#!/usr/bin/python
# -*- coding: utf-8 -*-
import numpy as num
import scipy as sp
import time
from kite import covariance_ext
from pyrocko import guts
from pyrocko.guts_array import Array
from kite.util import (Subject, property_cached, # noqa
trimMatrix, derampMatrix, squareMatrix)
__all__ = ['Covariance', 'CovarianceConfig']
NOISE_PATCH_MIN_PX = 1024
NOISE_PATCH_MAX_NAN = 0.6
noise_regimes = [
(1./2000, num.inf),
(1./2000, 1./500),
(1./500, 1./10),
(0, num.inf)]
[docs]def modelCovarianceExponential(distance, a, b):
"""Exponential function model to approximate a positive-definite covariance
We assume the following simple covariance model to describe the empirical
noise observations:
.. math::
cov(d) = c \\cdot e^{\\frac{-d}{b}}
:param distance: Distance between
:type distance: float or :class:`numpy.ndarray`
:param a: Linear model coefficient
:type a: float
:param b: Exponential model coefficient
:type b: float
:returns: Covariance at ``distance``
:rtype: :class:`numpy.ndarray`
"""
return a * num.exp(-distance/b)
[docs]def modelCovarianceExponentialCosine(distance, a, b, c, d):
r"""Exponential function model to approximate a positive-definite covariance
We assume the following simple covariance model to describe the empirical
noise observations:
.. math::
cov(d) = c \\cdot e^{\\frac{-d}{b}} \\cdot \cos{\\frac{d-c}{d}}
:param distance: Distance between
:type distance: float or :class:`numpy.ndarray`
:param a: Linear model coefficient
:type a: float
:param b: Exponential model coefficient
:type b: float
:param c: Cosinus distance correction
:type c: float
:param c: Cosinus coefficient
:type c: float
:returns: Covariance at ``distance``
:rtype: :class:`numpy.ndarray`
"""
return a * num.exp(-distance/b) * num.cos((distance-c)/d)
[docs]def modelPowerspec(k, beta, D):
"""Exponential linear model to estimate a log-linear power spectrum
We assume the following log-linear model for a measured power spectrum:
.. math::
pow(k) = \\frac{k^\\beta}{D}
:param k: Wavenumber
:type k: float or :class:`numpy.ndarray`
:param a: Exponential model factor
:type a: float
:param b: Fractional model factor
:type b: float
"""
return (k**beta)/D
[docs]class CovarianceConfig(guts.Object):
noise_coord = Array.T(
shape=(None,), dtype=num.float,
serialize_as='list',
optional=True,
help='Noise patch coordinates and size,')
model_coefficients = guts.Tuple.T(
optional=True,
help='Covariance model coefficients. Either two (exponential) '
'or three (exponential and cosine term) coefficients.'
'See also :func:`~kite.covariance.modelCovariance`.')
model_function = guts.StringChoice.T(
choices=['exponential', 'exponential_cosine'],
default='exponential',
help='Covariance approximation function.')
sampling_method = guts.StringChoice.T(
choices=['spectral', 'spatial'],
default='spatial',
help='Method for estimating the covariance and structure function.')
spatial_bins = guts.Int.T(
default=75,
help='Number of distance bins for spatial covariance sampling.')
spatial_pairs = guts.Int.T(
default=200000,
help='Number of random pairs for spatial covariance sampling.')
variance = guts.Float.T(
optional=True,
help='Variance of the model.')
adaptive_subsampling = guts.Bool.T(
default=True,
help='Adaptive subsampling flag for full covariance calculation.')
covariance_matrix = Array.T(
dtype=num.float,
optional=True,
serialize_as='base64',
help='Cached covariance matrix, '
'see :attr:`~kite.Covariance.covariance_matrix`.')
def __init__(self, *args, **kwargs):
if len(kwargs) != 0:
if 'a' in kwargs and 'b' in kwargs:
kwargs['model_coefficients'] = (
kwargs.pop('a'), kwargs.pop('b'))
guts.Object.__init__(self, *args, **kwargs)
[docs]class Covariance(object):
"""Construct the variance-covariance matrix of quadtree subsampled data.
Variance and covariance estimates are used to construct the weighting
matrix to be used later in an optimization.
Two different methods exist to propagate full-resolution data variances
and covariances of :class:`kite.Scene.displacement` to the
covariance matrix of the subsampled dataset:
1. The distance between :py:class:`kite.quadtree.QuadNode`
leaf focal points, :py:class:`kite.covariance.Covariance.matrix_focal`
defines the approximate covariance of the quadtree leaf pair.
2. The _accurate_ propagation of covariances by taking the mean of
every node pair pixel covariances. This process is computational
very expensive and can take a few minutes.
