python-microscopy 25.11.20__cp313-cp313-win_amd64.whl

PYME/Deconv/dec.py

@@ -0,0 +1,659 @@
+#!/usr/bin/python
+
+##################
+# dec.py
+#
+# Copyright David Baddeley, 2009
+# d.baddeley@auckland.ac.nz
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program.  If not, see <http://www.gnu.org/licenses/>.
+#
+##################
+
+from scipy.fftpack import fftn, ifftn, fftshift, ifftshift
+from scipy import ndimage
+import numpy as np
+
+import PYME.misc.fftw_compat as fftw3f
+from .wiener import resizePSF
+
+from . import fftwWisdom
+fftwWisdom.load_wisdom()
+
+NTHREADS = 8
+FFTWFLAGS = ['measure']
+
+
+
+class ICTMDeconvolution(object):
+    """Base deconvolution class, implementing a variant of the ICTM algorithm.
+    ie. find f such that:
+       ||Af-d||^2 + lamb^2||L(f - fdef)||^2
+    is minimised
+    
+    Note that this is nominally for Gaussian distributed noise, although can be
+    adapted by adding a weighting to the misfit term.
+
+    Derived classed should additionally define the following methods:
+    AFunc - the forward mapping (computes Af)
+    AHFunc - conjugate transpose of forward mapping (computes \bar{A}^T f)
+    LFunc - the likelihood function
+    LHFunc - conj. transpose of likelihood function
+
+    see dec_conv for an implementation of conventional image deconvolution with a
+    measured, spatially invariant PSF
+    """
+    def __init__(self):
+        #allocate some empty lists to track our progress in
+        self.tests=[]
+        self.ress = []
+        self.prefs = []
+        
+    def subsearch(self, f0, res, fdef, Afunc, Lfunc, lam, S):
+        """minimise in subspace - this is the bit which gets called on each iteration
+        to work out what the next step is going to be. See Inverse Problems text for details.
+        """
+        nsrch = np.size(S,1)
+        pref = Lfunc(f0-fdef)
+        w0 = np.dot(pref, pref)
+        c0 = np.dot(res,res)
+
+        AS = np.zeros((np.size(res), nsrch), 'f')
+        LS = np.zeros((np.size(pref), nsrch), 'f')
+
+        for k in range(nsrch):
+            AS[:,k] = Afunc(S[:,k])[self.mask]
+            LS[:,k] = Lfunc(S[:,k])
+
+        Hc = np.dot(np.transpose(AS), AS)
+        Hw = np.dot(np.transpose(LS), LS)
+        Gc = np.dot(np.transpose(AS), res)
+        Gw = np.dot(np.transpose(-LS), pref)
+
+        c = np.linalg.solve(Hc + pow(lam, 2)*Hw, Gc + pow(lam, 2)*Gw)
+
+        cpred = c0 + np.dot(np.dot(np.transpose(c), Hc), c) - np.dot(np.transpose(c), Gc)
+        wpred = w0 + np.dot(np.dot(np.transpose(c), Hw), c) - np.dot(np.transpose(c), Gw)
+
+        fnew = f0 + np.dot(S, c)
+
+        return (fnew, cpred, wpred)
+
+    def startGuess(self, data):
+        """starting guess for deconvolution - can be overridden in derived classes
+        but the data itself is usually a pretty good guess.
+        """
+        return data.copy()
+
+    def deconvp(self, args):
+        """ convenience function for deconvolving in parallel using processing.Pool.map"""
+        self.deconv(*args)
+    
+    def deconv(self, data, lamb, num_iters=10, weights=1, alpha=None, bg=0, pos=True):
+        """This is what you actually call to do the deconvolution.
+        parameters are:
+
+        data - the raw data
+        lamb - the regularisation parameter
+        num_iters - number of iterations (note that the convergence is fast when
+                    compared to many algorithms - e.g Richardson-Lucy - and the
+                    default of 10 will usually already give a reasonable result)
+        weights - a weighting on the residuals to deweight missing data and/or adjust the
+                  noise behaviour
+        alpha - PSF phase - hacked in for variable phase 4Pi deconvolution, should
+                really be refactored out into the dec_4pi classes.
