molecules and solids across structural and alchemical
space", Sandip De, Albert P. Bartók, Gábor Csányi and Michele Ceriotti,
Phys. Chem. Chem. Phys. 18, 13754 (2016),
https://doi.org/10.1039/c6cp00415f
Implementation (and some documentation) originally based on DScribe:
https://github.com/SINGROUP/dscribe.
"DScribe: Library of descriptors for machine learning in materials science",
Himanen, L., J{\"a}ger, M. O.J., Morooka, E. V., Federici
Canova, F., Ranawat, Y. S., Gao, D. Z., Rinke, P. and Foster, A. S.
Computer Physics Communications, 106949 (2019),
https://doi.org/10.1016/j.cpc.2019.106949
Args:
rcut (float): A cutoff for local region in angstroms. Should be
bigger than 1 angstrom.
nmax (int): The number of radial basis functions.
lmax (int): The maximum degree of spherical harmonics.
sigma (float): The standard deviation of the gaussians used to expand the
atomic density.
rbf (str): The radial basis functions to use. The available options are:
* "gto": Spherical gaussian type orbitals defined as :math:`g_{nl}(r) = \sum_{n'=1}^{n_\mathrm{max}}\,\\beta_{nn'l} r^l e^{-\\alpha_{n'l}r^2}`
* "polynomial": Polynomial basis defined as :math:`g_{n}(r) = \sum_{n'=1}^{n_\mathrm{max}}\,\\beta_{nn'} (r-r_\mathrm{cut})^{n'+2}`
periodic (bool): Determines whether the system is considered to be
periodic.
crossover (bool): Determines if crossover of atomic types should
be included in the power spectrum. If enabled, the power
spectrum is calculated over all unique species combinations Z
and Z'. If disabled, the power spectrum does not contain
cross-species information and is only run over each unique
species Z. Turned on by default to correspond to the original
definition
"""
@requires(
dscribe,
"SOAPFeaturizer requires DScribe. Install from github.com/SINGROUP/dscribe"
)
def __init__(
self,
rcut,
nmax,
lmax,
sigma,
periodic,
rbf="gto",
crossover=True,
):
self.rcut = rcut
self.nmax = nmax
self.lmax = lmax
self.sigma = sigma
self.rbf = rbf
self.periodic = periodic
self.crossover = crossover
self.adaptor = AseAtomsAdaptor()
self.length = None
self.atomic_numbers = None
self.soap = None
self.n_elements = None
def _check_fitted(self):
if not self.soap:
raise NotFittedError("Please fit SOAP before featurizing.")
def fit(self, X, y=None):
"""
Fit the SOAP featurizer to a dataframe.
Args:
X ([SiteCollection]): For example, a list of pymatgen Structures.
y : unused (added for consistency with overridden method signature)
Returns:
self
"""
# Check that pymatgen.Structures are provided
if not all([isinstance(struct, Structure) for struct in X]):
raise TypeError(
"This fit requires an array-like input of Pymatgen "
"Structures and sites!")
