def generate_distances(mols, args):
t = time.time()
if args.method in ["bob", "cm", "ucm"]:
X = np.asarray([mol.descriptor for mol in mols], dtype=float)
if args.exponents[0] == 1:
D = manhattan_distance(X.T, X.T)
else:
# prevent overflow in l2
X = X[:, (X > 1.5).all(0)]
D = l2_distance(X.T, X.T)
else:
#X = np.asarray([np.concatenate(mol.local_descriptor) for mol in mols])
#D = manhattan_distance(X.T, X.T)
# initialize
D = np.zeros((len(mols), len(mols)), dtype=float)
resids = np.unique(mols[0].residues)
for res in resids:
if args.ca:
indices = np.where((mols[0].residues == res)
& (mols[0].atomnames == "CA"))[0]
else:
indices = np.where(mols[0].residues == res)[0]
D_res = np.zeros((len(mols), len(mols)), dtype=float)
for i in indices:
X = np.asarray([mol.local_descriptor[i] for mol in mols])
if args.exponents[2] == 1:
D_res += manhattan_distance(X.T, X.T)**args.exponents[1]
else:
D_res += l2_distance(X.T, X.T)**args.exponents[1]
if args.exponents[1] == 2:
np.sqrt(D_res, out=D_res)
D += D_res**args.exponents[0]
if args.exponents[0] == 2:
np.sqrt(D, out=D)
print "Generated distances in %.1f seconds" % (time.time() - t)
return D
# Set hyper-parameters
sigma = 10**(4.2)
llambda = 10**(-10.0)
# print "Calculating K-matrix ...",
print(u"Calculating: K\u1D62\u2C7C = k(q\u1D62, q\u2C7C) ... ",
end="")
sys.stdout.flush()
start = time.time()
# Gaussian kernel usually better for atomic properties
# K = gaussian_kernel(X, X, sigma)
# Alternatively, just calculate the L2 distance, and convert
# to kernel matrix (e.g. for multiple sigmas, etc):
D = l2_distance(X, X)
D /= -2.0 * sigma * sigma
D = np.exp(K)
print("%7.2f seconds" % (time.time() - start))
for i in xrange(n_train):
K[i, i] += llambda
print(u"Calculating: \u03B1 = (K + \u03BBI)\u207B\u00B9 y ... ",
end="")
sys.stdout.flush()
start = time.time()
alpha = cho_solve(K, Y)
print("%7.2f seconds" % (time.time() - start))
sigma = float(sys.argv[1])
llambda = float(sys.argv[2])
xyz_file = sys.argv[3]
mol = Molecule()
mol.read_xyz(xyz_file)
mol.generate_atomic_coulomb_matrix()
Ys = np.zeros((mol.natoms))
s = (-0.5 / sigma**2)
print "Running predicition ..."
for a, atom in enumerate(mol.atomtypes):
Xs = np.reshape(mol.atomic_coulomb_matrix[a],
(len(mol.atomic_coulomb_matrix[a]), 1))
Ds = l2_distance(Xs, X[atom])
Ds *= s
np.array(np.exp(Ds, Ds))
Ys[a] = np.dot(Ds[0, :], alpha[atom])
error = np.sum(Ys)
Ys -= np.mean(Ys)
print "Final charges:"
for i, q in enumerate(Ys):
print "%s %12.8f" % (mol.atomtypes[i], q)