+        bg - Dummy variable
+        pos - Flag to turn positivity constraints on/off
+        """
+        #remember what shape we are
+        self.dataShape = data.shape
+        
+        if 'prep' in dir(self) and not '_F' in dir(self):
+            self.prep()
+
+        if not np.isscalar(weights):
+            self.mask = weights > 0
+        else:
+            self.mask = np.isfinite(data.ravel())
+
+        #if doing 4Pi dec, do some phase related precomputation
+        #TODO move this to the 4Pi_dec classes
+        if (not alpha is None):
+            self.alpha = alpha
+            self.e1 = fftshift(np.exp(1j*self.alpha))
+            self.e2 = fftshift(np.exp(2j*self.alpha))
+
+        #guess a starting estimate for the object
+        self.f = self.startGuess(data).ravel()
+
+        self.fs = self.f.reshape(self.shape)
+
+        #make things 1 dimensional
+        #self.f = self.f.ravel()
+        data = data.ravel()
+        self.res = 0*data
+        
+        if not np.isscalar(weights):
+            weights = weights / weights.mean()
+
+        #use 0 as the default solution - should probably be refactored like the starting guess
+        fdef = np.zeros(self.f.shape, 'f')
+
+        #initial search directions
+        S = np.zeros((np.size(self.f), 3), 'f')
+
+        #number of search directions
+        nsrch = 2
+        self.loopcount = 0
+
+        while (self.loopcount  < num_iters) and (not self._stop_cond()):
+            self.loopcount += 1
+            #the direction our prior/ Likelihood function wants us to go
+            pref = self.Lfunc(self.f - fdef)
+            
+            #print pref, self.f
+
+            #the residuals
+            #if you want to bodge non-gaussian noise you can multiply with
+            #a weighting function - eg: 1/sqrt(data + eps))
+            #note that 1/sqrt(data) by itself is not a good idea as it will give
+            #infinite weight to zeros. As most devices have some form of readout noise
+            #justifying the eps shouldn't be too tricky
+            self.res[:] = (weights*(data - self.Afunc(self.f)))
+            
+            #resulting search directions
+            #note that the use of the Likelihood fuction/prior as a search direction
+            #is where this method departs from the classical conjugate gradient approach
+            S[:,0] = self.Ahfunc(self.res)
+            S[:,1] = -self.Lhfunc(pref)
+            
+            #print norm(S[:,0]), norm(S[:,1])
+
+            #check to see if the two search directions are orthogonal
+            #this can be used as a measure of convergence and a stopping criteria
+            test = 1 - abs(np.dot(S[:,0], S[:,1])/(np.linalg.norm(S[:,0])*np.linalg.norm(S[:,1])))
+
+            #print & log some statistics
+            print(('Test Statistic %f' % (test,)))
+            self.tests.append(test)
+            self.ress.append(np.linalg.norm(self.res))
+            self.prefs.append(np.linalg.norm(pref))
+
+            #minimise along search directions to find new estimate
+            (fnew, cpred, spred) = self.subsearch(self.f, self.res[self.mask], fdef, self.Afunc, self.Lfunc, lamb, S[:, 0:nsrch])
+
+            #positivity constraint (not part of original algorithm & could be ommitted)
+            
+            if pos:
+                fnew = (fnew*(fnew > 0))
+
+            #add last step to search directions, as per classical conj. gradient
+            S[:,2] = (fnew - self.f)
+            nsrch = 3
+
+            #set the current estimate to out new estimate
+            self.f[:] = fnew
+
+        return np.real(self.fs)
+
+    def _stop_cond(self):
+        """Optional stopping condition to end deconvolution early."""
+        return False
+        
+    def sim_pic(self,data,alpha):
+        """Do the forward transform to simulate a picture. Currently with 4Pi cruft."""
+        self.alpha = alpha
+        self.e1 = fftshift(np.exp(1j*self.alpha))
+        self.e2 = fftshift(np.exp(2j*self.alpha))
+        
+        return self.Afunc(data)
+
+
+class DeconvMappingBase(object):
+    """
+    Base class for different types of deconvolution. Ultimate deconvolution classes will inherit from both this and a
+    a method class (e.g. ICTM Deconvolution)
+    
+    Provides implementations of the following:
+    
+    AFunc - the forward mapping (computes Af)
+    AHFunc - conjugate transpose of forward mapping (computes \bar{A}^T f)
+    LFunc - the likelihood function
+    LHFunc - conj. transpose of likelihood function
+    
+    Also, defines:
+    
+    psf_calc - does any pre-computation to get the PSF into a usable form
+    prep - allocate memory / setup FFTW plans etc ...