elements = set()
for s in X:
c = s.composition.elements
for e in c:
if e.Z not in elements:
elements.add(e.Z)
self.elements_sorted = sorted(list(elements))
self.atomic_numbers = elements
self.soap = SOAP_dscribe(species=self.atomic_numbers,
rcut=self.rcut,
nmax=self.nmax,
lmax=self.lmax,
sigma=self.sigma,
rbf=self.rbf,
periodic=self.periodic,
crossover=self.crossover,
average=False,
sparse=False)
self.length = self.soap.get_number_of_features()
return self
def featurize(self, struct, idx):
self._check_fitted()
s_ase = self.adaptor.get_atoms(struct)
return self.soap.create(s_ase, positions=[idx]).tolist()[0]
def feature_labels(self):
self._check_fitted()
labels = []
for zi in self.elements_sorted:
for zj in self.elements_sorted:
for l in range(self.lmax + 1):
for ni in range(self.nmax):
for nj in range(self.nmax):
if nj >= ni and zj >= zi:
labels.append(
"Z={},Z'={},l={},n={},n'={}".format(
zi, zj, l, ni, nj))
return labels
def citations(self):
return [
"@article{PhysRevB.87.184115,"
"title = {On representing chemical environments},"
"author = {Bart\'ok, Albert P. and Kondor, Risi and Cs\'anyi, "
"G\'abor},"
"journal = {Phys. Rev. B},"
"volume = {87},"
"issue = {18},"
"pages = {184115},"
"numpages = {16},"
"year = {2013},"
"month = {May},"
"publisher = {American Physical Society},"
"doi = {10.1103/PhysRevB.87.184115},"
"url = {https://link.aps.org/doi/10.1103/PhysRevB.87.184115}}",
"@Article{C6CP00415F,"
"author ={De, Sandip and Bartók, Albert P. and Csányi, Gábor"
" and Ceriotti, Michele},"
"title ={Comparing molecules and solids across structural and "
"alchemical space},"
"journal = {Phys. Chem. Chem. Phys.},"
"year = {2016},"
"volume = {18},"
"issue = {20},"
"pages = {13754-13769},"
"publisher = {The Royal Society of Chemistry},"
"doi = {10.1039/C6CP00415F},"
"url = {http://dx.doi.org/10.1039/C6CP00415F},}",
'@article{dscribe, '
'author = {Himanen, Lauri and J{\"a}ger, Marc O.~J. and '
'Morooka, Eiaki V. and Federici Canova, Filippo and Ranawat, '
'Yashasvi S. and Gao, David Z. and Rinke, Patrick and Foster, '
'Adam S.}, '
'title = {{DScribe: Library of descriptors for machine '
'learning in materials science}}, '
'journal = {Computer Physics Communications}, '
'year = {2019}, pages = {106949}, '
'doi = {https://doi.org/10.1016/j.cpc.2019.106949}}'
]
def implementors(self):
return ["Lauri Himanen and the DScribe team", "Alex Dunn"]
import ase, pickle
import numpy as np
from dscribe.utils import AverageKernel
from ase.build import molecule
from ase.collections import g2
import time
# Choose descriptor
descriptor = "SOAP"
# Compute local descriptors
all_atomtypes = [1, 6]
#all_atomtypes = []
if descriptor == "SOAP":
desc = SOAP(all_atomtypes, 8.0, 2, 0, periodic=False, crossover=True)
print(desc.get_number_of_features())
elif descriptor == "ACSF":
desc = ACSF(n_atoms_max=15,
types=[1, 6, 7, 8],
bond_params=[[
1,
2,
], [
4,
5,
]],
bond_cos_params=[1, 2, 3, 4],
ang4_params=[[1, 2, 3], [3, 1, 4], [4, 5, 6], [7, 8, 9]],
ang5_params=[[1, 2, 3], [3, 1, 4], [4, 5, 6], [7, 8, 9]],
flatten=False)
else:
def test_numerical(self):
"""Test numerical values against a naive python implementation."""
# Elaborate test system with multiple species, non-cubic cell, and
# close-by atoms.
a = 1
system = (Atoms(
symbols=["C", "H", "O"],
cell=[[0, a, a], [a, 0, a], [a, a, 0]],
scaled_positions=[
[0, 0, 0],
[1 / 3, 1 / 3, 1 / 3],
[2 / 3, 2 / 3, 2 / 3],
],
pbc=[True, True, True],
) * (3, 3, 3))
# view(system)
# Two centers: one in the middle, one on the edge.
centers = [np.sum(system.get_cell(), axis=0) / 2, [0, 0, 0]]
h = 0.0001
n_atoms = len(system)
