Skip to content

GitLab

Commit a403108e authored by Carl Poelking's avatar Carl Poelking
Browse files

Kernel adaptor for esoap.

parent e9857d56
Hide whitespace changes
Inline Side-by-side
src/soap/soapy/kernel.py
View file @ a403108e
......@@ -190,7 +190,7 @@ class KernelAdaptorSpecificUnique(object):
def adaptScalar(self, atomic, epsilon=1e-20):
X = reduce_xnklab_atomic(atomic, self.types)
X_mag = np.dot(X,X)**0.5
if X_mag < epsilon:
if X_mag < epsilon:
X_mag = 1.
X_norm = X/X_mag
return X, X_norm
......@@ -223,6 +223,37 @@ class KernelAdaptorSpecific(object):
X_norm = X/np.dot(X,X)**0.5
return X, X_norm
class KernelAdaptorGlobalSpecificEnergy(object):
def __init__(self, options, types_global):
self.types_global = types_global
return
def adapt(self, espectrum, return_pos_matrix=False, dtype='float64'):
dim_linear_ab = None
dim_linear_total = None
S = len(self.types_global)
X = None
for sa in self.types_global:
for sb in self.types_global:
powspec_ab = energy_spectrum.getPowerGlobal(sa, sb)
if powspec_ab != None:
xab = powspec_ab.array.real.flatten()
if not dim_linear_ab:
dim_linear_ab = xab.shape[0]
dim_linear_total = S**2*dim_linear_ab
X = np.zeros((S, S*dim_linear_ab), dtype=dtype)
else: assert dim_linear_ab == xab.shape[0]
isa = self.types_global.index(sa)
isb = self.types_global.index(sb)
i0 = isa
i1 = isa+1
j0 = isb*dim_linear_ab
j1 = j0+dim_linear_ab
X[i0:i1,j0:j1] = xab
X = X/np.dot(X,X)**0.5
if return_pos_matrix:
return X, np.array([[0.,0.,0.]]), ["global"]
else: return X.flatten()
class KernelAdaptorGlobalSpecific(object):
def __init__(self, options, types_global=None):
self.types = types_global
......@@ -270,13 +301,13 @@ class KernelAdaptorGeneric(object):
# if not pid in pid_list:
# logging.debug("Skip, adapt atomic, pid = %d" % pid)
# continue
Xi_unnorm, Xi_norm = self.adaptScalar(atomic_i)
dimX = Xi_norm.shape[0]
#Xi = np.real(atomic_i.getPower("","").array)
#dimX = Xi.shape[0]*Xi.shape[1]
#Xi = Xi.reshape((dimX))
#print "dim(X) =", dimX
if not IX.any():
IX = np.copy(Xi_norm) # TODO Is this necessary?
......@@ -299,7 +330,7 @@ class KernelAdaptorGeneric(object):
dxnkl_pid = atomic.getPowerGradGeneric(nb_pid)
dX_dx = dxnkl_pid.getArrayGradX()
dX_dy = dxnkl_pid.getArrayGradY()
dX_dz = dxnkl_pid.getArrayGradZ()
dX_dz = dxnkl_pid.getArrayGradZ()
dimX = dX_dx.shape[0]*dX_dx.shape[1]
dX_dx = np.real(dX_dx.reshape((dimX)))
dX_dy = np.real(dX_dy.reshape((dimX)))
......@@ -310,7 +341,7 @@ class KernelAdaptorGeneric(object):
dX_dy = dX_dy/mag_X - np.dot(X, dX_dy)/mag_X**3 * X
dX_dz = dX_dz/mag_X - np.dot(X, dX_dz)/mag_X**3 * X
return dX_dx, dX_dy, dX_dz
class KernelAdaptorGlobalGeneric(object):
def __init__(self, options, types_global=None):
return
......@@ -341,7 +372,7 @@ class KernelAdaptorGlobalGeneric(object):
dxnkl_pid = atomic.getPowerGradGeneric(nb_pid)
dX_dx = dxnkl_pid.getArrayGradX()
dX_dy = dxnkl_pid.getArrayGradY()
dX_dz = dxnkl_pid.getArrayGradZ()
dX_dz = dxnkl_pid.getArrayGradZ()
dimX = dX_dx.shape[0]*dX_dx.shape[1]
dX_dx = np.real(dX_dx.reshape((dimX)))
dX_dy = np.real(dX_dy.reshape((dimX)))
