def restore_cube(model: Image, psf: Image, residual=None, **kwargs) -> Image:
""" Restore the model image to the residuals
:params psf: Input PSF
:return: restored image
"""
assert isinstance(model, Image), model
assert image_is_canonical(model)
assert isinstance(psf, Image), psf
assert image_is_canonical(psf)
assert residual is None or isinstance(residual, Image), residual
assert image_is_canonical(residual)
restored = copy_image(model)
npixel = psf.data.shape[3]
sl = slice(npixel // 2 - 7, npixel // 2 + 8)
size = get_parameter(kwargs, "psfwidth", None)
if size is None:
# isotropic at the moment!
from scipy.optimize import minpack
try:
fit = fit_2dgaussian(psf.data[0, 0, sl, sl])
if fit.x_stddev <= 0.0 or fit.y_stddev <= 0.0:
log.debug(
'restore_cube: error in fitting to psf, using 1 pixel stddev'
)
size = 1.0
else:
size = max(fit.x_stddev, fit.y_stddev)
log.debug('restore_cube: psfwidth = %s' % (size))
except minpack.error as err:
log.debug('restore_cube: minpack error, using 1 pixel stddev')
size = 1.0
except ValueError as err:
log.debug(
'restore_cube: warning in fit to psf, using 1 pixel stddev')
size = 1.0
else:
log.debug('restore_cube: Using specified psfwidth = %s' % (size))
# By convention, we normalise the peak not the integral so this is the volume of the Gaussian
norm = 2.0 * numpy.pi * size**2
gk = Gaussian2DKernel(size)
for chan in range(model.shape[0]):
for pol in range(model.shape[1]):
restored.data[chan, pol, :, :] = norm * convolve_fft(
model.data[chan, pol, :, :],
gk,
normalize_kernel=False,
allow_huge=True)
if residual is not None:
restored.data += residual.data
return restored
def get_frequency_map(vis, im: Image = None):
""" Map channels from visibilities to image
"""
# Find the unique frequencies in the visibility
ufrequency = numpy.unique(vis.frequency)
vnchan = len(ufrequency)
if im is None:
spectral_mode = 'channel'
if vis.frequency_map is None:
vfrequencymap = get_rowmap(vis.frequency, ufrequency)
vis.frequencymap = vfrequencymap
else:
vfrequencymap = vis.frequency_map
assert min(
vfrequencymap
) >= 0, "Invalid frequency map: visibility channel < 0: %s" % str(
vfrequencymap)
elif im.data.shape[0] == 1 and vnchan >= 1:
assert image_is_canonical(im)
spectral_mode = 'mfs'
if vis.frequency_map is None:
vfrequencymap = numpy.zeros_like(vis.frequency, dtype='int')
vis.frequencymap = vfrequencymap
else:
vfrequencymap = vis.frequency_map
else:
assert image_is_canonical(im)
# We can map these to image channels
v2im_map = im.wcs.sub(['spectral']).wcs_world2pix(ufrequency,
0)[0].astype('int')
spectral_mode = 'channel'
nrows = len(vis.frequency)
row2vis = numpy.array(get_rowmap(vis.frequency, ufrequency))
vfrequencymap = [v2im_map[row2vis[row]] for row in range(nrows)]
assert min(vfrequencymap
) >= 0, "Invalid frequency map: image channel < 0 %s" % str(
vfrequencymap)
assert max(vfrequencymap) < im.shape[0], "Invalid frequency map: image channel > number image channels %s" % \
str(vfrequencymap)
return spectral_mode, vfrequencymap
def image_gather_channels(image_list: List[Image],
im: Image = None,
subimages=0) -> Image:
"""Gather a list of subimages back into an image using the channel_iterator
If the template image is not given then it will be formed assuming that the list has
been generated by image_scatter_channels with subimages = number of channels
:param image_list: List of subimages
:param im: Output image
:param subimages: Number of image partitions on each axis (2)
:return: list of subimages
"""
if im is None:
nchan = len(image_list)
_, npol, ny, nx = image_list[0].shape
im_shape = nchan, npol, ny, ny
im = create_image_from_array(
numpy.zeros(im_shape, dtype=image_list[0].data.dtype),
image_list[0].wcs, image_list[0].polarisation_frame)
assert image_is_canonical(im)
if subimages == 0:
subimages = len(image_list)
for i, slab in enumerate(image_channel_iter(im, subimages=subimages)):
slab.data[...] = image_list[i].data[...]
return im
def weight_blockvisibility(vis,
model,
gcfcf=None,
weighting="uniform",
robustness=0.0,
**kwargs):
""" Weight the visibility data
This is done collectively so the weights are summed over all vis_lists and then
corrected
:param vis_list:
:param model_imagelist: Model required to determine weighting parameters
:param weighting: Type of weighting
:param kwargs: Parameters for functions in graphs
:return: List of vis_graphs
"""
assert isinstance(vis, BlockVisibility), vis
assert image_is_canonical(model)
if gcfcf is None:
gcfcf = create_pswf_convolutionfunction(model)
griddata = create_griddata_from_image(model, vis)
griddata, sumwt = grid_blockvisibility_weight_to_griddata(
vis, griddata, gcfcf[1])
vis = griddata_blockvisibility_reweight(vis,
griddata,
gcfcf[1],
weighting=weighting,
robustness=robustness)
return vis
def predict_wstack_single(vis,
model,
remove=True,
gcfcf=None,
**kwargs) -> Visibility:
""" Predict using a single w slices.
This processes a single w plane, rotating out the w beam for the average w
The w-stacking or w-slicing approach is to partition the visibility data by slices in w. The measurement equation is
approximated as:
.. math::
V(u,v,w) =\\sum_i \\int \\frac{ I(l,m) e^{-2 \\pi j (w_i(\\sqrt{1-l^2-m^2}-1))})}{\\sqrt{1-l^2-m^2}} e^{-2 \\pi j (ul+vm)} dl dm
If images constructed from slices in w are added after applying a w-dependent image plane correction, the w term will be corrected.