n_comp = 3
# The maximum error depends on how big the system is. With a small
# system the error is smaller for non-periodic systems than the
# corresponding error when periodicity is turned on. The errors become
# equal (~1e-5) when the size of the system is increased.
for periodic in [False]:
for rbf in ["gto", "polynomial"]:
for average in ["off", "outer", "inner"]:
soap = SOAP(
species=[1, 8, 6],
rcut=3,
nmax=4,
lmax=4,
rbf=rbf,
sparse=False,
average=average,
crossover=True,
periodic=periodic,
dtype=
"float64", # The numerical derivatives require double precision
)
n_features = soap.get_number_of_features()
if average != "off":
n_centers = 1
derivatives_python = np.zeros(
(n_atoms, n_comp, n_features))
else:
n_centers = len(centers)
derivatives_python = np.zeros(
(n_centers, n_atoms, n_comp, n_features))
d0 = soap.create(system, centers)
coeffs = [-1.0 / 2.0, 1.0 / 2.0]
deltas = [-1.0, 1.0]
for i_atom in range(len(system)):
for i_center in range(n_centers):
for i_comp in range(3):
for i_stencil in range(2):
if average == "off":
i_cent = [centers[i_center]]
else:
i_cent = centers
system_disturbed = system.copy()
i_pos = system_disturbed.get_positions()
i_pos[i_atom,
i_comp] += h * deltas[i_stencil]
system_disturbed.set_positions(i_pos)
d1 = soap.create(system_disturbed, i_cent)
if average != "off":
derivatives_python[
i_atom,
i_comp, :] += (coeffs[i_stencil] *
d1 / h)
else:
derivatives_python[
i_center, i_atom,
i_comp, :] += (coeffs[i_stencil] *
d1[0, :] / h)
# Calculate with central finite difference implemented in C++.
# Try both cartesian centers and indices.
for c in [centers]:
derivatives_cpp, d_cpp = soap.derivatives(
system, positions=c, method="numerical")
# Test that descriptor values are correct
self.assertTrue(np.allclose(d0, d_cpp, atol=1e-6))
# Compare values
# print(np.abs(derivatives_python).max())
# print(derivatives_python[0,1,:,:])
# print(derivatives_cpp[0,0,:,:])
self.assertTrue(
np.allclose(derivatives_python,
derivatives_cpp,
atol=2e-5))
def test_parallel(self):
"""Tests parallel output validity for both dense and sparse output."""
for sparse in [False, True]:
desc = SOAP(
species=[1, 6, 7, 8],
rcut=5,
nmax=3,
lmax=3,
sigma=1,
periodic=False,
crossover=True,
average="off",
sparse=sparse,
)
n_features = desc.get_number_of_features()
samples = [molecule("CO"), molecule("NO"), molecule("OH")]
centers = [[0], [0], [0]]
# Determining number of jobs based on the amount of CPUs
# desc.derivatives(system=samples, n_jobs=-1, only_physical_cores=False)
# desc.derivatives(system=samples, n_jobs=-1, only_physical_cores=True)
# Perhaps most common scenario: more systems than jobs, using all
# centers and indices
der, des = desc.derivatives(
system=samples,
n_jobs=2,
)
self.assertTrue(der.shape == (3, 2, 2, 3, n_features))
self.assertTrue(des.shape == (3, 2, n_features))
assumed_der = np.empty((3, 2, 2, 3, n_features))
assumed_des = np.empty((3, 2, n_features))
desc.sparse = False
assumed_der[0, :], assumed_des[0, :] = desc.derivatives(samples[0],
n_jobs=1)
assumed_der[1, :], assumed_des[1, :] = desc.derivatives(samples[1],
n_jobs=1)
assumed_der[2, :], assumed_des[2, :] = desc.derivatives(samples[2],
n_jobs=1)
desc._sparse = sparse