:py:class:`kite.covariance.Covariance.matrix_focal`
:param quadtree: Quadtree to work on
:type quadtree: :class:`~kite.Quadtree`
:param config: Config object
:type config: :class:`~kite.covariance.CovarianceConfig`
"""
def __init__(self, scene, config=CovarianceConfig()):
self.evChanged = Subject()
self.evConfigChanged = Subject()
self.frame = scene.frame
self.quadtree = scene.quadtree
self.scene = scene
self._noise_data = None
self._powerspec1d_cached = None
self._powerspec2d_cached = None
self._powerspec3d_cached = None
self._noise_data_grid = None
self._initialized = False
self._nthreads = 0
self._log = scene._log.getChild('Covariance')
self.setConfig(config)
self.quadtree.evChanged.subscribe(self._clear)
self.scene.evConfigChanged.subscribe(self.setConfig)
def __call__(self, *args, **kwargs):
return self.getLeafCovariance(*args, **kwargs)
[docs] def setConfig(self, config=None):
""" Sets and updated the config of the instance
:param config: New config instance, defaults to configuration provided
by parent :class:`~kite.Scene`
:type config: :class:`~kite.covariance.CovarianceConfig`, optional
"""
if config is None:
config = self.scene.config.covariance
if self.scene.config.old_import:
self._log.warning('Old format - resetting noise patch coordinates')
config.covariance_matrix = None
config.noise_coord = None
self.config = config
if config.noise_coord is None\
and (config.model_coefficients is not None or
config.variance is not None):
self.noise_data # init data array
self.config.model_coefficients = config.model_coefficients
self.config.variance = config.variance
self._clear(config=False)
self.evConfigChanged.notify()
def _clear(self, config=True, spectrum=True):
if config:
self.config.model_coefficients = None
self.config.variance = None
self.config.covariance_matrix = None
if spectrum:
self.structure_spectral = None
self._powerspec1d_cached = None
self._powerspec2d_cached = None
self._noise_data_grid = None
self.covariance_matrix = None
self.covariance_matrix_focal = None
self.covariance_spectral = None
self.covariance_spatial = None
self.structure_spatial = None
self.weight_matrix = None
self.weight_matrix_focal = None
self._initialized = False
self.evChanged.notify()
@property
def nthreads(self):
""" Number of threads (CPU cores) to use for full covariance
calculation
Setting ``nthreads`` to ``0`` uses all available cores (default).
:setter: Sets the number of threads
:type: int
"""
return self._nthreads
@nthreads.setter
def nthreads(self, value):
self._nthreads = int(value)
@property
def finished_combinations(self):
return covariance_ext.get_finished_combinations()
@property
def noise_coord(self):
""" Coordinates of the noise patch in local coordinates.
:setter: Set the noise coordinates
:getter: Get the noise coordinates
:type: :class:`numpy.ndarray`, ``[llE, llN, sizeE, sizeN]``
"""
if self.config.noise_coord is None:
self.noise_data
return self.config.noise_coord
@noise_coord.setter
def noise_coord(self, values):
self.config.noise_coord = num.array(values)
@property
def noise_patch_size_km2(self):
"""
:getter: Noise patch size in :math:`km^2`.
:type: float
"""
if self.noise_coord is None:
return 0.
size = (self.noise_coord[2] * self.noise_coord[3])*1e-6
if self.noise_data.size < self.NOISE_PATCH_MIN_PX:
self._log.warning('Defined noise patch is instably small')
return size
@property
def noise_data(self, data):
""" Noise data we process to estimate the covariance
:setter: Set the noise patch to analyze the covariance.
:getter: If the noise data has not been set manually, we grab data
through :func:`~kite.Covariance.selectNoiseNode`.
:type: :class:`numpy.ndarray`
"""
return self._noise_data
@noise_data.getter
def noise_data(self):
if self._noise_data is not None:
return self._noise_data
elif self.config.noise_coord is not None:
self._log.debug('Selecting noise_data from config...')
llE, llN = self.scene.frame.mapENMatrix(
*self.config.noise_coord[:2])
sE, sN = self.scene.frame.mapENMatrix(
*self.config.noise_coord[2:])
slice_E = slice(llE, llE + sE)
slice_N = slice(llN, llN + sN)
covariance_matrix = self.config.covariance_matrix
self.noise_data = self.scene.displacement[slice_N, slice_E]
self.config.covariance_matrix = covariance_matrix
else:
self._log.debug('Selecting noise_data from Quadtree...')
