Merge branch 'lm_space_complex_transformsV2' into 'feature/field_multiple_space'

Lm space complex transforms v2



See merge request !27

nifty/field.py

@@ -5,7 +5,7 @@ from d2o import distributed_data_object,\
     STRATEGIES as DISTRIBUTION_STRATEGIES
 
 from nifty.config import about,\
-                         nifty_configuration as gc,\
+                         nifty_configuration as gc
 
 from nifty.field_types import FieldType
 

nifty/operators/fft_operator/fft_operator.py

 from nifty.config import about
 import nifty.nifty_utilities as utilities
+from nifty.spaces import RGSpace,\
+                         GLSpace,\
+                         HPSpace,\
+                         LMSpace
+
 from nifty.operators.linear_operator import LinearOperator
-from transformations import TransformationFactory
+from transformations import RGRGTransformation,\
+                            LMGLTransformation,\
+                            LMHPTransformation,\
+                            GLLMTransformation,\
+                            HPLMTransformation,\
+                            TransformationCache
 
 
 class FFTOperator(LinearOperator):
 
+    # ---Class attributes---
+
+    default_codomain_dictionary = {RGSpace: RGSpace,
+                                   HPSpace: LMSpace,
+                                   GLSpace: LMSpace,
+                                   LMSpace: HPSpace,
+                                   }
+
+    transformation_dictionary = {(RGSpace, RGSpace): RGRGTransformation,
+                                 (HPSpace, LMSpace): HPLMTransformation,
+                                 (GLSpace, LMSpace): GLLMTransformation,
+                                 (LMSpace, HPSpace): LMHPTransformation,
+                                 (LMSpace, GLSpace): LMGLTransformation
+                                 }
+
     # ---Overwritten properties and methods---
 
-    def __init__(self, domain=(), field_type=(), target=None):
+    def __init__(self, domain=(), field_type=(), target=None, module=None):
+
         super(FFTOperator, self).__init__(domain=domain,
                                           field_type=field_type)
 
+        # Initialize domain and target
         if len(self.domain) != 1:
             raise ValueError(about._errors.cstring(
                     'ERROR: TransformationOperator accepts only exactly one '
@@ -24,17 +51,30 @@ class FFTOperator(LinearOperator):
             ))
 
         if target is None:
-            target = utilities.get_default_codomain(self.domain[0])
-
+            target = (self.get_default_codomain(self.domain[0]), )
         self._target = self._parse_domain(target)
 
-        self._forward_transformation = TransformationFactory.create(
-            self.domain[0], self.target[0]
-        )
-
-        self._inverse_transformation = TransformationFactory.create(
-            self.target[0], self.domain[0]
-        )
+        # Create transformation instances
+        try:
+            forward_class = self.transformation_dictionary[
+                (self.domain[0].__class__, self.target[0].__class__)]
+        except KeyError:
+            raise TypeError(about._errors.cstring(
+                "ERROR: No forward transformation for domain-target pair "
+                "found."))
+        try:
+            backward_class = self.transformation_dictionary[
+                (self.target[0].__class__, self.domain[0].__class__)]
+        except KeyError:
+            raise TypeError(about._errors.cstring(
+                "ERROR: No backward transformation for domain-target pair "
+                "found."))
+
+        self._forward_transformation = TransformationCache.create(
+            forward_class, self.domain[0], self.target[0], module=module)
+
+        self._backward_transformation = TransformationCache.create(
+            backward_class, self.target[0], self.domain[0], module=module)
 
     def _times(self, x, spaces, types):
         spaces = utilities.cast_axis_to_tuple(spaces, len(x.domain))
@@ -69,7 +109,7 @@ class FFTOperator(LinearOperator):
         else:
             axes = x.domain_axes[spaces[0]]
 
-        new_val = self._inverse_transformation.transform(x.val, axes=axes)
+        new_val = self._backward_transformation.transform(x.val, axes=axes)
 
         if spaces is None:
             result_domain = self.domain
@@ -99,3 +139,22 @@ class FFTOperator(LinearOperator):
     @property
     def unitary(self):
         return True
+
+    # ---Added properties and methods---
+
+    @classmethod
+    def get_default_codomain(cls, domain):
+        domain_class = domain.__class__
+        try:
+            codomain_class = cls.default_codomain_dictionary[domain_class]
+        except KeyError:
+            raise TypeError(about._errors.cstring("ERROR: unknown domain"))
+
+        try:
+            transform_class = cls.transformation_dictionary[(domain_class,
+                                                             codomain_class)]
+        except KeyError:
+            raise TypeError(about._errors.cstring(
+                "ERROR: No transformation for domain-codomain pair found."))
+
+        return transform_class.get_codomain(domain)

nifty/operators/fft_operator/transformations/__init__.py

@@ -4,4 +4,4 @@ from hplmtransformation import HPLMTransformation
 from lmgltransformation import LMGLTransformation
 from lmhptransformation import LMHPTransformation
 
