def predict_2d_base(vis: Union[BlockVisibility, Visibility], model: Image,
**kwargs) -> Union[BlockVisibility, Visibility]:
""" Predict using convolutional degridding.
This is at the bottom of the layering i.e. all transforms are eventually expressed in terms of
this function. Any shifting needed is performed here.
:param vis: Visibility to be predicted
:param model: model image
:return: resulting visibility (in place works)
"""
if isinstance(vis, BlockVisibility):
log.debug("imaging.predict: coalescing prior to prediction")
avis = coalesce_visibility(vis, **kwargs)
else:
avis = vis
assert isinstance(avis, Visibility), avis
_, _, ny, nx = model.data.shape
padding = {}
if get_parameter(kwargs, "padding", False):
padding = {'padding': get_parameter(kwargs, "padding", False)}
spectral_mode, vfrequencymap = get_frequency_map(avis, model)
polarisation_mode, vpolarisationmap = get_polarisation_map(avis, model)
uvw_mode, shape, padding, vuvwmap = get_uvw_map(avis, model, **padding)
kernel_name, gcf, vkernellist = get_kernel_list(avis, model, **kwargs)
uvgrid = fft((pad_mid(model.data, int(round(padding * nx))) *
gcf).astype(dtype=complex))
avis.data['vis'] = convolutional_degrid(vkernellist,
avis.data['vis'].shape, uvgrid,
vuvwmap, vfrequencymap,
vpolarisationmap)
# Now we can shift the visibility from the image frame to the original visibility frame
svis = shift_vis_to_image(avis, model, tangent=True, inverse=True)
if isinstance(vis, BlockVisibility) and isinstance(svis, Visibility):
log.debug("imaging.predict decoalescing post prediction")
return decoalesce_visibility(svis)
else:
return svis
def predict_2d_base(vis: Visibility, model: Image, **kwargs) -> Visibility:
""" Predict using convolutional degridding.
This is at the bottom of the layering i.e. all transforms are eventually expressed in terms of
this function. Any shifting needed is performed here.
:param vis: Visibility to be predicted
:param model: model image
:return: resulting visibility (in place works)
"""
if type(vis) is not Visibility:
avis = coalesce_visibility(vis, **kwargs)
else:
avis = vis
_, _, ny, nx = model.data.shape
# print(model.shape)
spectral_mode, vfrequencymap = get_frequency_map(avis, model) # 可以并行
polarisation_mode, vpolarisationmap = get_polarisation_map(
avis, model, **kwargs) # 可以并行
uvw_mode, shape, padding, vuvwmap = get_uvw_map(avis, model,
**kwargs) # 可以并行
kernel_name, gcf, vkernellist = get_kernel_list(avis, model, **kwargs)
uvgrid = fft((pad_mid(model.data, int(round(padding * nx))) *
gcf).astype(dtype=complex))
avis.data['vis'] = convolutional_degrid(vkernellist,
avis.data['vis'].shape, uvgrid,
vuvwmap, vfrequencymap,
vpolarisationmap)
# Now we can shift the visibility from the image frame to the original visibility frame
svis = shift_vis_to_image(avis, model, tangent=True, inverse=True)
if type(vis) is not Visibility:
return decoalesce_visibility(svis)
else:
return svis
def invert_2d_base_timing(vis: Visibility, im: Image, dopsf: bool = False, normalize: bool = True, **kwargs) \
-> (Image, numpy.ndarray, tuple):
""" Invert using 2D convolution function, including w projection optionally
Use the image im as a template. Do PSF in a separate call.
This is at the bottom of the layering i.e. all transforms are eventually expressed in terms
of this function. . Any shifting needed is performed here.