+    
+    """
+    
+    def Afunc(self, f):
+        """ The forward mapping"""
+        raise NotImplementedError('Must be over-ridden in derived class')
+    
+    def Ahfunc(self, f):
+        """ The conjugate transpose of the forward mapping"""
+        raise NotImplementedError('Must be over-ridden in derived class')
+    
+    def Lfunc(self, f):
+        """ The likelihood function (ICTM deconvolution only)"""
+        raise NotImplementedError('Must be over-ridden in derived class')
+    
+    def Lhfunc(self, f):
+        """ The gonjugate transpose of likelihood function (ICTM deconvolution only)"""
+        raise NotImplementedError('Must be over-ridden in derived class')
+    
+    def psf_calc(self, psf, data_size):
+        """
+        do any pre-computation on the PSF. e.g. resizing to match the data shape and/or pre-calculating and
+        storing the OTF
+        """
+        raise NotImplementedError('Must be over-ridden in derived class')
+    
+    def prep(self):
+        """
+        Allocate memory and compute FFTW plans etc (this is separate from psf_calc as an aid to distributing Deconvolution
+        objects for parallel processing - psf_calc gets called before Deconv objects are passed, prep gets called when
+        the deconvolution is run.
+
+        """
+
+class ClassicMappingNP(DeconvMappingBase):
+    """Classical deconvolution with a stationary PSF - uses numpy FTTs rather than FFTW"""
+    def psf_calc(self, psf, data_size):
+        """Precalculate the OTF etc..."""
+
+        g = resizePSF(psf, data_size)
+
+        #keep track of our data shape
+        self.height = data_size[0]
+        self.width  = data_size[1]
+        self.depth  = data_size[2]
+
+        self.shape = data_size
+
+        self.g = g
+
+        #calculate OTF and conjugate transformed OTF
+        self.H = (fftn(g))
+        self.Ht = g.size*(ifftn(g))
+
+
+    def Afunc(self, f):
+        """Forward transform - convolve with the PSF"""
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+
+        F = fftn(fs)
+
+        d = ifftshift(ifftn(F*self.H))
+
+        d = np.real(d)
+        return np.ravel(d)
+
+    def Ahfunc(self, f):
+        """Conjugate transform - convolve with conj. PSF"""
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+
+        F = fftn(fs)
+        d = ifftshift(ifftn(F*self.Ht))
+        d = np.real(d)
+        return np.ravel(d)
+    
+    def Lfunc(self, f):
+        """convolve with an approximate 2nd derivative likelihood operator in 3D.
+        i.e. [[[0,0,0][0,1,0][0,0,0]],[[0,1,0][1,-6,1][0,1,0]],[[0,0,0][0,1,0][0,0,0]]]
+        """
+        #make our data 3D again
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+        a = -6*fs
+
+        a[:,:,0:-1] += fs[:,:,1:]
+        a[:,:,1:] += fs[:,:,0:-1]
+
+        a[:,0:-1,:] += fs[:,1:,:]
+        a[:,1:,:] += fs[:,0:-1,:]
+
+        a[0:-1,:,:] += fs[1:,:,:]
+        a[1:,:,:] += fs[0:-1,:,:]
+
+        #flatten data again
+        return np.ravel(np.cast['f'](a))
+
+    Lhfunc=Lfunc
+
+class dec_conv_slow(ICTMDeconvolution, ClassicMappingNP):
+    def __init__(self, *args, **kwargs):
+        ICTMDeconvolution.__init__(self, *args, **kwargs)
+
+class ClassicMappingFFTW(DeconvMappingBase):
+    """Classical deconvolution with a stationary PSF using FFTW for convolutions"""
+    def prep(self):
+        #allocate memory
+        self._F = fftw3f.create_aligned_array(self.FTshape, 'complex64')
+        self._r = fftw3f.create_aligned_array(self.shape, 'f4')
+
+        print('Creating plans for FFTs - this might take a while')
+        #calculate plans for other ffts
+        self._plan_r_F = fftw3f.Plan(self._r, self._F, 'forward', flags = FFTWFLAGS, nthreads=NTHREADS)
+        self._plan_F_r = fftw3f.Plan(self._F, self._r, 'backward', flags = FFTWFLAGS, nthreads=NTHREADS)
+        
+        fftwWisdom.save_wisdom()
+
+        print('Done planning')
+    
+    def psf_calc(self, psf, data_size):
+        """Precalculate the OTF etc..."""