-from transformation_factory import TransformationFactory
\ No newline at end of file
+from transformation_cache import TransformationCache
\ No newline at end of file

nifty/operators/fft_operator/transformations/gllmtransformation.py

 import numpy as np
-from transformation import Transformation
-from d2o import distributed_data_object
-from nifty.config import dependency_injector as gdi
-import nifty.nifty_utilities as utilities
+from nifty.config import dependency_injector as gdi,\
+                         about
 from nifty import GLSpace, LMSpace
+from slicing_transformation import SlicingTransformation
+import lm_transformation_factory as ltf
 
-gl = gdi.get('libsharp_wrapper_gl')
+libsharp = gdi.get('libsharp_wrapper_gl')
 
 
-class GLLMTransformation(Transformation):
+class GLLMTransformation(SlicingTransformation):
+
+    # ---Overwritten properties and methods---
+
     def __init__(self, domain, codomain=None, module=None):
         if 'libsharp_wrapper_gl' not in gdi:
-            raise ImportError("The module libsharp is needed but not available")
-
-        if codomain is None:
-            self.domain = domain
-            self.codomain = self.get_codomain(domain)
-        elif self.check_codomain(domain, codomain):
-            self.domain = domain
-            self.codomain = codomain
-        else:
-            raise ValueError("ERROR: Incompatible codomain!")
+            raise ImportError(about._errors.cstring(
+                "The module libsharp is needed but not available."))
 
-    @staticmethod
-    def get_codomain(domain):
+        super(GLLMTransformation, self).__init__(domain, codomain, module)
+
+    # ---Mandatory properties and methods---
+
+    @classmethod
+    def get_codomain(cls, domain):
         """
             Generates a compatible codomain to which transformations are
             reasonable, i.e.\  an instance of the :py:class:`lm_space` class.
@@ -38,96 +37,89 @@ class GLLMTransformation(Transformation):
             codomain : LMSpace
                 A compatible codomain.
         """
-        if domain is None:
-            raise ValueError('ERROR: cannot generate codomain for None')
 
         if not isinstance(domain, GLSpace):
-            raise TypeError('ERROR: domain needs to be a GLSpace')
+            raise TypeError(about._errors.cstring(
+                "ERROR: domain needs to be a GLSpace"))
 
         nlat = domain.nlat
         lmax = nlat - 1
         mmax = nlat - 1
         if domain.dtype == np.dtype('float32'):
-            return LMSpace(lmax=lmax, mmax=mmax, dtype=np.complex64)
+            return_dtype = np.float32
         else:
-            return LMSpace(lmax=lmax, mmax=mmax, dtype=np.complex128)
+            return_dtype = np.float64
+
+        result = LMSpace(lmax=lmax, mmax=mmax, dtype=return_dtype)
+        cls.check_codomain(domain, result)
+        return result
 
     @staticmethod
     def check_codomain(domain, codomain):
         if not isinstance(domain, GLSpace):
-            raise TypeError('ERROR: domain is not a GLSpace')
-
-        if codomain is None:
-            return False
+            raise TypeError(about._errors.cstring(
+                "ERROR: domain is not a GLSpace"))
 
         if not isinstance(codomain, LMSpace):
-            raise TypeError('ERROR: codomain must be a LMSpace.')
+            raise TypeError(about._errors.cstring(
+                "ERROR: codomain must be a LMSpace."))
 
         nlat = domain.nlat
         nlon = domain.nlon
         lmax = codomain.lmax
         mmax = codomain.mmax
 
-        if (nlon != 2 * nlat - 1) or (lmax != nlat - 1) or (lmax != mmax):
-            return False
+        if lmax != mmax:
+            raise ValueError(about._errors.cstring(
+                'ERROR: codomain has lmax != mmax.'))
 
-        return True
-
-    def transform(self, val, axes=None, **kwargs):
-        """
-        GL -> LM transform method.
+        if lmax != nlat - 1:
+            raise ValueError(about._errors.cstring(
+                'ERROR: codomain has lmax != nlat - 1.'))
 
