def get_least_polynomial_coefficients( v, num_dims, num_pts, k,
basis_indices = None ):
if basis_indices is None:
H = numpy.zeros( (num_pts, polynomial_space_dimension(num_dims,
k[num_pts-1,0] ) ),
numpy.double );
#print k
v_counter = 0;
for q in range( num_pts ):
current_size = polynomial_subspace_dimension( num_dims, k[q,0] );
v_inds = range( v_counter, v_counter + current_size );
previous_dimension = polynomial_space_dimension(num_dims, k[q,0]-1 );
H_cols = range(previous_dimension,previous_dimension+current_size);
#print H_cols, v_inds
H[q,H_cols] = v[v_inds];
v_counter += current_size;
else:
num_basis_indices = len( basis_indices )
H = numpy.zeros( ( num_pts, num_basis_indices ),
numpy.double );
v_counter = 0;
previous_dimension = 0
degree = k[0,0]
#print k, basis_indices
for q in range( num_pts ):
"""
if ( k[q] != degree or v_counter == 0 ):
degree = k[q,0]
current_size = 0
while ( q+current_size < num_pts and k[q+current_size]==degree ):
current_size += 1
if ( q +current_size == num_pts and k[q] != k[q-1]):
current_size = num_basis_indices - q
"""
if ( k[q] != k[q-1] or q == 0):
current_size = 0
for index in basis_indices:
if ( index.level_sum() == k[q] ):
current_size += 1
v_inds = range( v_counter, v_counter + current_size );
H_cols = range(previous_dimension,previous_dimension+current_size);
#print H_cols, v_inds
H[q,H_cols] = v[v_inds];
v_counter += current_size;
if ( q+1 < num_pts and k[q] != k[q+1]):
previous_dimension += current_size
return H
def cv_vs_error_study( build_pts, build_vals, domain,
test_pts, test_vals,
results_file = None,
cv_file = None, solver_type = 2 ):
num_dims = build_pts.shape[0]
if ( num_dims == 10 ):
max_order = 5
elif ( num_dims == 15 ):
max_order = 4
else:
max_order = 3
poly_1d = [ LegendrePolynomial1D() ]
basis = TensorProductBasis( num_dims, poly_1d )
pce = PCE( num_dims, order = 0, basis = basis, func_domain = domain )
orders = numpy.arange( 1, max_order + 1 )
solvers = numpy.array( [solver_type], numpy.int32 )
cv_params_grid_array = cartesian_product( [solvers,orders] )
cv_params_grid = []
for i in xrange( cv_params_grid_array.shape[0] ):
cv_params = {}
cv_params['solver'] = numpy.int32( cv_params_grid_array[i,0] )
cv_params['order'] = numpy.int32( cv_params_grid_array[i,1] )
num_pce_terms = polynomial_space_dimension( num_dims,
cv_params['order'] )
if ( cv_params['solver'] <= 1 and
num_pce_terms >= build_pts.shape[1] ):
cv_params['lambda'] = 1.e-12
cv_params_grid.append( cv_params )
# print cv_params_grid
# cv_iterator = LeaveOneOutCrossValidationIterator()
cv_iterator = KFoldCrossValidationIterator( num_folds = 20 )
CV = GridSearchCrossValidation( cv_iterator, pce,
use_predictor_cross_validation = True,
use_fast_predictor_cross_validation = True )
t0 = time.time()
CV.run( build_pts, build_vals, cv_params_grid )
time_taken = time.time() - t0
print 'cross validation took ', time_taken, ' seconds'
print "################"
print "Best cv params: ", CV.best_cv_params
print "Best cv score: ", CV.best_score
print "################"
for order in orders:
residual_norms = numpy.empty( len( CV.cv_params_set ), numpy.double )
scores = numpy.empty( len( CV.cv_params_set ), numpy.double )
k = 0
for i in xrange( len( CV.cv_params_set ) ):
if ( CV.cv_params_set[i]['order'] == order ):
residual_norms[k] = CV.cv_params_set[i]['norm_residual']
scores[k] = CV.scores[i]
k += 1
residual_norms.resize( k )
scores.resize( k )
pce = PCE( num_dims,
order = order,
basis = basis,
func_domain = domain )
V = pce.vandermonde( build_pts ).T
pce.set_solver( CV.best_cv_params['solver'] )
# pce.linear_solver.max_iterations = 3
sols, sol_metrics = pce.linear_solver.solve( V, build_vals )
from sklearn.linear_model import orthogonal_mp
l2_error = numpy.empty( ( sols.shape[1] ), numpy.double )
residuals = numpy.empty( ( sols.shape[1] ), numpy.double )
test_pts = numpy.random.uniform( 0., 1., ( num_dims, 1000 ) )
f = GenzModel( domain, 'oscillatory' )
# f.set_coefficients( 4.5, 'no-decay' )
f.set_coefficients( 4.5, 'quadratic-decay' )
test_vals = f( test_pts ).reshape( ( test_pts.shape[1], 1 ) )
for i in xrange( sols.shape[1] ):
coeff = sols[:,i]
pce.set_coefficients( coeff )
residuals[i] = numpy.linalg.norm( build_vals -
pce.evaluate_set( build_pts ) )
num_test_pts = test_pts.shape[1]
pce_vals_pred = pce.evaluate_set( test_pts ).T
error = test_vals.squeeze() - pce_vals_pred
l2_error[i] = numpy.linalg.norm( error ) / numpy.sqrt( num_test_pts )
import pylab
print residuals, l2_error
print residual_norms, scores
pylab.loglog( residuals, l2_error, label = str( order ) + 'true' )
pylab.loglog( residual_norms, scores, label = str( order )+'-cv' )
pylab.xlim([1e-3,10])
pylab.legend()
pylab.show()