node = self.selectNoiseNode()
self.noise_data = node.displacement
self.noise_coord = [node.llE, node.llN,
node.sizeE, node.sizeN]
return self.noise_data
@noise_data.setter
def noise_data(self, data):
data = data.copy()
data = derampMatrix(trimMatrix(data))
data[num.isnan(data)] = 0.
self._noise_data = data
self._clear()
@property
def noise_data_gridE(self):
return self._get_noise_data_grid()[0]
@property
def noise_data_gridN(self):
return self._get_noise_data_grid()[1]
def _get_noise_data_grid(self):
if self._noise_data_grid is None:
scene = self.scene
llE, llN = scene.frame.mapENMatrix(*self.noise_coord[:2])
sE, sN = scene.frame.mapENMatrix(*self.noise_coord[2:])
slice_E = slice(llE, llE + sE + 1)
slice_N = slice(llN, llN + sN + 1)
gridE = scene.frame.gridEmeter[slice_N, slice_E]
gridN = scene.frame.gridNmeter[slice_N, slice_E]
gridE = trimMatrix(self.noise_data, data=gridE)
gridN = trimMatrix(self.noise_data, data=gridN)
self._noise_data_grid = (gridE, gridN)
return self._noise_data_grid
[docs] def selectNoiseNode(self):
""" Choose noise node from quadtree
the biggest :class:`~kite.quadtree.QuadNode` from
:class:`~kite.Quadtree`.
:returns: A quadnode with the least signal.
:rtype: :class:`~kite.quadtree.QuadNode`
"""
t0 = time.time()
node_selection = [n for n in self.quadtree.nodes
if n.npixel > NOISE_PATCH_MIN_PX
and n.nan_fraction < NOISE_PATCH_MAX_NAN]
if not node_selection:
node_selection = self.quadtree.leaves
stdmax = max([n.std for n in node_selection])
lmax = max([n.std for n in node_selection])
def costFunction(n):
nl = num.log2(n.length)/num.log2(lmax)
ns = n.std/stdmax
return nl*(1.-ns)*(1.-n.nan_fraction)
fitness = num.array([costFunction(n) for n in node_selection])
self._log.debug('Fetched noise from Quadtree.nodes [%0.4f s]'
% (time.time() - t0))
return node_selection[num.argmin(fitness)]
def _mapLeaves(self, nx, ny):
""" Helper function returning appropriate
:class:`~kite.quadtree.QuadNode` and for maintaining
the internal mapping with the matrices.
:param nx: matrix x position
:type nx: int
:param ny: matrix y position
:type ny: int
:returns: tuple of :class:`~kite.quadtree.QuadNode` s for ``nx``
and ``ny``
:rtype: tuple
"""
leaf1 = self.quadtree.leaves[nx]
leaf2 = self.quadtree.leaves[ny]
self._leaf_mapping[leaf1.id] = nx
self._leaf_mapping[leaf2.id] = ny
return leaf1, leaf2
def isFullCovarianceCalculated(self):
if self.config.covariance_matrix is None:
return False
return True
@property_cached
def covariance_matrix(self):
""" Covariance matrix calculated from mean of all pixel pairs
inside the node pairs (full and accurate propagation).
:type: :class:`numpy.ndarray`,
size (:class:`~kite.Quadtree.nleaves` x
:class:`~kite.Quadtree.nleaves`)
"""
if not isinstance(self.config.covariance_matrix, num.ndarray):
self.config.covariance_matrix =\
self._calcCovarianceMatrix(method='full')
self.evChanged.notify()
elif self.config.covariance_matrix.ndim == 1:
try:
nl = self.quadtree.nleaves
self.config.covariance_matrix =\
self.config.covariance_matrix.reshape(nl, nl)
except ValueError:
self.config.covariance_matrix = None
return self.covariance_matrix
return self.config.covariance_matrix
@property_cached
def covariance_matrix_focal(self):
""" Approximate Covariance matrix from quadtree leaf pair
distance only. Fast, use for intermediate steps only and
finallly use approach :attr:`~kite.Covariance.covariance_matrix`.
:type: :class:`numpy.ndarray`,
size (:class:`~kite.Quadtree.nleaves` x
:class:`~kite.Quadtree.nleaves`)
"""
return self._calcCovarianceMatrix(method='focal')
@property_cached
def weight_matrix(self):
""" Weight matrix from full covariance :math:`cov^{-1}`.
:type: :class:`numpy.ndarray`,
size (:class:`~kite.Quadtree.nleaves` x
:class:`~kite.Quadtree.nleaves`)
"""
return num.linalg.inv(self.covariance_matrix)
@property_cached
def weight_matrix_L2(self):
""" Weight matrix from full covariance :math:`\\sqrt{cov^{-1}}`.