-        Parameters
-        ----------
-        val : np.ndarray or distributed_data_object
-            The value array which is to be transformed
+        if nlon != 2 * nlat - 1:
+            raise ValueError(about._errors.cstring(
+                'ERROR: domain has nlon != 2 * nlat - 1.'))
 
-        axes : None or tuple
-            The axes along which the transformation should take place
+        return None
 
-        """
-        if self.domain.discrete:
-            val = self.domain.weight(val, power=-0.5, axes=axes)
-
-        # shorthands for transform parameters
+    def _transformation_of_slice(self, inp, **kwargs):
         nlat = self.domain.nlat
         nlon = self.domain.nlon
         lmax = self.codomain.lmax
         mmax = self.codomain.mmax
 
-        if isinstance(val, distributed_data_object):
-            temp_val = val.get_full_data()
-        else:
-            temp_val = val
-
-        return_val = None
-
-        for slice_list in utilities.get_slice_list(temp_val.shape, axes):
-            if slice_list == [slice(None, None)]:
-                inp = temp_val
-            else:
-                if return_val is None:
-                    return_val = np.empty_like(temp_val)
-                inp = temp_val[slice_list]
-
-            if self.domain.dtype == np.dtype('float32'):
-                inp = gl.map2alm_f(inp,
-                                   nlat=nlat, nlon=nlon,
-                                   lmax=lmax, mmax=mmax)
-            else:
-                inp = gl.map2alm(inp,
-                                 nlat=nlat, nlon=nlon,
-                                 lmax=lmax, mmax=mmax)
-
-            if slice_list == [slice(None, None)]:
-                return_val = inp
-            else:
-                return_val[slice_list] = inp
-
-        if isinstance(val, distributed_data_object):
-            new_val = val.copy_empty(dtype=self.codomain.dtype)
-            new_val.set_full_data(return_val, copy=False)
+        if issubclass(inp.dtype.type, np.complexfloating):
+
+            [resultReal, resultImag] = [self.libsharpMap2Alm(x,
+                                                             nlat=nlat,
+                                                             nlon=nlon,
+                                                             lmax=lmax,
+                                                             mmax=mmax,
+                                                             **kwargs)
+                                        for x in (inp.real, inp.imag)]
+
+            [resultReal, resultImag] = [ltf.buildIdx(x, lmax=lmax)
+                                        for x in [resultReal, resultImag]]
+
+            result = self._combine_complex_result(resultReal, resultImag)
         else:
-            return_val = return_val.astype(self.codomain.dtype, copy=False)
+            result = self.libsharpMap2Alm(inp, nlat=nlat, nlon=nlon, lmax=lmax,
+                                          mmax=mmax)
+            result = ltf.buildIdx(result, lmax=lmax)
+
+        return result
 
-        return return_val
+    # ---Added properties and methods---
+
+    def libsharpMap2Alm(self, inp, **kwargs):
+        if inp.dtype == np.dtype('float32'):
+            return libsharp.map2alm_f(inp, **kwargs)
+        elif inp.dtype == np.dtype('float64'):
+            return libsharp.map2alm(inp, **kwargs)
+        else:
+            about.warnings.cprint("WARNING: performing dtype conversion for "
+                                  "libsharp compatibility.")
+            casted_inp = inp.astype(np.dtype('float64'), copy=False)
+            result = libsharp.map2alm(casted_inp, **kwargs)
+            return result
\ No newline at end of file

nifty/operators/fft_operator/transformations/hplmtransformation.py

 import numpy as np
-from transformation import Transformation
-from d2o import distributed_data_object
-from nifty.config import dependency_injector as gdi
-import nifty.nifty_utilities as utilities
+from nifty.config import dependency_injector as gdi,\
+                         about
 from nifty import HPSpace, LMSpace
+from slicing_transformation import SlicingTransformation
+
+import lm_transformation_factory as ltf
 
 hp = gdi.get('healpy')
 
 
-class HPLMTransformation(Transformation):
+class HPLMTransformation(SlicingTransformation):
+
+    # ---Overwritten properties and methods---
+
     def __init__(self, domain, codomain=None, module=None):
         if 'healpy' not in gdi:
-            raise ImportError("The module healpy is needed but not available")
-
-        if codomain is None:
-            self.domain = domain
-            self.codomain = self.get_codomain(domain)
-        elif self.check_codomain(domain, codomain):
-            self.domain = domain
-            self.codomain = codomain
-        else:
-            raise ValueError("ERROR: Incompatible codomain!")
+            raise ImportError(about._errors.cstring(
+                "The module healpy is needed but not available"))
 