+        
+        g = resizePSF(psf, data_size)
+
+        #keep track of our data shape
+        self.height = data_size[0]
+        self.width  = data_size[1]
+        self.depth  = data_size[2]
+
+        self.shape = data_size
+        
+        print('Calculating OTF')
+
+        self.FTshape = [self.shape[0], self.shape[1], int(self.shape[2]/2 + 1)]
+
+        self.g = g.astype('f4')
+        self.g2 = 1.0*self.g[::-1, ::-1, ::-1]
+
+        #allocate memory
+        self.H = fftw3f.create_aligned_array(self.FTshape, 'complex64')
+        self.Ht = fftw3f.create_aligned_array(self.FTshape, 'complex64')
+
+        #create plans & calculate OTF and conjugate transformed OTF
+        fftw3f.Plan(self.g, self.H, 'forward')()
+        fftw3f.Plan(self.g2, self.Ht, 'forward')()
+
+        self.Ht /= g.size
+        self.H /= g.size
+
+
+
+
+    def Lfunc(self, f):
+        """convolve with an approximate 2nd derivative likelihood operator in 3D.
+        i.e. [[[0,0,0][0,1,0][0,0,0]],[[0,1,0][1,-6,1][0,1,0]],[[0,0,0][0,1,0][0,0,0]]]
+        """
+        #make our data 3D again
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+        a = -6*fs
+
+        a[:,:,0:-1] += fs[:,:,1:]
+        a[:,:,1:] += fs[:,:,0:-1]
+
+        a[:,0:-1,:] += fs[:,1:,:]
+        a[:,1:,:] += fs[:,0:-1,:]
+
+        a[0:-1,:,:] += fs[1:,:,:]
+        a[1:,:,:] += fs[0:-1,:,:]
+
+        #flatten data again
+        return np.ravel(np.cast['f'](a))
+
+    Lhfunc=Lfunc
+
+    def Afunc(self, f):
+        """Forward transform - convolve with the PSF"""
+        #fs = reshape(f, (self.height, self.width, self.depth))
+        self._r[:] = f.reshape(self._r.shape)
+
+        #F = fftn(fs)
+
+        #d = ifftshift(ifftn(F*self.H));
+        self._plan_r_F()
+        self._F *= self.H
+        self._plan_F_r()
+
+        #d = real(d);
+        return np.ravel(ifftshift(self._r))
+
+    def Ahfunc(self, f):
+        """Conjugate transform - convolve with conj. PSF"""
+#        fs = reshape(f, (self.height, self.width, self.depth))
+#
+#        F = fftn(fs)
+#        d = ifftshift(ifftn(F*self.Ht));
+#        d = real(d);
+#        return ravel(d)
+        self._r[:] = f.reshape(self._r.shape)
+
+        self._plan_r_F()
+        self._F *= self.Ht
+        self._plan_F_r()
+
+        return np.ravel(ifftshift(self._r))
+
+class dec_conv(ICTMDeconvolution, ClassicMappingFFTW):
+    def __init__(self, *args, **kwargs):
+        ICTMDeconvolution.__init__(self, *args, **kwargs)
+
+class SpatialConvolutionMapping(DeconvMappingBase):
+    """Classical deconvolution using non-fft convolution - pot. faster for
+    v. small psfs. Note that PSF must be symetric"""
+    def psf_calc(self, psf, data_size):
+        g = psf/psf.sum()
+
+        #keep track of our data shape
+        self.height = data_size[0]
+        self.width  = data_size[1]
+        self.depth  = data_size[2]
+
+        self.shape = data_size
+
+        self.g = g
+
+        #calculate OTF and conjugate transformed OTF
+        #self.H = (fftn(g));
+        #self.Ht = g.size*(ifftn(g));
+
+
+    def Lfunc(self, f):
+        """convolve with an approximate 2nd derivative likelihood operator in 3D.