:param vis: Visibility to be predicted
:param model: model image
:return: resulting visibility (in place works)
"""
assert isinstance(vis, Visibility), vis
assert image_is_canonical(model)
vis.data['vis'][...] = 0.0
log.debug("predict_wstack_single: predicting using single w slice")
# We might want to do wprojection so we remove the average w
w_average = numpy.average(vis.w)
if remove:
vis.data['uvw'][..., 2] -= w_average
tempvis = copy_visibility(vis)
# Calculate w beam and apply to the model. The imaginary part is not needed
workimage = copy_image(model)
w_beam = create_w_term_like(model, w_average, vis.phasecentre)
# Do the real part
workimage.data = w_beam.data.real * model.data
vis = predict_2d(vis, workimage, gcfcf=gcfcf, **kwargs)
# and now the imaginary part
workimage.data = w_beam.data.imag * model.data
tempvis = predict_2d(tempvis, workimage, gcfcf=gcfcf, **kwargs)
vis.data['vis'] -= 1j * tempvis.data['vis']
if remove:
vis.data['uvw'][..., 2] += w_average
return vis
def invert_wstack_single(vis: Visibility,
im: Image,
dopsf,
normalize=True,
remove=True,
gcfcf=None,
**kwargs) -> (Image, numpy.ndarray):
"""Process single w slice
The w-stacking or w-slicing approach is to partition the visibility data by slices in w. The measurement equation is
approximated as:
.. math::
V(u,v,w) =\\sum_i \\int \\frac{ I(l,m) e^{-2 \\pi j (w_i(\\sqrt{1-l^2-m^2}-1))})}{\\sqrt{1-l^2-m^2}} e^{-2 \\pi j (ul+vm)} dl dm
If images constructed from slices in w are added after applying a w-dependent image plane correction, the w term will be corrected.
: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)
:returns: image, sum of weights
"""
assert image_is_canonical(im)
log.debug("invert_wstack_single: predicting using single w slice")
kwargs['imaginary'] = True
assert isinstance(vis, Visibility), vis
# We might want to do wprojection so we remove the average w
w_average = numpy.average(vis.w)
if remove:
vis.data['uvw'][..., 2] -= w_average
reWorkimage, sumwt, imWorkimage = invert_2d(vis,
im,
dopsf,
normalize=normalize,
gcfcf=gcfcf,
**kwargs)
if remove:
vis.data['uvw'][..., 2] += w_average
# Calculate w beam and apply to the model. The imaginary part is not needed
w_beam = create_w_term_like(im, w_average, vis.phasecentre)
reWorkimage.data = w_beam.data.real * reWorkimage.data - w_beam.data.imag * imWorkimage.data
return reWorkimage, sumwt
def get_polarisation_map(vis: Visibility, im: Image = None):
""" Get the mapping of visibility polarisations to image polarisations
"""
assert image_is_canonical(im)
if vis.polarisation_frame == im.polarisation_frame:
if vis.polarisation_frame == PolarisationFrame('stokesI'):
return "stokesI->stokesI", lambda pol: 0
elif vis.polarisation_frame == PolarisationFrame('stokesIQUV'):
return "stokesIQUV->stokesIQUV", lambda pol: pol
return "unknown", lambda pol: pol
def image_channel_iter(im: Image, subimages=1) -> collections.Iterable:
"""Create a image_channel_iter generator, returning images
The WCS is adjusted appropriately for each raster element. Hence this is a coordinate-aware
way to iterate through an image.
Provided we don't break reference semantics, memory should be conserved
To update the image in place::
for r in image_channel_iter(im, subimages=nchan):
r.data[...] = numpy.sqrt(r.data[...])
:param im: Image
:param subimages: Number of subimages
:returns: Generator of images
See also
:py:func:`rascil.processing_components.image.image_gather_channels`
:py:func:`rascil.processing_components.image.image_scatter_channels`
"""
assert image_is_canonical(im)
nchan, npol, ny, nx = im.shape
assert subimages <= nchan, "More subimages %d than channels %d" % (
subimages, nchan)
step = nchan // subimages
channels = numpy.array(range(0, nchan, step), dtype='int')
assert len(
channels
) == subimages, "subimages %d does not match length of channels %d" % (
subimages, len(channels))
for i, channel in enumerate(channels):
if i + 1 < len(channels):
channel_max = channels[i + 1]
else:
channel_max = nchan
# Adjust WCS
wcs = im.wcs.deepcopy()
wcs.wcs.crpix[3] -= channel
# Yield image from slice (reference!)
yield create_image_from_array(im.data[channel:channel_max, ...], wcs,
im.polarisation_frame)
def image_scatter_channels(im: Image, subimages=None) -> List[Image]:
"""Scatter an image into a list of subimages using the channels
:param im: Image
:param subimages: Number of channels
:return: list of subimages
See also
:py:func:`rascil.processing_components.image.iterators.image_channel_iter`
"""
assert image_is_canonical(im)
image_list = list()
if subimages is None:
subimages = im.shape[0]
for slab in image_channel_iter(im, subimages=subimages):
image_list.append(slab)
assert len(image_list) == subimages, "Too many subimages scattered"
return image_list
def predict_ng(bvis: BlockVisibility, model: Image,
**kwargs) -> BlockVisibility:
""" Predict using convolutional degridding.
Nifty-gridder version. https://gitlab.mpcdf.mpg.de/ift/nifty_gridder
In the imaging and pipeline workflows, this may be invoked using context='ng'.