-    @staticmethod
-    def get_codomain(domain):
+        super(HPLMTransformation, self).__init__(domain, codomain, module)
+
+    # ---Mandatory properties and methods---
+
+    @classmethod
+    def get_codomain(cls, domain):
         """
             Generates a compatible codomain to which transformations are
             reasonable, i.e.\  an instance of the :py:class:`lm_space` class.
@@ -38,87 +38,65 @@ class HPLMTransformation(Transformation):
             codomain : LMSpace
                 A compatible codomain.
         """
-        if domain is None:
-            raise ValueError('ERROR: cannot generate codomain for None')
 
         if not isinstance(domain, HPSpace):
-            raise TypeError('ERROR: domain needs to be a HPSpace')
+            raise TypeError(about._errors.cstring(
+                "ERROR: domain needs to be a HPSpace"))
 
         lmax = 3 * domain.nside - 1
         mmax = lmax
-        return LMSpace(lmax=lmax, mmax=mmax, dtype=np.dtype('complex128'))
+
+        result = LMSpace(lmax=lmax, mmax=mmax, dtype=np.dtype('float64'))
+        cls.check_codomain(domain, result)
+        return result
 
     @staticmethod
     def check_codomain(domain, codomain):
         if not isinstance(domain, HPSpace):
-            raise TypeError('ERROR: domain is not a HPSpace')
-
-        if codomain is None:
-            return False
+            raise TypeError(about._errors.cstring(
+                'ERROR: domain is not a HPSpace'))
 
         if not isinstance(codomain, LMSpace):
-            raise TypeError('ERROR: codomain must be a LMSpace.')
+            raise TypeError(about._errors.cstring(
+                'ERROR: codomain must be a LMSpace.'))
 
         nside = domain.nside
         lmax = codomain.lmax
         mmax = codomain.mmax
 
-        if (3 * nside - 1 != lmax) or (lmax != mmax):
-            return False
+        if 3 * nside - 1 != lmax:
+            raise ValueError(about._errors.cstring(
+                'ERROR: codomain has 3*nside-1 != lmax.'))
 
-        return True
+        if lmax != mmax:
+            raise ValueError(about._errors.cstring(
+                'ERROR: codomain has lmax != mmax.'))
 
-    def transform(self, val, axes=None, **kwargs):
-        """
-        HP -> LM transform method.
-
-        Parameters
-        ----------
-        val : np.ndarray or distributed_data_object
-            The value array which is to be transformed
+        return None
 
-        axes : None or tuple
-            The axes along which the transformation should take place
+    def _transformation_of_slice(self, inp, **kwargs):
+        lmax = self.codomain.lmax
+        mmax = self.codomain.mmax
 
-        """
-        # get by number of iterations from kwargs
-        niter = kwargs['niter'] if 'niter' in kwargs else 0
+        if issubclass(inp.dtype.type, np.complexfloating):
+            [resultReal, resultImag] = [hp.map2alm(x.astype(np.float64,
+                                                            copy=False),
+                                                   lmax=lmax,
+                                                   mmax=mmax,
+                                                   pol=True,
+                                                   use_weights=False,
+                                                   **kwargs)
+                                        for x in (inp.real, inp.imag)]
 
-        if self.domain.discrete:
-            val = self.domain.weight(val, power=-0.5, axes=axes)
+            [resultReal, resultImag] = [ltf.buildIdx(x, lmax=lmax)
+                                        for x in [resultReal, resultImag]]
 
-        # shorthands for transform parameters
-        lmax = self.codomain.lmax
-        mmax = self.codomain.mmax
+            result = self._combine_complex_result(resultReal, resultImag)
 