:type: :class:`numpy.ndarray`,
size (:class:`~kite.Quadtree.nleaves` x
:class:`~kite.Quadtree.nleaves`)
"""
incov = num.linalg.inv(self.covariance_matrix)
return sp.linalg.sqrtm(incov)
@property_cached
def weight_matrix_focal(self):
""" Approximated weight matrix from fast focal method
:math:`cov_{focal}^{-1}`.
:type: :class:`numpy.ndarray`,
size (:class:`~kite.Quadtree.nleaves` x
:class:`~kite.Quadtree.nleaves`)
"""
return num.linalg.inv(self.covariance_matrix_focal)
@property_cached
def weight_vector(self):
""" Weight vector from full covariance :math:`cov^{-1}`.
:type: :class:`numpy.ndarray`,
size (:class:`~kite.Quadtree.nleaves`)
"""
return num.sum(self.weight_matrix, axis=1)
@property_cached
def weight_vector_focal(self):
""" Weight vector from fast focal method
:math:`\\sqrt{cov_{focal}^{-1}}`.
:type: :class:`numpy.ndarray`,
size (:class:`~kite.Quadtree.nleaves`)
"""
return num.sum(self.weight_matrix_focal, axis=1)
def _calcCovarianceMatrix(self, method='focal', nthreads=None):
"""Constructs the covariance matrix.
:param method: Either ``focal`` point distances are used - this is
quick but only an approximation.
Or ``full``, where the full quadtree pixel distances matrices are
calculated , defaults to ``focal``
:type method: str, optional
:returns: Covariance matrix
:rtype: thon:numpy.ndarray
"""
self._initialized = True
if nthreads is None:
nthreads = self.nthreads
nl = len(self.quadtree.leaves)
self._leaf_mapping = {}
t0 = time.time()
if method == 'focal':
model = self.getModelFunction()
coords = self.quadtree.leaf_focal_points_meter
dist_matrix = num.sqrt(
(coords[:, 0] - coords[:, 0, num.newaxis])**2
+ (coords[:, 1] - coords[:, 1, num.newaxis])**2)
cov_matrix = model(dist_matrix, *self.covariance_model)
# adding variance
if self.variance < cov_matrix.max():
variance = cov_matrix.max()
else:
variance = self.variance
if self.quadtree.leaf_mean_px_var is not None:
self._log.debug(
'Adding variance from scene.displacement_px_var')
variance += self.quadtree.leaf_mean_px_var
num.fill_diagonal(cov_matrix, variance)
for nx, ny in num.nditer(num.triu_indices_from(dist_matrix)):
self._mapLeaves(nx, ny)
elif method == 'full':
leaf_map = num.empty((len(self.quadtree.leaves), 4),
dtype=num.uint32)
for nl, leaf in enumerate(self.quadtree.leaves):
leaf, _ = self._mapLeaves(nl, nl)
leaf_map[nl, 0], leaf_map[nl, 1] = (leaf._slice_rows.start,
leaf._slice_rows.stop)
leaf_map[nl, 2], leaf_map[nl, 3] = (leaf._slice_cols.start,
leaf._slice_cols.stop)
nleaves = self.quadtree.nleaves
cov_matrix = covariance_ext.covariance_matrix(
self.scene.frame.gridEmeter.filled(),
self.scene.frame.gridNmeter.filled(),
leaf_map,
self.covariance_model, self.variance,
nthreads, self.config.adaptive_subsampling)\
.reshape(nleaves, nleaves)
if self.quadtree.leaf_mean_px_var is not None:
self._log.debug(
'Adding variance from scene.displacement_px_var')
cov_matrix[num.diag_indices_from(cov_matrix)] +=\
self.quadtree.leaf_mean_px_var
else:
raise TypeError('Covariance calculation %s method not defined!'
% method)
self._log.debug('Created covariance matrix - %s mode [%0.4f s]' %
(method, time.time()-t0))
return cov_matrix
def isMatrixPosDefinite(self, full=False):
self._log.debug('Checking whether matrix is positive-definite')
if full:
matrix = self.covariance_matrix
else:
matrix = self.covariance_matrix_focal
try:
chol_decomp = num.linalg.cholesky(matrix)
except num.linalg.linalg.LinAlgError:
pos_def = False
else:
pos_def = ~num.all(num.iscomplex(chol_decomp))
finally:
if not pos_def:
self._log.warning('Covariance matrix is not positiv definite!')
return pos_def
@staticmethod
def _leafFocalDistance(leaf1, leaf2):
return num.sqrt((leaf1.focal_point[0]
- leaf2.focal_point[0])**2 +
(leaf1.focal_point[1]
- leaf2.focal_point[1])**2)
def _leafMapping(self, leaf1, leaf2):
if not isinstance(leaf1, str):
leaf1 = leaf1.id
if not isinstance(leaf2, str):
leaf2 = leaf2.id
if not self._initialized:
self.covariance_matrix_focal
try:
return self._leaf_mapping[leaf1], self._leaf_mapping[leaf2]
except KeyError as e:
raise KeyError('Unknown quadtree leaf with id %s' % e)
[docs] def getLeafCovariance(self, leaf1, leaf2):
"""Get the covariance between ``leaf1`` and ``leaf2`` from
distances.
