def FLCF(volume,
template,
mask=None,
stdV=None,
gpu=False,
mempool=None,
pinned_mempool=None):
'''Fast local correlation function
@param volume: target volume
@param template: template to be searched. It can have smaller size then target volume.
@param mask: template mask. If not given, a default sphere mask will be used.
@param stdV: standard deviation of the target volume under mask, which do not need to be calculated again when the mask is identical.
@return: the local correlation function
'''
meanT = meanUnderMask(template, mask, gpu=gpu)
temp = ((template - meanT) /
stdUnderMask(template, mask, meanT, gpu=gpu)) * mask
res = xp.fft.fftshift(
xp.fft.ifftn(xp.conj(xp.fft.fftn(temp)) *
xp.fft.fftn(volume))).real / stdV / mask.sum()
meanT = None
temp = None
mempool.free_all_blocks()
pinned_mempool.free_all_blocks()
def normalised_cross_correlation_mask(first, second, mask):
"""
Do cross correlation with a running mask based on numpy
@param first: The first dataset (numpy 2D)
@type first: numpy array 2D
@param second: The second dataset (numpy 2D)
@type second: numpy array 2D
@param mask: The mask
@type mask: numpy array 2D
@return: The cross correlation result
@returntype: numpy array 2D
@requires: the shape of first to be equal to the shape of second and the shape of the mask
"""
# assert first.shape == second.shape
# assert first.shape == mask.shape
# assert len(first.shape) == 2
import xp.fft as nf
a = norm_inside_mask(first, mask)
b = norm_inside_mask(second, mask)
return xp.real(
nf.fftshift(nf.ifftn(xp.multiply(nf.fftn(b), xp.conj(
nf.fftn(a)))))) / xp.sum(mask)
def normalised_cross_correlation(first, second, filter_mask=None, device=0):
"""
Do a cross correlation based on numpy
@param first: The first dataset (numpy 2D)
@type first: numpy array 2D
@param second: The second dataset (numpy 2D)
@type second: numpy array 2D
@param filter_mask: a filter which is used to filter image 'first'
@type filter_mask: numpy array 2D
@return: The cross correlation result
@returntype: numpy array 2D
@requires: the shape of first to be equal to the shape of second, and equal t the shape of the filter (if used of course)
"""
assert first.shape == second.shape
assert len(first.shape) == 2
if not (filter_mask is None): assert first.shape == filter_mask.shape
# if filter_mask is None:
ffirst = xp.fft.fft2(xp.array(first))
# else:
# ffirst = xp.fft.fftshift(xp.fft.fftshift(xp.fft.fftn(first)) * filter_mask)
ccmap = xp.real(
xp.fft.fftshift(
xp.fft.ifftn(xp.multiply(xp.fft.fftn(second),
xp.conj(ffirst))))) / first.size
return ccmap
def xcf(
volume,
template,
mask=None,
stdV=None,
):
"""
XCF: returns the non-normalised cross correlation function. The xcf
result is scaled only by the square of the number of elements.
@param volume : The search volume
@type volume: L{pytom_volume.vol}
@param template : The template searched (this one will be used for conjugate complex multiplication)
@type template: L{pytom_volume.vol}
@param mask: Will be unused, only for compatibility reasons with FLCF
@param stdV: Will be unused, only for compatibility reasons with FLCF
@return: XCF volume
@rtype: L{pytom_volume.vol}
@author: Thomas Hrabe
"""
# if mask:
# volume = volume * mask
# template = template * mask
# determine fourier transforms of volumes
if template.dtype in (xp.float64, xp.float32):
fvolume = xp.fft.fftn(volume)
else:
fvolume = volume
if template.dtype in (xp.float32, xp.float64):
ftemplate = xp.fft.fftn(template)
else:
ftemplate = template
# perform element wise - conjugate multiplication
ftemplate = xp.conj(ftemplate)
fresult = fvolume * ftemplate
# transform back to real space
result = abs(xp.fft.ifftn(fresult))
#xp.fft.iftshift(result)
return result
def meanVolUnderMask(volume, mask):
"""
meanUnderMask: calculate the mean volume under the given mask (Both should have the same size)
@param volume: input volume
@type volume: L{numpy.ndarray} L{cupy.ndarray}
@param mask: mask
@type mask: L{numpy.ndarray} L{cupy.ndarray}
@param p: non zero value numbers in the mask
@type p: L{int} or L{float}
@param gpu: Boolean that indicates if the input volumes stored on a gpu.
@type gpu: L{bool}
@return: the calculated mean volume under mask
@rtype: L{numpy.ndarray} L{cupy.ndarray}
@author: Gijs van der Schot
"""
res = xp.fft.fftshift(
xp.fft.irfftn(
xp.fft.rfftn(volume) * xp.conj(xp.fft.rfftn(mask)))) / mask.sum()
return res.real
def meanVolUnderMaskPlanned(volume, mask, ifftnP, fftnP, plan):
"""
meanUnderMask: calculate the mean volume under the given mask (Both should have the same size)
@param volume: input volume
@type volume: L{numpy.ndarray} L{cupy.ndarray}
@param mask: mask
@type mask: L{numpy.ndarray} L{cupy.ndarray}
@param p: non zero value numbers in the mask
@type p: L{int} or L{float}
@param gpu: Boolean that indicates if the input volumes stored on a gpu.