-        if isinstance(val, distributed_data_object):
-            temp_val = val.get_full_data()
-        else:
-            temp_val = val
-
-        return_val = None
-
-        for slice_list in utilities.get_slice_list(temp_val.shape, axes):
-            if slice_list == [slice(None, None)]:
-                inp = temp_val
-            else:
-                if return_val is None:
-                    return_val = np.empty_like(temp_val)
-                inp = temp_val[slice_list]
-
-            inp = hp.map2alm(inp.astype(np.float64, copy=False),
-                             lmax=lmax, mmax=mmax, iter=niter, pol=True,
-                             use_weights=False, datapath=None)
-
-            if slice_list == [slice(None, None)]:
-                return_val = inp
-            else:
-                return_val[slice_list] = inp
-
-        if isinstance(val, distributed_data_object):
-            new_val = val.copy_empty(dtype=self.codomain.dtype)
-            new_val.set_full_data(return_val, copy=False)
         else:
-            return_val = return_val.astype(self.codomain.dtype, copy=False)
+            result = hp.map2alm(inp.astype(np.float64, copy=False),
+                                lmax=lmax, mmax=mmax, pol=True,
+                                use_weights=False)
+            result = ltf.buildIdx(result, lmax=lmax)
 
-        return return_val
+        return result

nifty/operators/fft_operator/transformations/lm_transformation_factory.pyx

+import numpy as np
+cimport numpy as np
+
+def buildLm(inp, **kwargs):
+    if inp.dtype == np.dtype('float32'):
+        return _buildLm_f(inp, **kwargs)
+    else:
+        return _buildLm(inp, **kwargs)
+
+def buildIdx(inp, **kwargs):
+    if inp.dtype == np.dtype('complex64'):
+        return _buildIdx_f(inp, **kwargs)
+    else:
+        return _buildIdx(inp, **kwargs)
+
+cpdef np.ndarray[np.float32_t, ndim=1]  _buildIdx_f(np.ndarray[np
+.complex64_t, ndim=1] nr, np.int_t lmax):
+    cdef np.int size = (lmax+1)*(lmax+1)
+
+    cdef np.ndarray res=np.zeros([size], dtype=np.complex64)
+    cdef np.ndarray final=np.zeros([size], dtype=np.float32)
+    res[0:lmax+1] = nr[0:lmax+1]
+
+    for i in xrange(len(nr)-lmax-1):
+        res[i*2+lmax+1] = nr[i+lmax+1]
+        res[i*2+lmax+2] = np.conjugate(nr[i+lmax+1])
+    final = _realify_f(res, lmax)
+    return final
+
+cpdef np.ndarray[np.float32_t, ndim=1] _realify_f(np.ndarray[np.complex64_t, ndim=1] nr, np.int_t lmax):
+    cdef np.ndarray resi=np.zeros([len(nr)], dtype=np.float32)
+
+    resi[0:lmax+1] = np.real(nr[0:lmax+1])
+
+    for i in xrange(lmax+1,len(nr),2):
+        mi =  int(np.ceil(((2*lmax+1)-np.sqrt((2*lmax+1)*(2*lmax+1)-4*(i-lmax)+1))/2))
+        resi[i]=np.sqrt(2)*np.real(nr[i])*(-1)**(mi*mi)
+        resi[i+1]=np.sqrt(2)*np.imag(nr[i])*(-1)**(mi*mi)
+    return resi
+
+cpdef np.ndarray[np.float64_t, ndim=1]  _buildIdx(np.ndarray[np.complex128_t,
+ ndim=1] nr, np.int_t lmax):
+    cdef np.int size = (lmax+1)*(lmax+1)
+
+    cdef np.ndarray res=np.zeros([size], dtype=np.complex128)
+    cdef np.ndarray final=np.zeros([size], dtype=np.float64)
+    res[0:lmax+1] = nr[0:lmax+1]
+
+    for i in xrange(len(nr)-lmax-1):
+        res[i*2+lmax+1] = nr[i+lmax+1]
+        res[i*2+lmax+2] = np.conjugate(nr[i+lmax+1])
+    final = _realify(res, lmax)
+    return final
+
+cpdef np.ndarray[np.float64_t, ndim=1] _realify(np.ndarray[np.complex128_t, ndim=1] nr, np.int_t lmax):
+    cdef np.ndarray resi=np.zeros([len(nr)], dtype=np.float64)
+
+    resi[0:lmax+1] = np.real(nr[0:lmax+1])
+
+    for i in xrange(lmax+1,len(nr),2):
+        mi =  int(np.ceil(((2*lmax+1)-np.sqrt((2*lmax+1)*(2*lmax+1)-4*(i-lmax)+1))/2))
+        resi[i]=np.sqrt(2)*np.real(nr[i])*(-1)**(mi*mi)
+        resi[i+1]=np.sqrt(2)*np.imag(nr[i])*(-1)**(mi*mi)
+    return resi
+
+cpdef np.ndarray[np.complex64_t, ndim=1] _buildLm_f(np.ndarray[np.float32_t,