:param bvis: BlockVisibility to be predicted
:param model: model image
:return: resulting BlockVisibility (in place works)
"""
assert isinstance(bvis, BlockVisibility), bvis
assert image_is_canonical(model)
if model is None:
return bvis
nthreads = get_parameter(kwargs, "threads", 4)
epsilon = get_parameter(kwargs, "epsilon", 1e-12)
do_wstacking = get_parameter(kwargs, "do_wstacking", True)
verbosity = get_parameter(kwargs, "verbosity", 0)
newbvis = copy_visibility(bvis, zero=True)
# Extracting data from BlockVisibility
freq = bvis.frequency # frequency, Hz
nrows, nants, _, vnchan, vnpol = bvis.vis.shape
uvw = newbvis.data['uvw'].reshape([nrows * nants * nants, 3])
vist = numpy.zeros([vnpol, vnchan, nants * nants * nrows],
dtype='complex')
# Get the image properties
m_nchan, m_npol, ny, nx = model.data.shape
# Check if the number of frequency channels matches in bvis and a model
# assert (m_nchan == v_nchan)
assert (m_npol == vnpol)
fuvw = uvw.copy()
# We need to flip the u and w axes. The flip in w is equivalent to the conjugation of the
# convolution function grid_visibility to griddata
fuvw[:, 0] *= -1.0
fuvw[:, 2] *= -1.0
# Find out the image size/resolution
pixsize = numpy.abs(numpy.radians(model.wcs.wcs.cdelt[0]))
# Make de-gridding over a frequency range and pol fields
vis_to_im = numpy.round(model.wcs.sub([4]).wcs_world2pix(
freq, 0)[0]).astype('int')
mfs = m_nchan == 1
if mfs:
for vpol in range(vnpol):
vist[vpol, :, :] = ng.dirty2ms(
fuvw.astype(numpy.float64),
bvis.frequency.astype(numpy.float64),
model.data[0, vpol, :, :].T.astype(numpy.float64),
pixsize_x=pixsize,
pixsize_y=pixsize,
epsilon=epsilon,
do_wstacking=do_wstacking,
nthreads=nthreads,
verbosity=verbosity).T
else:
for vpol in range(vnpol):
for vchan in range(vnchan):
imchan = vis_to_im[vchan]
vist[vpol, vchan, :] = ng.dirty2ms(
fuvw.astype(numpy.float64),
numpy.array(freq[vchan:vchan + 1]).astype(
numpy.float64),
model.data[imchan, vpol, :, :].T.astype(numpy.float64),
pixsize_x=pixsize,
pixsize_y=pixsize,
epsilon=epsilon,
do_wstacking=do_wstacking,
nthreads=nthreads,
verbosity=verbosity)[:, 0]
vis = convert_pol_frame(vist.T,
model.polarisation_frame,
bvis.polarisation_frame,
polaxis=2)
newbvis.data['vis'] = vis.reshape([nrows, nants, nants, vnchan, vnpol])
# Now we can shift the visibility from the image frame to the original visibility frame
return shift_vis_to_image(newbvis, model, tangent=True, inverse=True)
def invert_ng(bvis: BlockVisibility,
model: Image,
dopsf: bool = False,
normalize: bool = True,
**kwargs) -> (Image, numpy.ndarray):
""" Invert using nifty-gridder module
https://gitlab.mpcdf.mpg.de/ift/nifty_gridder
Use the image im as a template. Do PSF in a separate call.
In the imaging and pipeline workflows, this may be invoked using context='ng'.
:param dopsf: Make the PSF instead of the dirty image
:param bvis: BlockVisibility to be inverted
:param im: image template (not changed)
:param normalize: Normalize by the sum of weights (True)
:return: (resulting image, sum of the weights for each frequency and polarization)
"""
assert image_is_canonical(model)
assert isinstance(bvis, BlockVisibility), bvis
im = copy_image(model)
nthreads = get_parameter(kwargs, "threads", 4)
epsilon = get_parameter(kwargs, "epsilon", 1e-12)
do_wstacking = get_parameter(kwargs, "do_wstacking", True)
verbosity = get_parameter(kwargs, "verbosity", 0)
sbvis = copy_visibility(bvis)
sbvis = shift_vis_to_image(sbvis, im, tangent=True, inverse=False)
freq = sbvis.frequency # frequency, Hz
nrows, nants, _, vnchan, vnpol = sbvis.vis.shape
# if dopsf:
# sbvis = fill_vis_for_psf(sbvis)
ms = sbvis.vis.reshape([nrows * nants * nants, vnchan, vnpol])
ms = convert_pol_frame(ms,
bvis.polarisation_frame,
im.polarisation_frame,
polaxis=2)
uvw = sbvis.uvw.reshape([nrows * nants * nants, 3])
wgt = sbvis.flagged_imaging_weight.reshape(
[nrows * nants * nants, vnchan, vnpol])
if epsilon > 5.0e-6:
ms = ms.astype("c8")
wgt = wgt.astype("f4")
# Find out the image size/resolution
npixdirty = im.nwidth
pixsize = numpy.abs(numpy.radians(im.wcs.wcs.cdelt[0]))
fuvw = uvw.copy()
# We need to flip the u and w axes.
fuvw[:, 0] *= -1.0
fuvw[:, 2] *= -1.0
nchan, npol, ny, nx = im.shape
im.data[...] = 0.0
sumwt = numpy.zeros([nchan, npol])