:param leaf1: Leaf one
:type leaf1: str of `leaf.id` or :class:`~kite.quadtree.QuadNode`
:param leaf2: Leaf two
:type leaf2: str of `leaf.id` or :class:`~kite.quadtree.QuadNode`
:returns: Covariance between ``leaf1`` and ``leaf2``
:rtype: float
"""
return self.covariance_matrix[self._leafMapping(leaf1, leaf2)]
[docs] def getLeafWeight(self, leaf, model='focal'):
""" Get the total weight of ``leaf``, which is the summation of
all single pair weights of :attr:`kite.Covariance.weight_matrix`.
.. math ::
w_{x} = \\sum_i W_{x,i}
:param model: ``Focal`` or ``full``, default ``focal``
:type model: str
:param leaf: A leaf from :class:`~kite.Quadtree`
:type leaf: :class:`~kite.quadtree.QuadNode`
:returns: Weight of the leaf
:rtype: float
"""
(nl, _) = self._leafMapping(leaf, leaf)
weight_mat = self.weight_matrix_focal
return num.mean(weight_mat, axis=0)[nl]
[docs] def syntheticNoise(self, shape=(1024, 1024), dEdN=None,
anisotropic=False, rstate=None):
"""Create random synthetic noise from data noise power spectrum.
This function uses the power spectrum of the data noise
(:attr:`noise_data`) (:func:`powerspecNoise`) to create synthetic
noise, e.g. to use it for data pertubation in optinmizations.
The default sampling distances are taken from
:attr:`kite.scene.Frame.dE` and :attr:`kite.scene.Frame.dN`. They can
be overwritten.
:param shape: shape of the desired noise patch.
Pixels in northing and easting (`nE`, `nN`),
defaults to `(1024, 1024)`.
:type shape: tuple, optional
:param dEdN: The sampling distance in easting, defaults to
(:attr:`kite.scene.Frame.dE`, :attr:`kite.scene.Frame.dN`).
:type dE: tuple, floats
:returns: synthetic noise patch
:rtype: :class:`numpy.ndarray`
"""
if (shape[0] + shape[1]) % 2 != 0:
# self._log.warning('Patch dimensions must be even, '
# 'ceiling dimensions!')
pass
nE = shape[1] + (shape[1] % 2)
nN = shape[0] + (shape[0] % 2)
if rstate is None:
rstate = num.random.RandomState()
rfield = rstate.rand(nN, nE)
spec = num.fft.fft2(rfield)
if not dEdN:
dE, dN = (self.scene.frame.dE, self.scene.frame.dN)
kE = num.fft.fftfreq(nE, dE)
kN = num.fft.fftfreq(nN, dN)
k_rad = num.sqrt(kN[:, num.newaxis]**2 + kE[num.newaxis, :]**2)
amp = num.zeros_like(k_rad)
if not anisotropic:
noise_pspec, k, _, _, _, _ = self.powerspecNoise2D()
k_bin = num.insert(k + k[0]/2, 0, 0)
for i in range(k.size):
k_min = k_bin[i]
k_max = k_bin[i+1]
r = num.logical_and(k_rad > k_min, k_rad <= k_max)
if i == (k.size-1):
r = k_rad > k_min
if not num.any(r):
continue
amp[r] = noise_pspec[i]
amp[k_rad == 0.] = self.variance
amp[k_rad > k.max()] = noise_pspec[num.argmax(k)]
amp = num.sqrt(amp * self.noise_data.size * num.pi * 4)
elif anisotropic:
interp_pspec, _, _, _, skE, skN = self.powerspecNoise3D()
kE = num.fft.fftshift(kE)
kN = num.fft.fftshift(kN)
mkE = num.logical_and(kE >= skE.min(), kE <= skE.max())
mkN = num.logical_and(kN >= skN.min(), kN <= skN.max())
mkRad = num.where( # noqa
k_rad < num.sqrt(kN[mkN].max()**2 + kE[mkE].max()**2))
res = interp_pspec(kN[mkN, num.newaxis],
kE[num.newaxis, mkE], grid=True)
print((amp.shape, res.shape))
print((kN.size, kE.size))
amp = res
amp = num.fft.fftshift(amp)
print((amp.min(), amp.max()))
spec *= amp
noise = num.abs(num.fft.ifft2(spec))
noise -= num.mean(noise)
# remove shape % 2 padding
return noise[:shape[0], :shape[1]]
[docs] def getQuadtreeNoise(self, rstate=None, gather=num.nanmedian):
'''Create noise for a :class:`~kite.quadtree.Quadtree`
Use :meth:`~kite.covariance.Covariance.getSyntheticNoise` to create
data-driven noise on each quadtree leaf, summarized by
:param gather: Function gathering leaf's noise realisation,
defaults to num.median.