@type gpu: L{bool}
@return: the calculated mean volume under mask
@rtype: L{numpy.ndarray} L{cupy.ndarray}
@author: Gijs van der Schot
"""
volume_fft = fftnP(volume.astype(xp.complex64), plan=plan)
mask_fft = fftnP(mask.astype(xp.complex64), plan=plan)
res = xp.fft.fftshift(ifftnP(volume_fft * xp.conj(mask_fft),
plan=plan)) / mask.sum()
return res.real
def normalised_cross_correlation(first, second, filter_mask=None):
"""
Do a cross correlation based on numpy
@param first: The first dataset (numpy 2D)
@type first: numpy array 2D
@param second: The second dataset (numpy 2D)
@type second: numpy array 2D
@param filter_mask: a filter which is used to filter image 'first'
@type filter_mask: numpy array 2D
@return: The cross correlation result
@returntype: numpy array 2D
@requires: the shape of first to be equal to the shape of second, and equal t the shape of the filter (if used of course)
"""
import sys
from pytom.tompy.io import write
assert first.shape == second.shape
assert len(first.shape) == 2
if not (filter_mask is None): assert first.shape == filter_mask.shape
from numpy.fft import fftshift, ifftn, fftn
import numpy as xp
prod = 1
for d in first.shape:
prod *= d
if filter_mask is None:
ffirst = xp.fft.fftn(first)
fsecond = xp.fft.fftn(second)
else:
ffirst = ((xp.fft.fftn(first)) * xp.fft.fftshift(filter_mask))
fsecond = xp.fft.fftn(second) * xp.fft.fftshift(filter_mask)
ccmap = xp.real(xp.fft.fftshift(xp.fft.ifftn(
fsecond * xp.conj(ffirst)))) / filter_mask.sum()
# write('real_ccmap.mrc', ccmap)
return ccmap
def fourier_reduced2full(data, isodd=False):
"""Return an Hermitian symmetried data.
"""
# get the original shape
sx = data.shape[0]
sy = data.shape[1]
if isodd:
sz = (data.shape[2] - 1) * 2 + 1
else:
sz = (data.shape[2] - 1) * 2
res = xp.zeros((sx, sy, sz), dtype=data.dtype)
res[:, :, 0:data.shape[2]] = data
# calculate the coodinates accordingly
szz = sz - data.shape[2]
x, y, z = xp.indices((sx, sy, szz))
ind = [xp.mod(sx - x, sx), xp.mod(sy - y, sy), szz - z]
# do the complex conjugate of the second part
res[:, :, data.shape[2]:] = xp.conj(data[tuple(ind)])
return res
def bandCC(volume, reference, band, verbose=False, shared=None, index=None):
"""
bandCC: Determines the normalised correlation coefficient within a band
@param volume: The volume
@type volume: L{pytom_volume.vol}
@param reference: The reference
@type reference: L{pytom_volume.vol}
@param band: [a,b] - specify the lower and upper end of band.
@return: First parameter - The correlation of the two volumes in the specified band.
Second parameter - The bandpass filter used.
@rtype: List - [float,L{pytom_freqweight.weight}]
@author: Thomas Hrabe
"""
if not index is None: print(index)
from pytom.tompy.filter import bandpass
from pytom.tompy.correlation import xcf
#from pytom.tompy.filter import vol_comp
if verbose:
print('lowest freq : ', band[0], ' highest freq', band[1])
vf, m = bandpass(volume,
band[0],
band[1],
returnMask=True,
fourierOnly=True)
rf = bandpass(reference, band[0], band[1], mask=m,
fourierOnly=True) #,vf[1])
#ccVolume = vol_comp(rf[0].shape[0],rf[0].shape[1],rf[0].shape[2])
#ccVolume.copyVolume(rf[0])
vf = vf.astype(xp.complex128)
ccVolume = rf.astype(vf.dtype)
ccVolume = ccVolume * xp.conj(vf)
#pytom_volume.conj_mult(ccVolume,vf[0])
cc = ccVolume.sum()
cc = cc.real
v = vf
r = rf
absV = xp.abs(v)
absR = xp.abs(r)
sumV = xp.sum(absV**2)
sumR = xp.sum(absR**2)
sumV = xp.abs(sumV)
sumR = xp.abs(sumR)
if sumV == 0:
sumV = 1
if sumR == 0:
sumR = 1
cc = cc / (xp.sqrt(sumV * sumR))
#numerical errors will be punished with nan
if abs(cc) > 1.1:
cc = float('nan')
return float(cc)