+        i.e. [[[0,0,0][0,1,0][0,0,0]],[[0,1,0][1,-6,1][0,1,0]],[[0,0,0][0,1,0][0,0,0]]]
+        """
+        #make our data 3D again
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+        a = -6*fs
+
+        a[:,:,0:-1] += fs[:,:,1:]
+        a[:,:,1:] += fs[:,:,0:-1]
+
+        a[:,0:-1,:] += fs[:,1:,:]
+        a[:,1:,:] += fs[:,0:-1,:]
+
+        a[0:-1,:,:] += fs[1:,:,:]
+        a[1:,:,:] += fs[0:-1,:,:]
+
+        #flatten data again
+        return np.ravel(np.cast['f'](a))
+
+    Lhfunc=Lfunc
+
+    def Afunc(self, f):
+        """Forward transform - convolve with the PSF"""
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+
+        d = ndimage.convolve(fs, self.g)
+
+        #d = real(d);
+        return np.ravel(d)
+
+    def Ahfunc(self, f):
+        """Conjugate transform - convolve with conj. PSF"""
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+
+        d = ndimage.correlate(fs, self.g)
+        
+        return np.ravel(d)
+
+class dec_bead(ICTMDeconvolution, SpatialConvolutionMapping):
+    def __init__(self, *args, **kwargs):
+        ICTMDeconvolution.__init__(self, *args, **kwargs)
+
+class dec_4pi(ICTMDeconvolution):
+    """Variable phase 4Pi deconvolution, as descibed in Baddeley et al, Applied Optics 2006.
+
+    NOTE: this is a translation of Matlab code, without extensive testing, and without maintenance in over 10 years. I'm not sure it is 100%
+    correct or working. Use at your own risk.
+    
+    Usage: 
+
+    dec = dec_4pi()
+    dec.psf_calc(psf, kz, data_size=data.shape)
+    result = dec.deconv(data, lamb, num_iters, weights, alpha)
+    
+    where:
+
+    psf - the corresponding *confocal/widefield* PSF *without* 4Pi axial modulation. Must be the same size (shape) as the data. 
+    kz - the axial wavenumber of the 4Pi illumination in voxels. This should be the wavelength/(refractive index * axial step size)
+    data - the raw data
+    lamb - the ICTM regularisation parameter. See ICTMDeconvolution for details. Around 0.1 is a good starting point,
+           vary on a log scale - i.e. 10^-1, 10^-2 etc ...
+    num_iters - number of iterations. 20 is a reasonably good starting point.
+    weights - a weighting on the residuals to deweight missing data and/or adjust the noise behavior (see ICTMDeconvolution).
+              Using the default of 1.0 is usually fine.
+    alpha - the phase of the 4Pi illumination. This is the magic parameter, specific to variable phase deconvolution. It should
+            be an array of the same dimensions as the data, containing a pixel-wise map of phase values in radians. In the original
+            paper, these were calculated by fitting axial PSF profiles to areas in the data where there was a high chance that the 
+            underlying object was sparse (e.g. a bead or a piece of membrane parallel to the coverslip), and interpolating these to 
+            the entire 3D volume. The code to do this was written in Matlab and might be lost to time. Without fitting profiles and
+            interpolating, you can sometimes get reasonable results by eye-balling the phase and assuming a linear variation with z, 
+            although this is not recommended for anything but exploratory use.
+
+    NOTE: Vicidomini et al, Optics Letters, 2009 describe an extension of the concept which uses a joint estimation of both the 
+    phase and the underlying object, avoiding the need to supply alpha. In principle this should be preferable, but I have no first-hand
+    experience of how reliable it is. Keeping phase estimation separate from deconvolution might have it's advantages in some special cases
+    (I could imagine, e.g., that a double membrane at roughly the peak-sidelobe distance could be problematic for the joint estimation approach).
+
+    NOTE: Some of the 4PI specific logic (specifically the pre-calculation of the phase exponentials self.e1 and self.e2) lives in the base 
+    ICTMDeconvolution class. This is a historic artifact (all the deconvolution code was originally written for the 4Pi project and has since
+    been generalised). This, along with setting alpha, should be refactored into this class. 
+
+    NOTE: This should really be refactored as a on top of DeconvMappingBase to enable use with Richardson-Lucy deconvolution as well as ICTM, which
+    should be preferable for a microscope noise model.
+
+    """
+    
+    def psf_calc(self, psf, kz, data_size):
+        """Pre calculate OTFs etc ..."""