......@@ -353,7 +384,6 @@ class KernelAdaptorGlobalGeneric(object):
dX_dz = dX_dz/mag_X - np.dot(X, dX_dz)/mag_X**3 * X
return dX_dx, dX_dy, dX_dz
class KernelFunctionGaussianDiff(object):
def __init__(self, options):
self.delta = float(options['kernel.gaussian_delta'])
......@@ -393,7 +423,7 @@ class KernelFunctionDot(object):
def computeDot(self, IX, X, xi, delta):
return delta**2 * np.dot(IX,X)**xi
def compute(self, IX, X):
return self.computeDot(IX, X, self.xi, self.delta)
return self.computeDot(IX, X, self.xi, self.delta)
def computeDerivativeOuter(self, IX, X):
c = self.computeDot(IX, X, self.xi-1, self.delta)
return self.xi*np.diag(c).dot(IX)
......@@ -450,7 +480,7 @@ class KernelFunctionDotHarmonicDist(object):
D = (1.-C+self.eps)**0.5
IC = self.kfctdot.computeDerivativeOuter(IX, X)
return 2*(D-self.d0)*0.5/D*(-1.)*IC
class KernelFunctionDotLj(object):
def __init__(self, options):
self.sigma = float(options.get('kernel.lj_sigma'))
......@@ -466,7 +496,7 @@ class KernelFunctionDotLj(object):
IC = self.kfctdot.computeDerivativeOuter(IX, X)
return (-6.*self.sigma**12*(1./D)**7 + 3.*self.sigma**6*(1./D)**4)*IC
def computeBlockDot(self, IX, return_distance=False):
return self.kfctdot.computeBlock(IX, return_distance)
return self.kfctdot.computeBlock(IX, return_distance)
class KernelFunctionDot3Harmonic(object):
def __init__(self, options):
......@@ -521,21 +551,22 @@ class KernelFunctionDot3HarmonicDist(object):
return 2*(D_A - D_B)*(0.5/D_A*(-1.)*IC_A - 0.5/D_B*(-1.)*IC_B) + 2*(D_A - D_AB)*0.5/D_A*(-1.)*IC_A + 2*(D_B - D_AB)*0.5/D_B*(-1.)*IC_B + 2*(D_A-self.d0)*1./D_A*(-1.)*IC_A + 2*(D_B-self.d0)*1./D_B*(-1.)*IC_B
KernelAdaptorFactory = {
'generic': KernelAdaptorGeneric,
'generic': KernelAdaptorGeneric,
'specific': KernelAdaptorSpecific,
'specific-unique': KernelAdaptorSpecificUnique,
'global-generic': KernelAdaptorGlobalGeneric,
'global-specific': KernelAdaptorGlobalSpecific
'global-specific': KernelAdaptorGlobalSpecific,
'global-specific-energy': KernelAdaptorGlobalSpecificEnergy
}
KernelFunctionFactory = {
KernelFunctionFactory = {
'gaussian-diff': KernelFunctionGaussianDiff,
'dot': KernelFunctionDot,
'dot': KernelFunctionDot,
'dot-shifted': KernelFunctionDotShifted,
'dot-harmonic': KernelFunctionDotHarmonic,
'dot-harmonic': KernelFunctionDotHarmonic,
'dot-harmonic-dist': KernelFunctionDotHarmonicDist,
'dot-lj': KernelFunctionDotLj,
'dot-3-harmonic': KernelFunctionDot3Harmonic,
'dot-3-harmonic': KernelFunctionDot3Harmonic,
'dot-3-harmonic-dist': KernelFunctionDot3HarmonicDist
}
......@@ -543,7 +574,7 @@ class KernelPotential(object):
def __init__(self, options):
logging.info("Construct kernel potential ...")
self.basis = soap.Basis(options)
self.options = options
self.options = options
# CORE DATA
self.IX = None
self.alpha = None
......@@ -551,7 +582,7 @@ class KernelPotential(object):
self.labels = []
# KERNEL
logging.info("Choose kernel function ...")
self.kernelfct = KernelFunctionFactory[options.get('kernel.type')](options)
self.kernelfct = KernelFunctionFactory[options.get('kernel.type')](options)
# ADAPTOR
logging.info("Choose adaptor ...")