if sparse:
der = der.todense()
des = des.todense()
self.assertTrue(np.allclose(assumed_der, der))
self.assertTrue(np.allclose(assumed_des, des))
# More systems than jobs, using all centers and indices, not requesting
# descriptors
der = desc.derivatives(
system=samples,
return_descriptor=False,
n_jobs=2,
)
self.assertTrue(der.shape == (3, 2, 2, 3, n_features))
assumed_der = np.empty((3, 2, 2, 3, n_features))
desc._sparse = False
assumed_der[0, :] = desc.derivatives(samples[0],
n_jobs=1,
return_descriptor=False)
assumed_der[1, :] = desc.derivatives(samples[1],
n_jobs=1,
return_descriptor=False)
assumed_der[2, :] = desc.derivatives(samples[2],
n_jobs=1,
return_descriptor=False)
desc._sparse = sparse
if sparse:
der = der.todense()
self.assertTrue(np.allclose(assumed_der, der))
# More systems than jobs, using custom indices as centers
der, des = desc.derivatives(
system=samples,
positions=centers,
n_jobs=2,
)
self.assertTrue(der.shape == (3, 1, 2, 3, n_features))
self.assertTrue(des.shape == (3, 1, n_features))
assumed_der = np.empty((3, 1, 2, 3, n_features))
assumed_des = np.empty((3, 1, n_features))
desc._sparse = False
assumed_der[0, :], assumed_des[0, :] = desc.derivatives(samples[0],
centers[0],
n_jobs=1)
assumed_der[1, :], assumed_des[1, :] = desc.derivatives(samples[1],
centers[1],
n_jobs=1)
assumed_der[2, :], assumed_des[2, :] = desc.derivatives(samples[2],
centers[2],
n_jobs=1)
desc._sparse = sparse
if sparse:
der = der.todense()
des = des.todense()
self.assertTrue(np.allclose(assumed_der, der))
self.assertTrue(np.allclose(assumed_des, des))
# More systems than jobs, using custom cartesian centers
centers = [[[0, 1, 2]], [[2, 1, 0]], [[1, 2, 0]]]
der, des = desc.derivatives(
system=samples,
positions=centers,
n_jobs=2,
)
self.assertTrue(der.shape == (3, 1, 2, 3, n_features))
self.assertTrue(des.shape == (3, 1, n_features))
assumed_der = np.empty((3, 1, 2, 3, n_features))
assumed_des = np.empty((3, 1, n_features))
desc._sparse = False
assumed_der[0, :], assumed_des[0, :] = desc.derivatives(samples[0],
centers[0],
n_jobs=1)
assumed_der[1, :], assumed_des[1, :] = desc.derivatives(samples[1],
centers[1],
n_jobs=1)
assumed_der[2, :], assumed_des[2, :] = desc.derivatives(samples[2],
centers[2],
n_jobs=1)
desc._sparse = sparse
if sparse:
der = der.todense()
des = des.todense()
self.assertTrue(np.allclose(assumed_der, der))
self.assertTrue(np.allclose(assumed_des, des))
# Includes
includes = [[0], [0], [0]]
der, des = desc.derivatives(
system=samples,
include=includes,
n_jobs=2,
)
self.assertTrue(der.shape == (3, 2, 1, 3, n_features))
self.assertTrue(des.shape == (3, 2, n_features))
assumed_der = np.empty((3, 2, 1, 3, n_features))
assumed_des = np.empty((3, 2, n_features))
desc._sparse = False
assumed_der[0, :], assumed_des[0, :] = desc.derivatives(
samples[0], include=includes[0], n_jobs=1)
assumed_der[1, :], assumed_des[1, :] = desc.derivatives(
samples[1], include=includes[1], n_jobs=1)
assumed_der[2, :], assumed_des[2, :] = desc.derivatives(
samples[2], include=includes[2], n_jobs=1)
desc._sparse = sparse
if sparse:
der = der.todense()
des = des.todense()
self.assertTrue(np.allclose(assumed_der, der))
self.assertTrue(np.allclose(assumed_des, des))