+        g = psf;
+        
+        self.height = data_size[0]
+        self.width = data_size[1]
+        self.depth = data_size[2]
+        
+        (x, y, z) = np.mgrid[-np.floor(self.height / 2.0):(np.ceil(self.height / 2.0)),
+                    -np.floor(self.width / 2.0):(np.ceil(self.width / 2.0)),
+                    -np.floor(self.depth / 2.0):(np.ceil(self.depth / 2.0))]
+        
+        gs = np.shape(g)
+        g = g[int(np.floor((gs[0] - self.height) / 2)):int(self.height + np.floor((gs[0] - self.height) / 2)),
+            int(np.floor((gs[1] - self.width) / 2)):int(self.width + np.floor((gs[1] - self.width) / 2)),
+            int(np.floor((gs[2] - self.depth) / 2)):int(self.depth + np.floor((gs[2] - self.depth) / 2))]
+        
+        g = abs(ifftshift(ifftn(abs(fftn(g)))))
+        g = (g / sum(sum(sum(g))))
+        
+        self.g = g;
+        
+        self.H = fftn(g).astype('f')
+        self.Ht = ifftn(g).astype('f')
+        
+        tk = 2 * kz * z
+        
+        t = g * np.exp(1j * tk)
+        self.He = np.cast['F'](fftn(t));
+        self.Het = np.cast['F'](ifftn(t));
+        
+        tk = 2 * tk
+        
+        t = g * np.exp(1j * tk)
+        self.He2 = np.cast['F'](fftn(t));
+        self.He2t = np.cast['F'](ifftn(t));
+    
+    def Lfunc(self, f):
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+        a = -6 * fs
+        
+        a[:, :, 0:-1] += fs[:, :, 1:]
+        a[:, :, 1:] += fs[:, :, 0:-1]
+        
+        a[:, 0:-1, :] += fs[:, 1:, :]
+        a[:, 1:, :] += fs[:, 0:-1, :]
+        
+        a[0:-1, :, :] += fs[1:, :, :]
+        a[1:, :, :] += fs[0:-1, :, :]
+        
+        return np.ravel(np.cast['f'](a))
+    
+    Lhfunc = Lfunc
+    
+    def Afunc(self, f):
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+        
+        F = fftn(fs)
+        
+        d_1 = ifftshift(ifftn(F * self.H));
+        
+        d_e = ifftshift(ifftn(F * self.He));
+        
+        d_e2 = ifftshift(ifftn(F * self.He2));
+        
+        d = (1.5 * np.real(d_1) + 2 * np.real(d_e * self.e1) + 0.5 * np.real(d_e2 * self.e2))
+        
+        d = np.real(d);
+        return np.ravel(d)
+    
+    def Ahfunc(self, f):
+        fs = np.reshape(f, (self.height, self.width, self.depth))
+        
+        F = fftn(fs)
+        
+        d_1 = ifftshift(ifftn(F * self.Ht));
+        
+        d_e = ifftshift(ifftn(F * self.Het));
+        
+        d_e2 = ifftshift(ifftn(F * self.He2t));
+        
+        d = (1.5 * d_1 + 2 * np.real(d_e * np.exp(1j * self.alpha)) + 0.5 * np.real(d_e2 * np.exp(2 * 1j * self.alpha)));
+        
+        d = np.real(d);
+        return np.ravel(d)
+
+
+class dec_4pi_c(dec_4pi):
+    def prepare(self):
+        return cDec.prepare(np.shape(self.H))
+    
+    def cleanup(self):
+        return cDec.cleanup()
+    
+    def Afunc(self, f):
+        return np.ravel(ifftshift(np.reshape(cDec.fw_map(np.cast['F'](f), np.cast['F'](self.alpha), np.cast['F'](self.H), np.cast['F'](self.He), np.cast['F'](self.He2), np.cast['F'](self.e1), np.cast['F'](self.e2)), np.shape(self.alpha))))
+    
+    def Ahfunc(self, f):
+        #return cDec.fw_map(f,self.alpha, self.Ht, self.Het, self.He2t, self.e1, self.e2)
+        return np.ravel(ifftshift(np.reshape(cDec.fw_map(np.cast['F'](f), np.cast['F'](self.alpha), np.cast['F'](self.Ht), np.cast['F'](self.Het), np.cast['F'](self.He2t), np.cast['F'](self.e1), np.cast['F'](self.e2)), np.shape(self.alpha))))
+    
+    def Lfunc(self,f):
+        return cDec.Lfunc(f, np.shape(self.alpha))
+    
+    Lhfunc = Lfunc
+  
+