:type normalisation: callable, optional
:returns: Array of noise level at each quadtree leaf.
:rtype: :class:`numpy.ndarray`
'''
qt = self.quadtree
syn_noise = self.syntheticNoise(
shape=self.scene.displacement.shape,
rstate=rstate)
syn_noise[self.scene.displacement_mask] = num.nan
noise_quadtree_arr = num.full(qt.nleaves, num.nan)
for il, lv in enumerate(qt.leaves):
noise_quadtree_arr[il] = gather(
syn_noise[lv._slice_rows, lv._slice_cols])
return noise_quadtree_arr
def powerspecNoise1D(self, data=None, ndeg=512, nk=512):
if self._powerspec1d_cached is None:
self._powerspec1d_cached = self._powerspecNoise(
data, norm='1d', ndeg=ndeg, nk=nk)
return self._powerspec1d_cached
def powerspecNoise2D(self, data=None, ndeg=512, nk=512):
if self._powerspec2d_cached is None:
self._powerspec2d_cached = self._powerspecNoise(
data, norm='2d', ndeg=ndeg, nk=nk)
return self._powerspec2d_cached
def powerspecNoise3D(self, data=None):
if self._powerspec3d_cached is None:
self._powerspec3d_cached = self._powerspecNoise(
data, norm='3d')
return self._powerspec3d_cached
def _powerspecNoise(self, data=None, norm='1d', ndeg=512, nk=512):
"""Get the noise power spectrum from
:attr:`kite.Covariance.noise_data`.
:param data: Overwrite Covariance.noise_data, defaults to `None`
:type data: :class:`numpy.ndarray`, optional
:returns: `(power_spec, k, f_spectrum, kN, kE)`
:rtype: tuple
"""
if data is None:
noise = self.noise_data
else:
noise = data.copy()
if norm not in ['1d', '2d', '3d']:
raise AttributeError('norm must be either 1d, 2d or 3d')
# noise = squareMatrix(noise)
shift = num.fft.fftshift
spectrum = shift(num.fft.fft2(noise, axes=(0, 1), norm=None))
power_spec = (num.abs(spectrum)/spectrum.size)**2
kE = shift(num.fft.fftfreq(power_spec.shape[1],
d=self.quadtree.frame.dE))
kN = shift(num.fft.fftfreq(power_spec.shape[0],
d=self.quadtree.frame.dN))
k_rad = num.sqrt(kN[:, num.newaxis]**2 + kE[num.newaxis, :]**2)
power_spec[k_rad == 0.] = 0.
power_interp = sp.interpolate.RectBivariateSpline(kN, kE, power_spec)
# def power1d(k):
# theta = num.linspace(-num.pi, num.pi, ndeg, False)
# power = num.empty_like(k)
# for i in range(k.size):
# kE = num.cos(theta) * k[i]
# kN = num.sin(theta) * k[i]
# power[i] = num.median(power_interp.ev(kN, kE)) * k[i]\
# * num.pi * 4
# return power
def power1d(k):
theta = num.linspace(-num.pi, num.pi, ndeg, False)
power = num.empty_like(k)
for i in range(k.size):
kE = num.cos(theta) * k[i]
kN = num.sin(theta) * k[i]
power[i] = num.median(power_interp.ev(kN, kE))
return power
def power2d(k):
""" Mean 2D Power works! """
theta = num.linspace(-num.pi, num.pi, ndeg, False)
power = num.empty_like(k)
for i in range(k.size):
kE = num.sin(theta) * k[i]
kN = num.cos(theta) * k[i]
power[i] = num.median(power_interp.ev(kN, kE))
# Median is more stable than the mean here
return power
def power3d(k):
return power_interp
power = power1d
if norm == '2d':
power = power2d
elif norm == '3d':
power = power3d
k_rad = num.sqrt(kN[:, num.newaxis]**2 + kE[num.newaxis, :]**2)
k = num.linspace(k_rad[k_rad > 0].min(),
k_rad.max(), nk)
dk = 1./k.min() / (2. * nk)
return power(k), k, dk, spectrum, kE, kN
def _powerCosineTransform(self, p_spec):
"""Calculating the cosine transform of the power spectrum.