# Excludes
excludes = [[0], [0], [0]]
der, des = desc.derivatives(
system=samples,
exclude=excludes,
n_jobs=2,
)
self.assertTrue(der.shape == (3, 2, 1, 3, n_features))
self.assertTrue(des.shape == (3, 2, n_features))
assumed_der = np.empty((3, 2, 1, 3, n_features))
assumed_des = np.empty((3, 2, n_features))
desc._sparse = False
assumed_der[0, :], assumed_des[0, :] = desc.derivatives(
samples[0], exclude=excludes[0], n_jobs=1)
assumed_der[1, :], assumed_des[1, :] = desc.derivatives(
samples[1], exclude=excludes[1], n_jobs=1)
assumed_der[2, :], assumed_des[2, :] = desc.derivatives(
samples[2], exclude=excludes[2], n_jobs=1)
desc._sparse = sparse
if sparse:
der = der.todense()
des = des.todense()
self.assertTrue(np.allclose(assumed_der, der))
self.assertTrue(np.allclose(assumed_des, des))
# Test averaged output
desc.average = "inner"
positions = [[0], [0, 1], [1]]
der, des = desc.derivatives(
system=samples,
positions=positions,
n_jobs=2,
)
self.assertTrue(der.shape == (3, 1, 2, 3, n_features))
self.assertTrue(des.shape == (3, 1, n_features))
desc._sparse = False
assumed_der = np.empty((3, 1, 2, 3, n_features))
assumed_des = np.empty((3, 1, n_features))
assumed_der[0, :], assumed_des[0, :] = desc.derivatives(
samples[0], positions=positions[0], n_jobs=1)
assumed_der[1, :], assumed_des[1, :] = desc.derivatives(
samples[1], positions=positions[1], n_jobs=1)
assumed_der[2, :], assumed_des[2, :] = desc.derivatives(
samples[2], positions=positions[2], n_jobs=1)
desc._sparse = sparse
if sparse:
der = der.todense()
des = des.todense()
self.assertTrue(np.allclose(assumed_der, der))
self.assertTrue(np.allclose(assumed_des, des))
# Variable size list output, as the systems have a different size
desc.average = "off"
samples = [molecule("CO"), molecule("NO2"), molecule("OH")]
der, des = desc.derivatives(
system=samples,
n_jobs=2,
)
self.assertTrue(isinstance(der, list))
self.assertTrue(der[0].shape == (2, 2, 3, n_features))
self.assertTrue(der[1].shape == (3, 3, 3, n_features))
self.assertTrue(der[2].shape == (2, 2, 3, n_features))
desc._sparse = False
for i in range(len(samples)):
assumed_der, assumed_des = desc.derivatives(samples[i],
n_jobs=1)
i_der = der[i]
i_des = des[i]
if sparse:
i_der = i_der.todense()
i_des = i_des.todense()
self.assertTrue(np.allclose(assumed_der, i_der))
self.assertTrue(np.allclose(assumed_des, i_des))
desc._sparse = sparse
temp = df['temp']
load = df['load']
print('Got training targets from spread sheet')
species = set()
for i in range(len(atoms)):
species.update(atoms[i].get_chemical_symbols())
soap = SOAP(species=species,
periodic=True,
rcut=5,
nmax=1,
lmax=1,
average="outer")
print('Training hardness model...')
soap.get_number_of_features()
feature_vectors = soap.create(atoms, n_jobs=1)
feature_tensor = th.tensor(feature_vectors)
#print('DONE, SOAP descriptors ready to use')
feature_pd = pd.DataFrame(feature_vectors) #feature pd is the soap descriptors
#generate compositional descriptors
df = pd.read_excel(
'/Users/ziyanzhang/Downloads/dscribe/examples/hv_temp_cif_labels.xlsx')
class Vectorize_Formula:
def __init__(self):
elem_dict = pd.read_excel(
'/Users/ziyanzhang/Desktop/subgroup/elementsnew.xlsx'