# There's a latent problem here with the weights.
# wgt = numpy.real(convert_pol_frame(wgt, bvis.polarisation_frame, im.polarisation_frame, polaxis=2))
# Set up the conversion from visibility channels to image channels
vis_to_im = numpy.round(model.wcs.sub([4]).wcs_world2pix(
freq, 0)[0]).astype('int')
# Nifty gridder likes to receive contiguous arrays so we transpose
# at the beginning
mfs = nchan == 1
if dopsf:
mst = ms.T
mst[...] = 0.0
mst[0, ...] = 1.0
wgtt = wgt.T
if mfs:
dirty = ng.ms2dirty(fuvw.astype(numpy.float64),
bvis.frequency.astype(numpy.float64),
numpy.ascontiguousarray(mst[0, :, :].T),
numpy.ascontiguousarray(wgtt[0, :, :].T),
npixdirty,
npixdirty,
pixsize,
pixsize,
epsilon,
do_wstacking=do_wstacking,
nthreads=nthreads,
verbosity=verbosity)
sumwt[0, :] += numpy.sum(wgtt[0, 0, :].T, axis=0)
im.data[0, :] += dirty.T
else:
for vchan in range(vnchan):
ichan = vis_to_im[vchan]
frequency = numpy.array(freq[vchan:vchan + 1]).astype(
numpy.float64)
dirty = ng.ms2dirty(
fuvw.astype(numpy.float64),
frequency.astype(numpy.float64),
numpy.ascontiguousarray(mst[0,
vchan, :][...,
numpy.newaxis]),
numpy.ascontiguousarray(wgtt[0,
vchan, :][...,
numpy.newaxis]),
npixdirty,
npixdirty,
pixsize,
pixsize,
epsilon,
do_wstacking=do_wstacking,
nthreads=nthreads,
verbosity=verbosity)
sumwt[ichan, :] += numpy.sum(wgtt[0, ichan, :].T, axis=0)
im.data[ichan, :] += dirty.T
else:
mst = ms.T
wgtt = wgt.T
for pol in range(npol):
if mfs:
dirty = ng.ms2dirty(
fuvw.astype(numpy.float64),
bvis.frequency.astype(numpy.float64),
numpy.ascontiguousarray(mst[pol, :, :].T),
numpy.ascontiguousarray(wgtt[pol, :, :].T),
npixdirty,
npixdirty,
pixsize,
pixsize,
epsilon,
do_wstacking=do_wstacking,
nthreads=nthreads,
verbosity=verbosity)
sumwt[0, pol] += numpy.sum(wgtt[pol, 0, :].T, axis=0)
im.data[0, pol] += dirty.T
else:
for vchan in range(vnchan):
ichan = vis_to_im[vchan]
frequency = numpy.array(freq[vchan:vchan + 1]).astype(
numpy.float64)
dirty = ng.ms2dirty(fuvw.astype(numpy.float64),
frequency.astype(numpy.float64),
numpy.ascontiguousarray(
mst[pol,
vchan, :][...,
numpy.newaxis]),
numpy.ascontiguousarray(
wgtt[pol,
vchan, :][...,
numpy.newaxis]),
npixdirty,
npixdirty,
pixsize,
pixsize,
epsilon,
do_wstacking=do_wstacking,
nthreads=nthreads,
verbosity=verbosity)
sumwt[ichan, pol] += numpy.sum(wgtt[pol, ichan, :].T,
axis=0)
im.data[ichan, pol] += dirty.T
if normalize:
im = normalize_sumwt(im, sumwt)
return im, sumwt
def deconvolve_cube(dirty: Image,
psf: Image,
prefix='',
**kwargs) -> (Image, Image):
""" Clean using a variety of algorithms
The algorithms available are:
hogbom: Hogbom CLEAN See: Hogbom CLEAN A&A Suppl, 15, 417, (1974)
hogbom-complex: Complex Hogbom CLEAN of stokesIQUV image
msclean: MultiScale CLEAN See: Cornwell, T.J., Multiscale CLEAN (IEEE Journal of Selected Topics in Sig Proc,
2008 vol. 2 pp. 793-801)
mfsmsclean, msmfsclean, mmclean: MultiScale Multi-Frequency See: U. Rau and T. J. Cornwell,
“A multi-scale multi-frequency deconvolution algorithm for synthesis imaging in radio interferometry,” A&A 532,
A71 (2011).
For example::
comp, residual = deconvolve_cube(dirty, psf, niter=1000, gain=0.7, algorithm='msclean',
scales=[0, 3, 10, 30], threshold=0.01)
For the MFS clean, the psf must have number of channels >= 2 * nmoment
:param dirty: Image dirty image
:param psf: Image Point Spread Function
:param window_shape: Window image (Bool) - clean where True
:param mask: Window in the form of an image, overrides window_shape
:param algorithm: Cleaning algorithm: 'msclean'|'hogbom'|'mfsmsclean'
:param gain: loop gain (float) 0.7
:param threshold: Clean threshold (0.0)
:param fractional_threshold: Fractional threshold (0.01)
:param scales: Scales (in pixels) for multiscale ([0, 3, 10, 30])
:param nmoment: Number of frequency moments (default 3)
:param findpeak: Method of finding peak in mfsclean: 'Algorithm1'|'ASKAPSoft'|'CASA'|'RASCIL', Default is RASCIL.