The cosine transform of the power spectrum is an estimate
of the data covariance (see Hanssen, 2001)."""
cos = sp.fftpack.idct(p_spec, type=3)
return cos
[docs] def setSamplingMethod(self, method):
""" Set the sampling method """
assert method in CovarianceConfig.sampling_method.choices
self.config.sampling_method = method
self._clear(config=True, spectrum=False)
self.evChanged.notify()
self._log.debug('Changed sampling method to %s' % method)
[docs] def setSpatialBins(self, nbins):
""" Set number of spatial bins """
self.config.spatial_bins = nbins
self._clear(config=True, spectrum=False)
self.evChanged.notify()
self._log.debug('Changed spatial distance bins to %s' % nbins)
[docs] def setSpatialPairs(self, npairs):
""" Set number of random spatial pairs """
self.config.spatial_pairs = npairs
self._clear(config=True, spectrum=False)
self.evChanged.notify()
self._log.debug('Changed random pairs to %s' % npairs)
def setModelFunction(self, model):
assert model in CovarianceConfig.model_function.choices
self.config.model_function = model
self._clear(config=True, spectrum=True)
self.evChanged.notify()
self._log.debug('Changed model function to %s' % model)
def getModelFunction(self):
if self.config.model_function == 'exponential':
return modelCovarianceExponential
if self.config.model_function == 'exponential_cosine':
return modelCovarianceExponentialCosine
@property_cached
def covariance_spectral(self):
""" Covariance function estimated directly from the power spectrum of
displacement noise patch using the cosine transform.
:type: tuple, :class:`numpy.ndarray` (covariance, distance) """
power_spec, k, dk, _, _, _ = self.powerspecNoise1D()
# power_spec -= self.variance
d = num.arange(1, power_spec.size+1) * dk
cov = self._powerCosineTransform(power_spec)
return cov, d
@property_cached
def covariance_spatial(self):
self._log.debug('Estimating covariance (spatial)...')
nbins = self.config.spatial_bins
npairs = self.config.spatial_pairs
noise_data = self.noise_data.ravel()
grdE = self.noise_data_gridE
grdN = self.noise_data_gridN
max_distance = min(abs(grdE.min() - grdE.max()),
abs(grdN.min() - grdN.max()))
dist_bins = num.linspace(0, max_distance, nbins + 1)
grdE = grdE.ravel()
grdN = grdN.ravel()
# Select random coordinates
rstate = num.random.RandomState(noise_data.size)
rand_idx = rstate.randint(0, noise_data.size, (2, npairs))
idx0 = rand_idx[0, :]
idx1 = rand_idx[1, :]
distances = num.sqrt((grdN[idx0] - grdN[idx1])**2 +
(grdE[idx0] - grdE[idx1])**2)
cov_all = noise_data[idx0] * noise_data[idx1]
vario_all = (noise_data[idx0] - noise_data[idx1])**2
bins = num.digitize(distances, dist_bins, right=True)
bin_distances = dist_bins[1:] - dist_bins[1]/2
covariance = num.full_like(bin_distances, fill_value=num.nan)
variance = num.full_like(bin_distances, fill_value=num.nan)
for ib in range(nbins):
selection = bins == ib
if selection.sum() != 0:
covariance[ib] = num.nanmean(cov_all[selection])
variance[ib] = num.nanmean(vario_all[selection])/2
self._structure_spatial = (variance[~num.isnan(variance)],
bin_distances[~num.isnan(variance)])
covariance[0] = num.nan
return (covariance[~num.isnan(covariance)],
bin_distances[~num.isnan(covariance)])
[docs] def getCovariance(self):
""" Calculate the covariance function
:return: The covariance and distance
:rtype: tuple
"""
if self.config.sampling_method == 'spatial':
return self.covariance_spatial
elif self.config.sampling_method == 'spectral':
return self.covariance_spectral
@property
def covariance_model(self, regime=0):
""" Covariance model parameters for
:func:`~kite.covariance.modelCovariance` retrieved
from :attr:`~kite.Covariance.getCovarianceFunction`.
.. note:: using this function implies several model
fits: (1) fit of the spectrum and (2) the cosine transform.
Not sure about the consequences, if this is useful and/or
meaningful.
:getter: Get the parameters.