:param vis: Visibility to be inverted
:param im: image template (not changed)
:param dopsf: Make the psf instead of the dirty image
:param normalize: Normalize by the sum of weights (True)
:return: resulting image
"""
opt = get_parameter(kwargs, 'opt', False)
if not opt:
log.debug('Using original algorithm')
else:
log.debug('Using optimized algorithm')
if not isinstance(vis, Visibility):
svis = coalesce_visibility(vis, **kwargs)
else:
svis = copy_visibility(vis)
if dopsf:
svis.data['vis'] = numpy.ones_like(svis.data['vis'])
svis = shift_vis_to_image(svis, im, tangent=True, inverse=False)
nchan, npol, ny, nx = im.data.shape
padding = {}
if get_parameter(kwargs, "padding", False):
padding = {'padding': get_parameter(kwargs, "padding", False)}
spectral_mode, vfrequencymap = get_frequency_map(svis, im, opt)
polarisation_mode, vpolarisationmap = get_polarisation_map(svis, im)
uvw_mode, shape, padding, vuvwmap = get_uvw_map(svis, im, **padding)
kernel_name, gcf, vkernellist = get_kernel_list(svis, im, **kwargs)
# Optionally pad to control aliasing
imgridpad = numpy.zeros(
[nchan, npol,
int(round(padding * ny)),
int(round(padding * nx))],
dtype='complex')
# Use original algorithm
if not opt:
time_grid = -time.time()
imgridpad, sumwt = convolutional_grid(vkernellist, imgridpad,
svis.data['vis'],
svis.data['imaging_weight'],
vuvwmap, vfrequencymap,
vpolarisationmap)
time_grid += time.time()
# Use optimized algorithm
else:
time_grid = -time.time()
kernel_indices, kernels = vkernellist
ks0, ks1, ks2, ks3 = kernels[0].shape
kernels_c = numpy.zeros((len(kernels), ks0, ks1, ks2, ks3),
dtype=kernels[0].dtype)
for i in range(len(kernels)):
kernels_c[i, ...] = kernels[i]
vfrequencymap_c = numpy.array(vfrequencymap, dtype=numpy.int32)
sumwt = numpy.zeros((imgridpad.shape[0], imgridpad.shape[1]),
dtype=numpy.float64)
convolutional_grid_c(imgridpad, sumwt, native_order(svis.data['vis']),
native_order(svis.data['imaging_weight']),
native_order(kernels_c),
native_order(kernel_indices),
native_order(vuvwmap),
native_order(vfrequencymap_c))
time_grid += time.time()
# Fourier transform the padded grid to image, multiply by the gridding correction
# function, and extract the unpadded inner part.
# Normalise weights for consistency with transform
sumwt /= float(padding * int(round(padding * nx)) * ny)
imaginary = get_parameter(kwargs, "imaginary", False)
if imaginary:
log.debug("invert_2d_base: retaining imaginary part of dirty image")
result = extract_mid(ifft(imgridpad) * gcf, npixel=nx)
resultreal = create_image_from_array(result.real, im.wcs)
resultimag = create_image_from_array(result.imag, im.wcs)
if normalize:
resultreal = normalize_sumwt(resultreal, sumwt)
resultimag = normalize_sumwt(resultimag, sumwt)
return resultreal, sumwt, resultimag
else:
# Use original algorithm
if not opt:
time_ifft = -time.time()
inarr = ifft(imgridpad)
time_ifft += time.time()
# Use optimized algorithm
else:
time_ifft = -time.time()
inarr = numpy.zeros(imgridpad.shape, dtype=imgridpad.dtype)
ifft_c(inarr, imgridpad)
time_ifft += time.time()
result = extract_mid(numpy.real(inarr) * gcf, npixel=nx)
resultimage = create_image_from_array(result, im.wcs)
if normalize:
resultimage = normalize_sumwt(resultimage, sumwt)
return resultimage, sumwt, (time_grid, time_ifft)
def predict_2d_base_timing(vis: Visibility, model: Image,
**kwargs) -> (Visibility, tuple):
""" Predict using convolutional degridding.
This is at the bottom of the layering i.e. all transforms are eventually expressed in terms of
this function. Any shifting needed is performed here.