:return: component image, residual image
See also
:py:func:`rascil.processing_components.arrays.cleaners.hogbom`
:py:func:`rascil.processing_components.arrays.cleaners.hogbom_complex`
:py:func:`rascil.processing_components.arrays.cleaners.msclean`
:py:func:`rascil.processing_components.arrays.cleaners.msmfsclean`
"""
assert isinstance(dirty, Image), dirty
assert image_is_canonical(dirty)
assert isinstance(psf, Image), psf
assert image_is_canonical(psf)
window_shape = get_parameter(kwargs, 'window_shape', None)
if window_shape == 'quarter':
log.info("deconvolve_cube %s: window is inner quarter" % prefix)
qx = dirty.shape[3] // 4
qy = dirty.shape[2] // 4
window = numpy.zeros_like(dirty.data)
window[..., (qy + 1):3 * qy, (qx + 1):3 * qx] = 1.0
log.info(
'deconvolve_cube %s: Cleaning inner quarter of each sky plane' %
prefix)
elif window_shape == 'no_edge':
edge = get_parameter(kwargs, 'window_edge', 16)
nx = dirty.shape[3]
ny = dirty.shape[2]
window = numpy.zeros_like(dirty.data)
window[..., (edge + 1):(ny - edge), (edge + 1):(nx - edge)] = 1.0
log.info(
'deconvolve_cube %s: Window omits %d-pixel edge of each sky plane'
% (prefix, edge))
elif window_shape is None:
log.info("deconvolve_cube %s: Cleaning entire image" % prefix)
window = None
else:
raise ValueError("Window shape %s is not recognized" % window_shape)
mask = get_parameter(kwargs, 'mask', None)
if isinstance(mask, Image):
if window is not None:
log.warning(
'deconvolve_cube %s: Overriding window_shape with mask image' %
(prefix))
window = mask.data
psf_support = get_parameter(kwargs, 'psf_support',
max(dirty.shape[2] // 2, dirty.shape[3] // 2))
if (psf_support <= psf.shape[2] // 2) and (
(psf_support <= psf.shape[3] // 2)):
centre = [psf.shape[2] // 2, psf.shape[3] // 2]
psf.data = psf.data[..., (centre[0] - psf_support):(centre[0] +
psf_support),
(centre[1] - psf_support):(centre[1] +
psf_support)]
log.info('deconvolve_cube %s: PSF support = +/- %d pixels' %
(prefix, psf_support))
log.info('deconvolve_cube %s: PSF shape %s' %
(prefix, str(psf.data.shape)))
algorithm = get_parameter(kwargs, 'algorithm', 'msclean')
if algorithm == 'msclean':
log.info(
"deconvolve_cube %s: Multi-scale clean of each polarisation and channel separately"
% prefix)
gain = get_parameter(kwargs, 'gain', 0.7)
assert 0.0 < gain < 2.0, "Loop gain must be between 0 and 2"
thresh = get_parameter(kwargs, 'threshold', 0.0)
assert thresh >= 0.0
niter = get_parameter(kwargs, 'niter', 100)
assert niter > 0
scales = get_parameter(kwargs, 'scales', [0, 3, 10, 30])
fracthresh = get_parameter(kwargs, 'fractional_threshold', 0.01)
assert 0.0 < fracthresh < 1.0
comp_array = numpy.zeros_like(dirty.data)
residual_array = numpy.zeros_like(dirty.data)
for channel in range(dirty.data.shape[0]):
for pol in range(dirty.data.shape[1]):
if psf.data[channel, pol, :, :].max():
log.info(
"deconvolve_cube %s: Processing pol %d, channel %d" %
(prefix, pol, channel))
if window is None:
comp_array[channel, pol, :, :], residual_array[channel, pol, :, :] = \
msclean(dirty.data[channel, pol, :, :], psf.data[channel, pol, :, :],
None, gain, thresh, niter, scales, fracthresh, prefix)
else:
comp_array[channel, pol, :, :], residual_array[channel, pol, :, :] = \
msclean(dirty.data[channel, pol, :, :], psf.data[channel, pol, :, :],
window[channel, pol, :, :], gain, thresh, niter, scales, fracthresh,
prefix)
else:
log.info(
"deconvolve_cube %s: Skipping pol %d, channel %d" %
(prefix, pol, channel))
comp_image = create_image_from_array(comp_array, dirty.wcs,
dirty.polarisation_frame)
residual_image = create_image_from_array(residual_array, dirty.wcs,
dirty.polarisation_frame)
elif algorithm == 'msmfsclean' or algorithm == 'mfsmsclean' or algorithm == 'mmclean':
findpeak = get_parameter(kwargs, "findpeak", 'RASCIL')
log.info(
"deconvolve_cube %s: Multi-scale multi-frequency clean of each polarisation separately"
% prefix)
nmoment = get_parameter(kwargs, "nmoment", 3)
assert nmoment >= 1, "Number of frequency moments must be greater than or equal to one"
nchan = dirty.shape[0]
assert nchan > 2 * (nmoment -
1), "Require nchan %d > 2 * (nmoment %d - 1)" % (
nchan, 2 * (nmoment - 1))
dirty_taylor = calculate_image_frequency_moments(dirty,
nmoment=nmoment)
if nmoment > 1:
psf_taylor = calculate_image_frequency_moments(psf,
nmoment=2 * nmoment)
else:
psf_taylor = calculate_image_frequency_moments(psf, nmoment=1)
psf_peak = numpy.max(psf_taylor.data)
dirty_taylor.data /= psf_peak
psf_taylor.data /= psf_peak
log.info("deconvolve_cube %s: Shape of Dirty moments image %s" %
(prefix, str(dirty_taylor.shape)))
log.info("deconvolve_cube %s: Shape of PSF moments image %s" %
(prefix, str(psf_taylor.shape)))
gain = get_parameter(kwargs, 'gain', 0.7)
assert 0.0 < gain < 2.0, "Loop gain must be between 0 and 2"
thresh = get_parameter(kwargs, 'threshold', 0.0)
assert thresh >= 0.0
niter = get_parameter(kwargs, 'niter', 100)
assert niter > 0
scales = get_parameter(kwargs, 'scales', [0, 3, 10, 30])
fracthresh = get_parameter(kwargs, 'fractional_threshold', 0.1)
assert 0.0 < fracthresh < 1.0
comp_array = numpy.zeros(dirty_taylor.data.shape)
residual_array = numpy.zeros(dirty_taylor.data.shape)
for pol in range(dirty_taylor.data.shape[1]):
# Always use the Stokes I PSF
if psf_taylor.data[0, 0, :, :].max():
log.info("deconvolve_cube %s: Processing pol %d" %
(prefix, pol))
if window is None:
comp_array[:, pol, :, :], residual_array[:, pol, :, :] = \