self.adaptor = KernelAdaptorFactory[options.get('kernel.adaptor')](options)
......@@ -562,7 +593,7 @@ class KernelPotential(object):
def computeKernelMatrix(self, return_distance=False):
return self.kernelfct.computeBlock(self.IX, return_distance)
def computeDotKernelMatrix(self, return_distance=False):
# Even if the, e.g., dot-harmonic kernel is used for optimization,
# Even if the, e.g., dot-harmonic kernel is used for optimization,
# the dot-product based kernel matrix may still be relevant
return self.kernelfct.computeBlockDot(self.IX, return_distance)
def nConfigs(self):
......@@ -635,7 +666,7 @@ class KernelPotential(object):
# TODO The kernel policy should take care of this:
if self.use_global_spectrum:
spectrum.computeGlobal()
IX_acqu = self.adaptor.adapt(spectrum)
IX_acqu = self.adaptor.adapt(spectrum)
n_acqu = IX_acqu.shape[0]
dim_acqu = IX_acqu.shape[1]
logging.info("Compute energy from %d atomic environments ..." % n_acqu)
......@@ -684,18 +715,18 @@ class KernelPotential(object):
# Force on neighbour
logging.info(" -> Nb %d" % (nb_pid))
dX_dx, dX_dy, dX_dz = self.adaptor.adaptGradients(atomic, nb_pid, X_unnorm)
force_x = -alpha_dIC.dot(dX_dx)
force_y = -alpha_dIC.dot(dX_dy)
force_z = -alpha_dIC.dot(dX_dz)
forces[nb_pid-1][0] += force_x
forces[nb_pid-1][1] += force_y
forces[nb_pid-1][2] += force_z
#print forces
#raw_input('...')
return forces
def restore_positions(structure, positions):
idx = 0
for part in structure:
......@@ -711,7 +742,7 @@ def perturb_positions(structure, exclude_pid=[]):
dz = np.random.uniform(-1.,1.)
part.pos = part.pos + 0.1*np.array([dx,dy,dz])
return [ part.pos for part in structure ]
def random_positions(structure, exclude_pid=[], b0=-1., b1=1., x=True, y=True, z=True):
for part in structure:
if part.id in exclude_pid: continue
......@@ -733,7 +764,7 @@ def apply_force_step(structure, forces, scale, constrain_particles=[]):
idx = -1
for part in structure:
idx += 1
if part.id in constrain_particles:
if part.id in constrain_particles:
print "Skip force step, pid =", part.id
continue
#if np.random.uniform(0.,1.) > 0.5: continue
......@@ -750,7 +781,7 @@ def apply_force_norm_step(structure, forces, scale, constrain_particles=[]):
idx = -1
for part in structure:
idx += 1
if part.id in constrain_particles:
if part.id in constrain_particles:
print "Skip force step, pid =", part.id
continue
#if np.random.uniform(0.,1.) > 0.5: continue
......@@ -763,8 +794,8 @@ def evaluate_energy(positions, structure, kernelpot, opt_pids, verbose=False, of
pid_pos = positions.reshape((opt_pids.shape[0],3))
for pidx, pid in enumerate(opt_pids):
pos = pid_pos[pidx,:]
particle = structure.getParticle(pid)
particle.pos = pos
particle = structure.getParticle(pid)
particle.pos = pos
for part in structure:
if verbose: print part.id, part.type, part.pos
# Evaluate energy function
......@@ -782,8 +813,8 @@ def evaluate_energy_gradient(positions, structure, kernelpot, opt_pids, verbose=
pid_pos = positions.reshape((opt_pids.shape[0],3))
for pidx, pid in enumerate(opt_pids):
pos = pid_pos[pidx,:]
particle = structure.getParticle(pid)
particle.pos = pos
particle = structure.getParticle(pid)
particle.pos = pos
for part in structure:
if verbose: print part.id, part.type, part.pos
# Evaluate forces
......@@ -797,4 +828,3 @@ def evaluate_energy_gradient(positions, structure, kernelpot, opt_pids, verbose=
if verbose: print gradients_short
#forces[2] = 0. # CONSTRAIN TO Z=0 PLANE
return gradients_short
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment