def __init__(self, M, b, Ds=[], hypers=[]):
M = lo.aslinearoperator(M)
Ds = [lo.aslinearoperator(D) for D in Ds]
self.M = M
self.b = b
self.Ds = Ds
self.hypers = hypers
def __init__(self, M, b, Ds=[], hypers=[]):
M = lo.aslinearoperator(M)
Ds = [lo.aslinearoperator(D) for D in Ds]
self.M = M
self.b = b
self.Ds = Ds
self.hypers = hypers
def rls(A, b, Ds=[], hypers=[], optimizer=spl.cgs, **kargs):
"""Regularized Least Square
Inputs:
-------
M : model matrix, (needs matvec and rmatvec methods)
Dx : priors, (need matvec and rmatvec methods)
b : input vector
hypers: hyperparameters (scalars)
**kargs : parameters of of the least square optimizer
Outputs:
--------
x : solution
conv : convergence status
"""
verbose = getattr(kargs, 'verbose', True)
callback = getattr(kargs, 'callback', None)
# normalize hyperparameters
if callback is None:
kargs['callback'] = CallbackFactory(verbose=verbose)
A = lo.aslinearoperator(A)
X = A.T * A
for h, D in zip(hypers, Ds):
D = lo.aslinearoperator(D)
X += h * D.T * D
out, info = optimizer(X, A.T * b, **kargs)
return out
def rls(A, b, Ds=[], hypers=[], optimizer=spl.cgs, **kargs):
"""Regularized Least Square
Inputs:
-------
M : model matrix, (needs matvec and rmatvec methods)
Dx : priors, (need matvec and rmatvec methods)
b : input vector
hypers: hyperparameters (scalars)
**kargs : parameters of of the least square optimizer
Outputs:
--------
x : solution
conv : convergence status
"""
verbose = getattr(kargs, 'verbose', True)
callback = getattr(kargs, 'callback', None)
if callback is None:
kargs['callback'] = CallbackFactory(verbose=verbose)
A = lo.aslinearoperator(A)
X = A.T * A
for h, D in zip(hypers, Ds):
D = lo.aslinearoperator(D)
X += h * D.T * D
out, info = optimizer(X, A.T * b, **kargs)
return out
def __init__(self, M, b, Ds=[], hypers=[], norms=[], dnorms=[], **kargs):
M = lo.aslinearoperator(M)
Ds = [lo.aslinearoperator(D) for D in Ds]
self.M = M
self.b = b
self.Ds = Ds
self.hypers = hypers
self.norms = norms
self.dnorms = dnorms
def __init__(self, M, b, Ds=[], hypers=[], norms=[], dnorms=[], **kargs):
M = lo.aslinearoperator(M)
Ds = [lo.aslinearoperator(D) for D in Ds]
self.M = M
self.b = b
self.Ds = Ds
self.hypers = hypers
self.norms = norms
self.dnorms = dnorms
def rirls(M, y, Ds=[], tol1=1e-5, maxiter1=10, p=1, optimizer=spl.cgs, **kargs):
""" Regularized Iteratively Reweighted Least Square
"""
M = lo.aslinearoperator(M)
Ds = [lo.aslinearoperator(D) for D in Ds]
x0 = M.T * y
x = copy(x0)
r = M * x - y
tol1 /= np.sum(np.abs(r) ** p)
for i in xrange(maxiter1):
print("\n outer loop " + str(i + 1) + "\n")
A = M.T * M
for D in Ds:
rd = D * x
w = np.abs(rd) ** (p - 2)
w[np.where(1 - np.isfinite(w))] = 0 # inf
A += D.T * lo.diag(w) * D
x, t = optimizer(A, x0, **kargs)
if np.sum(np.abs(y - M * x) ** p) < tol1:
break
r = M * x - y
return x
def rirls(M, y, Ds=[], tol1=1e-5, maxiter1=10, p=1, optimizer=spl.cgs, **kargs):
""" Regularized Iteratively Reweighted Least Square
"""
M = lo.aslinearoperator(M)
Ds = [lo.aslinearoperator(D) for D in Ds]
x0 = M.T * y
x = copy(x0)
r = M * x - y
tol1 /= np.sum(np.abs(r) ** p)
for i in xrange(maxiter1):
print("\n outer loop " + str(i + 1) + "\n")
A = M.T * M
for D in Ds:
rd = D * x
w = np.abs(rd) ** (p - 2)
w[np.where(1 - np.isfinite(w))] = 0 # inf
A += D.T * lo.diag(w) * D
x, t = optimizer(A, x0, **kargs)
if np.sum(np.abs(y - M * x) ** p) < tol1:
break
r = M * x - y
return x
def irls(A0, x0, tol1=1e-5, maxiter1=10, p=1, optimizer=spl.cgs, **kargs):
""" Iteratively Reweighted Least Square
"""
A0 = lo.aslinearoperator(A0)
out = A0.T * x0
tol1 /= np.sum(np.abs(x0 - A0 * out) ** p)
for i in xrange(maxiter1):
print("\n outer loop " + str(i + 1) + "\n")
w = np.abs(x0 - A0 * out) ** (p - 2)
A = A0.T * lo.diag(w) * A0
x = A0.T * (w * x0)
out, t = optimizer(A, x, **kargs)
if np.sum(np.abs(x0 - A0 * out) ** p) < tol1:
break
return out
def irls(A0, x0, tol1=1e-5, maxiter1=10, p=1, optimizer=spl.cgs, **kargs):
""" Iteratively Reweighted Least Square
"""
A0 = lo.aslinearoperator(A0)
out = A0.T * x0
tol1 /= np.sum(np.abs(x0 - A0 * out) ** p)
for i in xrange(maxiter1):
print("\n outer loop " + str(i + 1) + "\n")
w = np.abs(x0 - A0 * out) ** (p - 2)
A = A0.T * lo.diag(w) * A0
x = A0.T * (w * x0)
out, t = optimizer(A, x, **kargs)
if np.sum(np.abs(x0 - A0 * out) ** p) < tol1:
break
return out
def gacg(M,
y,
Ds=[],
hypers=[],
norms=[],
dnorms=[],
tol=1e-6,
x0=None,
maxiter=None,
callback=None,
**kwargs):
"""Generalized approximate conjugate gradient
Approximate conjugate gradient is a gradient method with a polak
ribiere update.
It is generalized since it does not assume quadratic norm.
Inputs:
-------
M : model matrix, (needs matvec and rmatvec methods)
Ds : priors, (need matvec and rmatvec methods)
norms : norms of the likelihood and the priors
dnorms : derivation of the norm
b : input vector
hypers: hyperparameters (scalars)
Outputs:
--------
x : solution
"""
if callback is None:
callback = CallbackFactory(verbose=True, criterion=True)
# ensure linear operators are passed
M = lo.aslinearoperator(M)
Ds = [lo.aslinearoperator(D) for D in Ds]
# first guess
if x0 is None:
x = M.T * y
else:
x = copy(x0)
# normalize hyperparameters
hypers = normalize_hyper(hypers, y, x)
# tolerance
r = M * x - y
rd = [D * x for D in Ds]
J = criterion(hypers=hypers, norms=norms, Ds=Ds, r=r, rd=rd)
Jnorm = copy(J)
resid = 2 * tol
# maxiter
if maxiter is None:
maxiter = x.size
iter_ = 0
# main loop
while iter_ < maxiter and resid > tol:
iter_ += 1
# gradient
g, ng = gradient(hypers=hypers, dnorms=dnorms, M=M, Ds=Ds, r=r, rd=rd)
# descent direction
if (iter_ % 10) == 1:
d = -g
else:
b = ng / ng0
d = -g + b * d
g0 = copy(g)
ng0 = copy(ng)
# step
if norms[0] == norm2:
a = quadratic_optimal_step(d, g, M, hypers, Ds)
else:
a = backtracking_line_search(d,
g,
M,
hypers,
Ds,
x,
norms=norms,
y=y,
f0=J)
# update
x += a * d
# residual
r = M * x - y
rd = [D * x for D in Ds]
# criterion
J_old = copy(J)
J = criterion(hypers=hypers, norms=norms, Ds=Ds, r=r, rd=rd)
#resid = J / Jnorm
resid = (J_old - J) / Jnorm
callback(x)
# define output
if resid > tol:
info = resid
else:
info = 0
return x #, info
# deglitch uctod = uncompress(ctod, C, factor) uctod.mask = tod.mask uctod = tm.filter_median(uctod, length=deglitch_filter_length) uctod_mask = tm.deglitch_l2mad(uctod, projection) uctod = uncompress(ctod, C, factor) uctod.mask = uctod_mask uctod = tm.interpolate_linear(uctod) ctod = compress(uctod, C, factor) # filtering ctod = tm.filter_median(ctod, length=filter_length / factor) cov = noise_covariance(ctod, obs) S = cov.aslinearoperator() # model with compression M = lo.aslinearoperator(model.aslinearoperator()) M = C * M # noise covariance cov = noise_covariance(ctod, obs) S = lo.aslinearoperator(cov.aslinearoperator()) #S = C * S * C.T #S = None # backprojection backmap = (M.T * ctod.flatten()).reshape(projection.shapein) # priors Ds = [lo.diff(backmap.shape, axis=0, dtype=np.float64),] Ds.append(lo.diff(backmap.shape, axis=1, dtype=np.float64)) # weights weights = projection.transpose(tod.ones(tod.shape)) # masking the map MM = lo.mask(weights == 0)
def gacg(M, y, Ds=[], hypers=[], norms=[], dnorms=[], tol=1e-6, x0=None, maxiter=None,
callback=None, **kwargs):
"""Generalized approximate conjugate gradient
Approximate conjugate gradient is a gradient method with a polak
ribiere update.
It is generalized since it does not assume quadratic norm.
Inputs:
-------
M : model matrix, (needs matvec and rmatvec methods)
Ds : priors, (need matvec and rmatvec methods)
norms : norms of the likelihood and the priors
dnorms : derivation of the norm
b : input vector
hypers: hyperparameters (scalars)
Outputs:
--------
x : solution
"""
if callback is None:
callback = CallbackFactory(verbose=True)
# ensure linear operators are passed
M = lo.aslinearoperator(M)
Ds = [lo.aslinearoperator(D) for D in Ds]
# first guess
if x0 is None:
x = M.T * y
else:
x = copy(x0)
# tolerance
r = M * x - y
rd = [D * x for D in Ds]
J = criterion(hypers=hypers, norms=norms, Ds=Ds, r=r, rd=rd)
Jnorm = copy(J)
resid = 2 * tol
# maxiter
if maxiter is None:
maxiter = x.size
iter_ = 0
# main loop
while iter_ < maxiter and resid > tol:
iter_ += 1
# gradient
g, ng = gradient(hypers=hypers, dnorms=dnorms, M=M, Ds=Ds, r=r, rd=rd)
# descent direction
if (iter_ % 10) == 1:
d = - g
else:
b = ng / ng0
d = - g + b * d
g0 = copy(g)
ng0 = copy(ng)
# step
if norms[0] == norm2:
a = quadratic_optimal_step(d, g, M, hypers, Ds)
else:
a = backtracking_line_search(d, g, M, hypers, Ds,
x, norms, f0=J)
# update
x += a * d
# residual
r = M * x - y
rd = [D * x for D in Ds]
# criterion
J = criterion(hypers=hypers, norms=norms, Ds=Ds, r=r, rd=rd)
resid = J / Jnorm
callback(x)
# define output
if resid > tol:
info = resid
else:
info = 0
return x#, info
detector_policy='remove')
tod = obs.get_tod()
header = obs.get_map_header()
header.update('CDELT1', resolution / 3600)
header.update('CDELT2', resolution / 3600)
npix = 5
good_npix = False
projection = tm.Projection(obs, header=header,
resolution=resolution,
oversampling=False,
npixels_per_sample=npix)
model = projection
#C = csh.averaging(tod.shape, factor=factor)
compression_shape = [tod.size / factor, tod.size, ]
C = tm.CompressionAverage(compression_factor=factor).aslinearoperator(compression_shape)
C = lo.aslinearoperator(C)
# compress data
ctod = compress(tod, C, factor)
# uncompress for preprocessing
uctod = uncompress(ctod, C, factor)
# XXX need mask ...
uctod.mask = tod.mask
# deglitching
uctod = tm.filter_median(uctod, length=10)
uctod.mask = tm.deglitch_l2mad(uctod, projection)
masking = tm.Masking(uctod.mask)
model = masking * projection
# filtering
ctod = tm.filter_median(ctod, length=filter_length / factor)
cov = noise_covariance(ctod, obs)
S = cov.aslinearoperator()
"scan_angle":0.,
"scan_nlegs":1,
"scan_length":4.,
"compression_factor":1,
"scan_speed":60.0
}
pointing = tm.pacs_create_scan(ra0, dec0, **pointing_params)
# create obs object
obs = tm.PacsSimulation(pointing, "red", policy_bad_detector="keep")
# to define my compression matrix I need only 8 frames
# so I mask the others :
obs.pointing.removed[:] = True
obs.pointing.removed[201:209] = False
# create projector
projection = tm.Projection(obs)
P = lo.aslinearoperator(projection)
# define masking of coefficients
# ------------------------------
# number of coefficients to keep
factor = 8.
nk = P.shape[0] / factor
# coverage map
w = P.T * np.ones(P.shape[0])
# sort coverage map coefficients
ws = np.sort(w)
# find the nk largest coverage elements
threshold = ws[-nk]
mask = w < threshold
# define decimation Linear Operator according to coverage map coef
# selection.
tod = pacs.get_tod()
# projector
projection = tm.Projection(pacs, resolution=3.2,
oversampling=True, npixels_per_sample=6)
masking = tm.Masking(tod.mask)
compression = tm.CompressionAverage(pacs.compression_factor)
model = masking * compression * projection
# naive map
naive = model.transpose(tod)
# coverage map
coverage = model.transpose(tod.ones(tod.shape))
# noise covariance
length = 2**np.ceil(np.log2(np.array(tod.nsamples) + 200))
invNtt = tm.InvNtt(length, pacs.get_filter_uncorrelated())
fft = tm.Fft(length)
padding = tm.Padding(left=invNtt.ncorrelations,
right=length - tod.nsamples - invNtt.ncorrelations)
weight = padding.T * fft.T * invNtt * fft * padding
W = lo.aslinearoperator(weight.aslinearoperator())
# transform to lo
P = lo.aslinearoperator(model.aslinearoperator())
# priors
Ds = [lo.diff(naive.shape, axis=axis) for axis in xrange(naive.ndim)]
# inversion
hypers = [1e6, 1e6, ]
y = tod.flatten()
M = P.T * W * P + np.sum([h * D.T * D for h, D in zip(hypers, Ds)])
x, conv = spl.cgs(M, P.T * W * y, callback=lo.CallbackFactory(verbose=True))
sol = naive.zeros(naive.shape)
sol[:] = x.reshape(sol.shape)
#uctod[:] = (C.T * ctod.flatten()).reshape(tod.shape)
# deglitching
projection = tm.Projection(pacs,
header=header,
resolution=3.,
npixels_per_sample=6)
uctod.mask = tm.deglitch_l2mad(uctod, projection)
#ctod = compression.direct(uctod)
# model
masking = tm.Masking(uctod.mask)
model = masking * projection
# remove drift
#ctod = tm.filter_median(ctod, length=3000 / 8.)
# first map
M = C * lo.aslinearoperator(model.aslinearoperator())
#P = lo.aslinearoperator(projection.aslinearoperator())
#C = csh.averaging(tod.shape, factor=8)
#I = lo.mask(uctod.mask)
#M = C * I.T * I * P
#M = C * P
backmap = (M.T * ctod.flatten()).reshape(projection.shapein)
#weights = (M.T * np.ones(ctod.size)).reshape(projection.shapein)
weights = projection.transpose(tod.ones(tod.shape))
MM = lo.mask(weights == 0)
M = M * MM.T
# define algo
# priors
Dx = lo.diff(backmap.shape, axis=0, dtype=np.float64) * MM.T
Dy = lo.diff(backmap.shape, axis=1, dtype=np.float64) * MM.T
#Dw = lo.pywt_lo.wavedec2(backmap.shape, "haar", level=3)
naive = backmap / coverage
# mask according to coverage (everything that is covered by less than 10.)
mask = Masking(coverage < 10.)
# The model is the masking of the sky map then the projection
# This is basically matrix multiplication
model = projection * mask
# Performing inversion
# ---------------------
# with TAMASIS
x_tm = mapper_rls(tod, model, hyper=1e-1, tol=1e-10, maxiter=100)
# with lo routines
# transform to lo
H = lo.aslinearoperator(model * mask)
# smoothness priors
Ds = [lo.diff(backmap.shape, axis=axis) for axis in (0, 1)]
# inversion
y = tod.ravel() # requires 1d input
x_lo = lo.acg(H, y, Ds, 1e-1 * np.ones(3), tol=1e-10, maxiter=100)
x_lo.resize(backmap.shape) # output is 1d so need reshaping
# with sparsity assumptions (using Huber algorithm)
x_h = lo.hacg(H,
y,
Ds,
1e1 * np.ones(3),
np.asarray((None, 1e-6, 1e-6, 1e-6)),
x0=x_lo.flatten(),
tol=1e-7,
header = pacs.get_map_header()
#header['NAXIS1'] = 192
#header['NAXIS2'] = 192
#header['CRPIX1'] = 96
#header['CRPIX2'] = 96
header.update('CDELT1', resolution / 3600)
header.update('CDELT2', resolution / 3600)
# data
tod = pacs.get_tod()
y = tod.flatten()
# remove bad pixels (by updating mask !)
#tod = remove_bad_pixels(tod)
# compress data
factor = 8
compression = tm.CompressionAverage(factor)
C = lo.aslinearoperator(compression.aslinearoperator(shape=(tod.size / factor, tod.size)))
ctod = compression.direct(tod)
# uncompress for deglitching
uctod = tod.copy(tod.shape)
y0, t = lo.spl.cgs(C.T * C, C.T * ctod.flatten())
uctod[:] = y0.reshape(tod.shape)
# deglitching
projection = tm.Projection(pacs, header=header, resolution=3., npixels_per_sample=5)
uctod.mask = tm.deglitch_l2mad(uctod, projection)
ctod = compression.direct(uctod)
# model
masking = tm.Masking(uctod.mask)
model = compression * masking * projection
# remove drift
#ctod = tm.filter_median(ctod, length=3000 / 8.)
# first map
coverage = projection.T(np.ones(tod.shape))
# naive map
naive = backmap / coverage
# mask according to coverage (everything that is covered by less than 10.)
mask = Masking(coverage < 10.)
# The model is the masking of the sky map then the projection
# This is basically matrix multiplication
model = projection * mask
# Performing inversion
# ---------------------
# with TAMASIS
x_tm = mapper_rls(tod, model, hyper=1e-1, tol=1e-10, maxiter=100)
# with lo routines
# transform to lo
H = lo.aslinearoperator(model * mask)
# smoothness priors
Ds = [lo.diff(backmap.shape, axis=axis) for axis in (0, 1)]
# inversion
y = tod.ravel() # requires 1d input
x_lo = lo.acg(H, y, Ds, 1e-1 * np.ones(3), tol=1e-10, maxiter=100)
x_lo.resize(backmap.shape) # output is 1d so need reshaping
# with sparsity assumptions (using Huber algorithm)
x_h = lo.hacg(H, y, Ds, 1e1 * np.ones(3),
np.asarray((None, 1e-6, 1e-6, 1e-6)),
x0=x_lo.flatten(), tol=1e-7, maxiter=200)
x_h.resize(backmap.shape)
header = pacs.get_map_header() header['NAXIS1'] = 192 header['NAXIS2'] = 192 header['CRPIX1'] = 96 header['CRPIX2'] = 96 # data tod = pacs.get_tod() # remove bad pixels (by updating mask !) #tod = remove_bad_pixels(tod) # deglitching projection = Projection(pacs, header=header, resolution=3., npixels_per_sample=6) deglitch_l2mad(tod, projection) # model masking = Masking(tod.mask) model = masking * projection P = lo.aslinearoperator(model.aslinearoperator()) # derive filter #tod = filter_median(tod, length=9999) #kernel = filt.kernel_from_tod(tod, length=1000) kernel = (1 + (10. / np.arange(500)) ** .25) kernel = np.concatenate((kernel[::-1], kernel)) #kern = np.mean(kernel, axis=0) N = filt.kernels_convolve(tod.shape, 1 / np.sqrt(kernel)) # apply to data yn = N * tod.flatten() # apply to model M = N * P # first map backmap = model.transpose(tod) # define algo # priors
npix = 5
good_npix = False
projection = tm.Projection(obs,
header=header,
resolution=resolution,
oversampling=False,
npixels_per_sample=npix)
model = projection
#C = csh.averaging(tod.shape, factor=factor)
compression_shape = [
tod.size / factor,
tod.size,
]
C = tm.CompressionAverage(
compression_factor=factor).aslinearoperator(compression_shape)
C = lo.aslinearoperator(C)
# compress data
ctod = compress(tod, C, factor)
# uncompress for preprocessing
uctod = uncompress(ctod, C, factor)
# XXX need mask ...
uctod.mask = tod.mask
# deglitching
uctod = tm.filter_median(uctod, length=10)
uctod.mask = tm.deglitch_l2mad(uctod, projection)
masking = tm.Masking(uctod.mask)
model = masking * projection
# filtering
ctod = tm.filter_median(ctod, length=filter_length / factor)
cov = noise_covariance(ctod, obs)
S = cov.aslinearoperator()
'cam_angle': 0.,
"scan_angle": 0.,
"scan_nlegs": 1,
"scan_length": 400.,
"scan_speed": 60.0,
"compression_factor": 1 # we will compress later
}
pointing = tm.pacs_create_scan(ra0, dec0, **pointing_params)
# create obs object
obs = tm.PacsSimulation(pointing, "red", policy_bad_detector="keep")
# we dont need 200 first and last frames
obs.pointing.removed[200:] = True
obs.pointing.removed[-201:] = True
# create projector
projection = tm.Projection(obs, npixels_per_sample=4)
P = lo.aslinearoperator(projection)
# simulate data
x0 = tm.gaussian(projection.shapein, 3) # map with gaussian source
tod0 = projection(x0) # data
n = np.random.randn(*tod0.shape) # noise
nsr = 1e-2
tod = tod0 + nsr * n # noisy data
y = tod.ravel() # as 1d array
# load compression matrix
filename = os.path.join(os.getenv("HOME"), "data", "pacs",
"mmc_cam_angle_0_scan_angle_0_speed60.fits")
c = fa.FitsArray(file=filename).astype(np.float64)
cmm = csh.compression.AnyOperator(c)
C = cmm((projection.shapeout[0], projection.shapeout[1][0]), )
# compress
import lo.pywt_lo as pywt_lo
# data
pacs = PacsObservation(filename=tamasis_dir + 'tests/frames_blue.fits',
fine_sampling_factor=1,
keep_bad_detectors=False)
tod = pacs.get_tod()
# projector
projection = Projection(pacs,
resolution=3.2,
oversampling=False,
npixels_per_sample=6)
model = projection
# naive map
backmap = model.transpose(tod)
# mask
M = lo.mask(backmap < 0).T
#M = lo.identity(2 * (backmap.size,))
# transform to lo
P = lo.aslinearoperator(model.aslinearoperator()) * M
# prior
Dx = lo.diff(backmap.shape, axis=0, dtype=np.float64) * M
Dy = lo.diff(backmap.shape, axis=1, dtype=np.float64) * M
Dw = pywt_lo.wavelet2(backmap.shape, "haar") * M
# inversion
y = tod.flatten()
callback = lo.CallbackFactory(verbose=True)
x = it.rirls(P, (Dx, Dy, Dw), y, p=1.5, maxiter=100, callback=callback)
sol = backmap.zeros(backmap.shape)
sol[:] = (M * x).reshape(sol.shape)
header = pacs.get_map_header()
#header['NAXIS1'] = 192
#header['NAXIS2'] = 192
#header['CRPIX1'] = 96
#header['CRPIX2'] = 96
header.update('CDELT1', resolution / 3600)
header.update('CDELT2', resolution / 3600)
# data
tod = pacs.get_tod()
y = tod.flatten()
# remove bad pixels (by updating mask !)
#tod = remove_bad_pixels(tod)
# compress data
factor = 8
compression = tm.CompressionAverage(factor)
C = lo.aslinearoperator(
compression.aslinearoperator(shape=(tod.size / factor, tod.size)))
ctod = compression.direct(tod)
# uncompress for deglitching
uctod = tod.copy(tod.shape)
y0, t = lo.spl.cgs(C.T * C, C.T * ctod.flatten())
uctod[:] = y0.reshape(tod.shape)
# deglitching
projection = tm.Projection(pacs,
header=header,
resolution=3.,
npixels_per_sample=5)
uctod.mask = tm.deglitch_l2mad(uctod, projection)
ctod = compression.direct(uctod)
# model
masking = tm.Masking(uctod.mask)
model = compression * masking * projection