msmfsclean(dirty_taylor.data[:, pol, :, :], psf_taylor.data[:, 0, :, :],
None, gain, thresh, niter, scales, fracthresh, findpeak, prefix)
else:
log.info(
'deconvolve_cube %s: Clean window has %d valid pixels'
% (prefix, int(numpy.sum(window[0, pol]))))
comp_array[:, pol, :, :], residual_array[:, pol, :, :] = \
msmfsclean(dirty_taylor.data[:, pol, :, :], psf_taylor.data[:, 0, :, :],
window[0, pol, :, :], gain, thresh, niter, scales, fracthresh,
findpeak, prefix)
else:
log.info("deconvolve_cube %s: Skipping pol %d" % (prefix, pol))
comp_image = create_image_from_array(comp_array, dirty_taylor.wcs,
dirty.polarisation_frame)
residual_image = create_image_from_array(residual_array,
dirty_taylor.wcs,
dirty.polarisation_frame)
return_moments = get_parameter(kwargs, "return_moments", False)
if not return_moments:
log.info("deconvolve_cube %s: calculating spectral cubes" % prefix)
comp_image = calculate_image_from_frequency_moments(
dirty, comp_image)
residual_image = calculate_image_from_frequency_moments(
dirty, residual_image)
else:
log.info("deconvolve_cube %s: constructed moment cubes" % prefix)
elif algorithm == 'hogbom':
log.info(
"deconvolve_cube %s: Hogbom clean of each polarisation and channel separately"
% prefix)
gain = get_parameter(kwargs, 'gain', 0.1)
assert 0.0 < gain < 2.0, "Loop gain must be between 0 and 2"
thresh = get_parameter(kwargs, 'threshold', 0.0)
assert thresh >= 0.0
niter = get_parameter(kwargs, 'niter', 100)
assert niter > 0
fracthresh = get_parameter(kwargs, 'fractional_threshold', 0.1)
assert 0.0 < fracthresh < 1.0
comp_array = numpy.zeros(dirty.data.shape)
residual_array = numpy.zeros(dirty.data.shape)
for channel in range(dirty.data.shape[0]):
for pol in range(dirty.data.shape[1]):
if psf.data[channel, pol, :, :].max():
log.info(
"deconvolve_cube %s: Processing pol %d, channel %d" %
(prefix, pol, channel))
if window is None:
comp_array[channel, pol, :, :], residual_array[channel, pol, :, :] = \
hogbom(dirty.data[channel, pol, :, :], psf.data[channel, pol, :, :],
None, gain, thresh, niter, fracthresh, prefix)
else:
comp_array[channel, pol, :, :], residual_array[channel, pol, :, :] = \
hogbom(dirty.data[channel, pol, :, :], psf.data[channel, pol, :, :],
window[channel, pol, :, :], gain, thresh, niter, fracthresh, prefix)
else:
log.info(
"deconvolve_cube %s: Skipping pol %d, channel %d" %
(prefix, pol, channel))
comp_image = create_image_from_array(comp_array, dirty.wcs,
dirty.polarisation_frame)
residual_image = create_image_from_array(residual_array, dirty.wcs,
dirty.polarisation_frame)
elif algorithm == 'hogbom-complex':
log.info(
"deconvolve_cube_complex: Hogbom-complex clean of each polarisation and channel separately"
)
gain = get_parameter(kwargs, 'gain', 0.1)
assert 0.0 < gain < 2.0, "Loop gain must be between 0 and 2"
thresh = get_parameter(kwargs, 'threshold', 0.0)
assert thresh >= 0.0
niter = get_parameter(kwargs, 'niter', 100)
assert niter > 0
fracthresh = get_parameter(kwargs, 'fractional_threshold', 0.1)
assert 0.0 <= fracthresh < 1.0
comp_array = numpy.zeros(dirty.data.shape)
residual_array = numpy.zeros(dirty.data.shape)
for channel in range(dirty.data.shape[0]):
for pol in range(dirty.data.shape[1]):
if pol == 0 or pol == 3:
if psf.data[channel, pol, :, :].max():
log.info(
"deconvolve_cube_complex: Processing pol %d, channel %d"
% (pol, channel))
if window is None:
comp_array[channel, pol, :, :], residual_array[channel, pol, :, :] = \
hogbom(dirty.data[channel, pol, :, :], psf.data[channel, pol, :, :],
None, gain, thresh, niter, fracthresh)
else:
comp_array[channel, pol, :, :], residual_array[channel, pol, :, :] = \
hogbom(dirty.data[channel, pol, :, :], psf.data[channel, pol, :, :],
window[channel, pol, :, :], gain, thresh, niter, fracthresh)
else:
log.info(
"deconvolve_cube_complex: Skipping pol %d, channel %d"
% (pol, channel))
if pol == 1:
if psf.data[channel, 1:2, :, :].max():
log.info(
"deconvolve_cube_complex: Processing pol 1 and 2, channel %d"
% (channel))
if window is None:
comp_array[channel, 1, :, :], comp_array[
channel, 2, :, :], residual_array[
channel, 1, :, :], residual_array[
channel, 2, :, :] = hogbom_complex(
dirty.data[channel, 1, :, :],
dirty.data[channel, 2, :, :],
psf.data[channel, 1, :, :],
psf.data[channel, 2, :, :], None,
gain, thresh, niter, fracthresh)
else:
comp_array[channel, 1, :, :], comp_array[
channel, 2, :, :], residual_array[
channel, 1, :, :], residual_array[
channel, 2, :, :] = hogbom_complex(
dirty.data[channel, 1, :, :],
dirty.data[channel, 2, :, :],
psf.data[channel, 1, :, :],
psf.data[channel, 2, :, :],
window[channel, pol, :, :], gain,
thresh, niter, fracthresh)
else:
log.info(
"deconvolve_cube_complex: Skipping pol 1 and 2, channel %d"
% (channel))
if pol == 2:
continue
comp_image = create_image_from_array(
comp_array,
dirty.wcs,
polarisation_frame=PolarisationFrame('stokesIQUV'))
residual_image = create_image_from_array(
residual_array,
dirty.wcs,
polarisation_frame=PolarisationFrame('stokesIQUV'))
else:
raise ValueError('deconvolve_cube %s: Unknown algorithm %s' %
(prefix, algorithm))
return comp_image, residual_image
def image_raster_iter(im: Image,
facets=1,
overlap=0,
taper='flat',
make_flat=False) -> collections.Iterable:
"""Create an image_raster_iter generator, returning images, optionally with overlaps
The WCS is adjusted appropriately for each raster element. Hence this is a coordinate-aware
way to iterate through an image.
Provided we don't break reference semantics, memory should be conserved. However make_flat
creates a new set of images and thus reference semantics dont hold.
To update the image in place::
for r in image_raster_iter(im, facets=2):
r.data[...] = numpy.sqrt(r.data[...])
If the overlap is greater than zero, we choose to keep all images the same size so the
other ring of facets are ignored. So if facets=4 and overlap > 0 then the iterator returns
(facets-2)**2 = 4 images.
A taper is applied in the overlap regions. None implies a constant value, linear is a ramp, and
quadratic is parabolic at the ends.
:param im: Image
:param facets: Number of image partitions on each axis (2)
:param overlap: overlap in pixels
:param taper: method of tapering at the edges: 'flat' or 'linear' or 'quadratic' or 'tukey'
:param make_flat: Make the flat images
:returns: Generator of images
See also
:py:func:`rascil.processing_components.image.image_gather_facets`
:py:func:`rascil.processing_components.image.image_scatter_facets`
"""
assert image_is_canonical(im)
nchan, npol, ny, nx = im.shape
assert facets <= ny, "Cannot have more raster elements than pixels"
assert facets <= nx, "Cannot have more raster elements than pixels"
assert facets >= 1, "Facets cannot be zero or less"
assert overlap >= 0, "Overlap must be zero or greater"
if facets == 1:
yield im
else:
assert overlap < (nx // facets), "Overlap in facets is too large"
assert overlap < (ny // facets), "Overlap in facets is too large"
# Step between facets
sx = nx // facets + overlap
sy = ny // facets + overlap
# Size of facet
dx = sx + overlap
dy = sy + overlap
# Step between facets
sx = nx // facets + overlap
sy = ny // facets + overlap
# Size of facet
dx = nx // facets + 2 * overlap
dy = nx // facets + 2 * overlap
def taper_linear():
t = numpy.ones(dx)
ramp = numpy.arange(0, overlap).astype(float) / float(overlap)
t[:overlap] = ramp
t[(dx - overlap):dx] = 1.0 - ramp
result = numpy.outer(t, t)
return result
def taper_quadratic():
t = numpy.ones(dx)
ramp = numpy.arange(0, overlap).astype(float) / float(overlap)
quadratic_ramp = numpy.ones(overlap)
quadratic_ramp[0:overlap // 2] = 2.0 * ramp[0:overlap // 2]**2
quadratic_ramp[overlap //
2:] = 1 - 2.0 * ramp[overlap // 2:0:-1]**2
t[:overlap] = quadratic_ramp
t[(dx - overlap):dx] = 1.0 - quadratic_ramp
result = numpy.outer(t, t)
return result
def taper_tukey():
xs = numpy.arange(dx) / float(dx)
r = 2 * overlap / dx
t = [tukey_filter(x, r) for x in xs]
result = numpy.outer(t, t)
return result
i = 0
for fy in range(facets):
y = ny // 2 + sy * (fy - facets // 2) - overlap // 2
for fx in range(facets):
x = nx // 2 + sx * (fx - facets // 2) - overlap // 2
if (x >= 0) and (x + dx) <= nx and (y >= 0) and (y + dy) <= ny:
# Adjust WCS
wcs = im.wcs.deepcopy()
wcs.wcs.crpix[0] -= x
wcs.wcs.crpix[1] -= y
# yield image from slice (reference!)
subim = create_image_from_array(
im.data[..., y:y + dy, x:x + dx], wcs,
im.polarisation_frame)
if overlap > 0 and make_flat:
flat = create_empty_image_like(subim)
if taper == 'linear':
flat.data[..., :, :] = taper_linear()
elif taper == 'quadratic':
flat.data[..., :, :] = taper_quadratic()
elif taper == 'tukey':
flat.data[..., :, :] = taper_tukey()
else:
flat.data[...] = 1.0
yield flat
else:
yield subim
i += 1
def invert_ng(bvis: BlockVisibility,
model: Image,
dopsf: bool = False,
normalize: bool = True,
**kwargs) -> (Image, numpy.ndarray):
""" Invert using nifty-gridder module
https://gitlab.mpcdf.mpg.de/ift/nifty_gridder
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 bvis: BlockVisibility to be inverted
:param im: image template (not changed)
:param normalize: Normalize by the sum of weights (True)
:return: (resulting image, sum of the weights for each frequency and polarization)
"""
assert image_is_canonical(model)
assert isinstance(bvis, BlockVisibility), bvis
im = copy_image(model)
nthreads = get_parameter(kwargs, "threads", 4)
epsilon = get_parameter(kwargs, "epsilon", 1e-12)
do_wstacking = get_parameter(kwargs, "do_wstacking", True)
verbosity = get_parameter(kwargs, "verbosity", 0)
sbvis = copy_visibility(bvis)
sbvis = shift_vis_to_image(sbvis, im, tangent=True, inverse=False)
vis = bvis.vis
freq = sbvis.frequency # frequency, Hz
nrows, nants, _, vnchan, vnpol = vis.shape
uvw = sbvis.uvw.reshape([nrows * nants * nants, 3])
ms = vis.reshape([nrows * nants * nants, vnchan, vnpol])
wgt = sbvis.imaging_weight.reshape(
[nrows * nants * nants, vnchan, vnpol])
if dopsf:
ms[...] = 1.0 + 0.0j
if epsilon > 5.0e-6:
ms = ms.astype("c8")
wgt = wgt.astype("f4")
# Find out the image size/resolution
npixdirty = im.nwidth
pixsize = numpy.abs(numpy.radians(im.wcs.wcs.cdelt[0]))
fuvw = uvw.copy()
# We need to flip the u and w axes.
fuvw[:, 0] *= -1.0
fuvw[:, 2] *= -1.0
nchan, npol, ny, nx = im.shape
im.data[...] = 0.0
sumwt = numpy.zeros([nchan, npol])
ms = convert_pol_frame(ms,
bvis.polarisation_frame,
im.polarisation_frame,
polaxis=2)