:param vis: Visibility to be predicted
:param model: model image
:return: resulting visibility (in place works)
"""
if not isinstance(vis, Visibility):
avis = coalesce_visibility(vis, **kwargs)
else:
avis = vis
_, _, ny, nx = model.data.shape
opt = get_parameter(kwargs, 'opt', False)
if not opt:
log.debug('Using original algorithm')
else:
log.debug('Using optimized algorithm')
padding = {}
if get_parameter(kwargs, "padding", False):
padding = {'padding': get_parameter(kwargs, "padding", False)}
spectral_mode, vfrequencymap = get_frequency_map(avis, model, opt)
polarisation_mode, vpolarisationmap = get_polarisation_map(avis, model)
uvw_mode, shape, padding, vuvwmap = get_uvw_map(avis, model, **padding)
kernel_name, gcf, vkernellist = get_kernel_list(avis, model, **kwargs)
inarr = (pad_mid(model.data, int(round(padding * nx))) *
gcf).astype(dtype=complex)
# Use original algorithm
if not opt:
time_fft = -time.time()
uvgrid = fft(inarr)
time_fft += time.time()
time_degrid = -time.time()
vt = convolutional_degrid(vkernellist, avis.data['vis'].shape, uvgrid,
vuvwmap, vfrequencymap, vpolarisationmap)
time_degrid += time.time()
# Use optimized algorithm
else:
time_fft = -time.time()
uvgrid = numpy.zeros(inarr.shape, dtype=inarr.dtype)
fft_c(uvgrid, inarr)
time_fft += time.time()
time_degrid = -time.time()
kernel_indices, kernels = vkernellist
ks0, ks1, ks2, ks3 = kernels[0].shape
kernels_c = numpy.zeros((len(kernels), ks0, ks1, ks2, ks3),
dtype=kernels[0].dtype)
for i in range(len(kernels)):
kernels_c[i, ...] = kernels[i]
vfrequencymap_c = numpy.array(vfrequencymap, dtype=numpy.int32)
vt = numpy.zeros(avis.data['vis'].shape, dtype=numpy.complex128)
convolutional_degrid_c(vt, native_order(kernels_c),
native_order(kernel_indices),
native_order(uvgrid), native_order(vuvwmap),
native_order(vfrequencymap_c))
time_degrid += time.time()
avis.data['vis'] = vt
# Now we can shift the visibility from the image frame to the original visibility frame
svis = shift_vis_to_image(avis, model, tangent=True, inverse=True)
if not isinstance(vis, Visibility):
svis = decoalesce_visibility(svis)
return svis, (time_degrid, time_fft)
def invert_2d_base(vis: Visibility, im: Image, dopsf: bool = False, normalize: bool = True, **kwargs) \
-> (Image, numpy.ndarray):
""" Invert using 2D convolution function, including w projection optionally
Use the image im as a template. Do PSF in a separate call.
This is at the bottom of the layering i.e. all transforms are eventually expressed in terms
of this function. . Any shifting needed is performed here.
:param vis: Visibility to be inverted
:param im: image template (not changed)
:param dopsf: Make the psf instead of the dirty image
:param normalize: Normalize by the sum of weights (True)
:return: resulting image
"""
if type(vis) is not Visibility:
svis = coalesce_visibility(vis, **kwargs)
else:
svis = copy_visibility(vis)
if dopsf:
svis.data['vis'] = numpy.ones_like(svis.data['vis'])
svis = shift_vis_to_image(svis, im, tangent=True, inverse=False)
nchan, npol, ny, nx = im.data.shape
spectral_mode, vfrequencymap = get_frequency_map(svis, im)
polarisation_mode, vpolarisationmap = get_polarisation_map(
svis, im, **kwargs)
uvw_mode, shape, padding, vuvwmap = get_uvw_map(svis, im, **kwargs)
kernel_name, gcf, vkernellist = get_kernel_list(svis, im, **kwargs)
# Optionally pad to control aliasing
imgridpad = numpy.zeros(
[nchan, npol,
int(round(padding * ny)),
int(round(padding * nx))],
dtype='complex')
imgridpad, sumwt = convolutional_grid(vkernellist, imgridpad,
svis.data['vis'],
svis.data['imaging_weight'], vuvwmap,
vfrequencymap, vpolarisationmap)
# Fourier transform the padded grid to image, multiply by the gridding correction
# function, and extract the unpadded inner part.
# Normalise weights for consistency with transform
sumwt /= float(padding * int(round(padding * nx)) * ny)
imaginary = get_parameter(kwargs, "imaginary", False)
if imaginary:
log.debug("invert_2d_base: retaining imaginary part of dirty image")
result = extract_mid(ifft(imgridpad) * gcf, npixel=nx)
resultreal = create_image_from_array(result.real, im.wcs)
resultimag = create_image_from_array(result.imag, im.wcs)
if normalize:
resultreal = normalize_sumwt(resultreal, sumwt)
resultimag = normalize_sumwt(resultimag, sumwt)
return resultreal, sumwt, resultimag
else:
result = extract_mid(numpy.real(ifft(imgridpad)) * gcf, npixel=nx)
resultimage = create_image_from_array(result, im.wcs)
if normalize:
resultimage = normalize_sumwt(resultimage, sumwt)
return resultimage, sumwt