:type: tuple, ``a`` and ``b``
"""
if self.config.model_coefficients is None:
covariance, distance = self.getCovariance()
model = self.getModelFunction()
if self.config.model_function == 'exponential':
coeff = (num.mean(covariance), num.mean(distance))
elif self.config.model_function == 'exponential_cosine':
coeff = (num.mean(covariance), num.mean(distance),
num.mean(distance)*-.1, .1)
func = self.getModelFunction()
# Testing penalty function
def model(*args):
distance, a, b, c, d = args
res = func(*args)
penalty = 0.
if distance[-1]/b > (distance[-1]+c)/d:
penalty = (b-d) * coeff[0]
self._log.warning('Penalty %f' % penalty)
return res + penalty
# Overwrite with pure model function
model = self.getModelFunction() # noqa
try:
coeff, _ = sp.optimize.curve_fit(
model,
distance,
covariance,
p0=coeff)
except RuntimeError:
self._log.warning('Could not fit the %s'
' covariance model'
% self.config.model_function)
finally:
self.config.model_coefficients = tuple(map(float, coeff))
return self.config.model_coefficients
@property
def covariance_model_rms(self):
"""
:getter: RMS missfit between :class:`~kite.Covariance.covariance_model`
and :class:`~kite.Covariance.getCovarianceFunction`
:type: float
"""
cov, d = self.getCovariance()
model = self.getModelFunction()
cov_mod = model(d, *self.covariance_model)
return num.sqrt(num.mean((cov - cov_mod)**2))
@property_cached
def structure_spatial(self):
self.covariance_spatial
return self._structure_spatial
@property_cached
def structure_spectral(self):
""" Structure function derived from ``noise_patch``
:type: tuple, :class:`numpy.ndarray` (structure_spectral, distance)
Adapted from
http://clouds.eos.ubc.ca/~phil/courses/atsc500/docs/strfun.pdf
"""
power_spec, k, dk, _, _, _ = self.powerspecNoise1D()
d = num.arange(1, power_spec.size+1) * dk
def structure_spectral(power_spec, d, k):
struc_func = num.zeros_like(k)
for i, d in enumerate(d):
for ik, tk in enumerate(k):
# struc_func[i] += (1. - num.cos(tk*d))*power_spec[ik]
struc_func[i] += (1. - sp.special.j0(tk*d))*power_spec[ik]
struc_func *= 2./1
return struc_func
struc_func = structure_spectral(power_spec, d, k)
return struc_func, d
[docs] def getStructure(self, method=None):
""" Get the structure function
:param method: Either `spatial` or `spectral`, if `None`
the method is taken from config
:type method: str (optional)
:return: (variance, distance)
:rtype: tuple
"""
if method is None:
method = self.config.sampling_method
if method == 'spatial':
return self.structure_spatial
elif method == 'spectral':
return self.structure_spectral
@property
def variance(self):
""" Variance of data noise estimated from the
high-frequency end of power spectrum.
:setter: Set the variance manually
:getter: Retrieve the variance
:type: float
"""
return self.config.variance
@variance.setter
def variance(self, value):
self.config.variance = float(value)
# self._clear(config=False, spectrum=False, spatial=False)
self.evChanged.notify()
@variance.getter
def variance(self):
if self.config.variance is None and \
self.config.sampling_method == 'spatial':
structure_spatial, dist = self.structure_spatial
last_20p = -int(structure_spatial.size * .2)
self.config.variance = float(
num.mean(structure_spatial[(last_20p):]))
elif (self.config.variance is None and
self.config.sampling_method == 'spectral'):
power_spec, k, dk, spectrum, _, _ = self.powerspecNoise1D()
cov, _ = self.covariance_spectral
ma = self.covariance_model[0]
# print(cov[1])
ps = power_spec * spectrum.size
# print(spectrum.size)
# print(num.mean(ps[-int(ps.size/9.):-1]))
var = num.median(ps[-int(ps.size/9.):]) + ma
self.config.variance = float(var)
return self.config.variance
[docs] def export_weight_matrix(self, filename):
""" Export the full :attr:`~kite.Covariance.weight_matrix` to an ASCII
file. The data can be loaded through :func:`numpy.loadtxt`.
:param filename: path to export to
:type filename: str
"""
self._log.debug('Exporting Covariance.weight_matrix to %s' % filename)
header = 'Exported kite.Covariance.weight_matrix, '\
'for more information visit https://pyrocko.org\n'\
'\nThe matrix is symmetric and ordered by QuadNode.id:\n'
header += ', '.join([l.id for l in self.quadtree.leaves])
num.savetxt(filename, self.weight_matrix, header=header)
@property_cached
def plot(self):
""" Simple overview plot to summarize the covariance estimations.
"""
from kite.plot2d import CovariancePlot
return CovariancePlot(self)
@property_cached
def plot_syntheticNoise(self):
""" Simple overview plot to summarize the covariance estimations.
"""
from kite.plot2d import SyntheticNoisePlot
return SyntheticNoisePlot(self)