# There's a latent problem here with the weights.
# wgt = numpy.real(convert_pol_frame(wgt, bvis.polarisation_frame, im.polarisation_frame, polaxis=2))
# Set up the conversion from visibility channels to image channels
vis_to_im = numpy.round(model.wcs.sub([4]).wcs_world2pix(
freq, 0)[0]).astype('int')
for vchan in range(vnchan):
ichan = vis_to_im[vchan]
for pol in range(npol):
# Nifty gridder likes to receive contiguous arrays
ms_1d = numpy.array([
ms[row, vchan:vchan + 1, pol]
for row in range(nrows * nants * nants)
],
dtype='complex')
ms_1d.reshape([ms_1d.shape[0], 1])
wgt_1d = numpy.array([
wgt[row, vchan:vchan + 1, pol]
for row in range(nrows * nants * nants)
])
wgt_1d.reshape([wgt_1d.shape[0], 1])
dirty = ng.ms2dirty(fuvw,
freq[vchan:vchan + 1],
ms_1d,
wgt_1d,
npixdirty,
npixdirty,
pixsize,
pixsize,
epsilon,
do_wstacking=do_wstacking,
nthreads=nthreads,
verbosity=verbosity)
sumwt[ichan, pol] += numpy.sum(wgt[:, vchan, pol])
im.data[ichan, pol] += dirty.T
if normalize:
im = normalize_sumwt(im, sumwt)
return im, sumwt
def invert_timeslice_single(vis: Visibility,
im: Image,
dopsf,
normalize=True,
remove=True,
gcfcf=None,
**kwargs) -> (Image, numpy.ndarray):
"""Process single time slice
Extracted for re-use in parallel version
The w-term can be viewed as a time-variable distortion. Approximating the array as instantaneously
co-planar, we have that w can be expressed in terms of u,v:
.. math::
w = a u + b v
Transforming to a new coordinate system:
.. math::
l' = l + a ( \\sqrt{1-l^2-m^2}-1))
.. math::
m' = m + b ( \\sqrt{1-l^2-m^2}-1))
Ignoring changes in the normalisation term, we have:
.. math::
V(u,v,w) =\\int \\frac{ I(l',m')} { \\sqrt{1-l'^2-m'^2}} e^{-2 \\pi j (ul'+um')} dl' dm'
:param vis: Visibility to be inverted
:param im: image template (not changed)
:param dopsf: Make the psf instead of the dirty image
:param gcfcf: (Grid correction function, convolution function)
:param normalize: Normalize by the sum of weights (True)
:returns: image, sum of weights
"""
assert isinstance(vis, Visibility), vis
assert image_is_canonical(im)
uvw = vis.uvw
vis, p, q = fit_uvwplane(vis, remove=remove)
workimage, sumwt = invert_2d(vis,
im,
dopsf,
normalize=normalize,
gcfcf=gcfcf,
**kwargs)
# Work image is distorted. We describe the distortion by putting the olbiquity parameters in
# the wcs. The output image should be described as having zero olbiquity parameters.
if numpy.abs(p) > 1e-7 or numpy.abs(q) > 1e-7:
# Note that this has to be zero relative in first element, one relative in second!!!!
workimage.wcs.wcs.set_pv([(0, 1, -p), (0, 2, -q)])
finalimage, footprint = reproject_image(workimage, im.wcs, im.shape)
finalimage.data[footprint.data <= 0.0] = 0.0
finalimage.wcs.wcs.set_pv([(0, 1, 0.0), (0, 2, 0.0)])
if remove:
vis.data['uvw'][...] = uvw
return finalimage, sumwt
else:
if remove:
vis.data['uvw'][...] = uvw
return workimage, sumwt
def predict_timeslice_single(vis: Visibility,
model: Image,
predict=predict_2d,
remove=True,
gcfcf=None,
**kwargs) -> Visibility:
""" Predict using a single time slices.
This fits a single plane and corrects the image geometry.
The w-term can be viewed as a time-variable distortion. Approximating the array as instantaneously
co-planar, we have that w can be expressed in terms of u,v:
.. math::
w = a u + b v
Transforming to a new coordinate system:
.. math::
l' = l + a ( \\sqrt{1-l^2-m^2}-1))
.. math::
m' = m + b ( \\sqrt{1-l^2-m^2}-1))
Ignoring changes in the normalisation term, we have:
.. math::
V(u,v,w) =\\int \\frac{ I(l',m')} { \\sqrt{1-l'^2-m'^2}} e^{-2 \\pi j (ul'+um')} dl' dm'
:param vis: Visibility to be predicted
:param model: model image
:param predict:
:param remove: Remove fitted w (so that wprojection will do the right thing)
:param gcfcf: (Grid correction function, convolution function)
:return: resulting visibility (in place works)
"""
assert image_is_canonical(model)
assert isinstance(vis, Visibility), vis
vis.data['vis'][...] = 0.0
# Fit and remove best fitting plane for this slice
uvw = vis.uvw
avis, p, q = fit_uvwplane(vis, remove=remove)
# We want to describe work image as distorted. We describe the distortion by putting
# the olbiquity parameters in the wcs. The input model should be described as having
# zero olbiquity parameters.
# Note that this has to be zero relative in first element, one relative in second!!!
if numpy.abs(p) > 1e-7 or numpy.abs(q) > 1e-7:
newwcs = model.wcs.deepcopy()
newwcs.wcs.set_pv([(0, 1, -p), (0, 2, -q)])
workimage, footprintimage = reproject_image(model,
newwcs,
shape=model.shape)
workimage.data[footprintimage.data <= 0.0] = 0.0
workimage.wcs.wcs.set_pv([(0, 1, -p), (0, 2, -q)])
# Now we can do the predict
vis = predict(avis, workimage, gcfcf=gcfcf, **kwargs)
else:
vis = predict(avis, model, gcfcf=gcfcf, **kwargs)
if remove:
avis.data['uvw'][...] = uvw
return vis
