def test_tvgen_nd():
"""Test that the general solver does not crash for high-d tensors"""
for _ in range(20):
dims = np.random.randint(3, 5)
shape = np.random.randint(2, 10, size=dims)
x = np.random.randn(*shape)
w = np.random.randn(dims)
tvgen(x, w, list(range(1, dims+1)), np.ones(dims))
def test_tvgen_nd():
"""Test that the general solver does not crash for high-d tensors"""
for _ in range(20):
dims = np.random.randint(3, 5)
shape = np.random.randint(2, 10, size=dims)
x = np.random.randn(*shape)
w = np.random.randn(dims)
tvgen(x, w, list(range(1, dims + 1)), np.ones(dims))
def TTVOneIteration(x, z, t, imgs, A, s0, args):
tikh = args.tikh * s0
# SGD
r = x.reshape([-1, x.shape[-1]]).T
c = imgs.reshape([-1, imgs.shape[-1]]).T
q1 = A @ r - c
q2 = tikh * r
q1 = A.T @ q1
q2 = tikh * q2
aq1 = A @ q1
aq2 = tikh * q2
s = np.sum((q1 + q2)**2) / (np.sum(aq1**2) + np.sum(aq2**2))
zNew = r - s * (q1 + q2)
zNew = np.reshape(zNew.T, x.shape)
# TV
zNew = prox_tv.tvgen(zNew, args.betas, [1,2,3], [1,1,1], n_threads = 8, max_iters = 10)
# acceleration
tNew = (1 + np.sqrt(1 + 4 * t * t)) / 2
xNew = zNew + ((t - 1) / tNew) * (zNew - z)
return xNew, zNew, tNew
def test_tvgen_2d():
"""Tests that the general solver returns correct 2d solutions"""
for _ in range(20):
x = _generate2D()
w = 20 * np.random.rand()
specific = tv1_2d(x, w, max_iters=1000)
general = tvgen(x, [w, w], [1, 2], [1, 1], max_iters=1000)
assert np.allclose(specific, general, atol=1e-2)
def test_tvgen_2d():
"""Tests that the general solver returns correct 2d solutions"""
for _ in range(20):
x = _generate2d()
w = 20*np.random.rand()
specific = tv1_2d(x, w, max_iters=1000)
general = tvgen(x, [w, w], [1, 2], [1, 1], max_iters=1000)
assert np.allclose(specific, general, atol=1e-2)
def test_tvgen_1d():
"""Tests that the general solver returns correct 1d solutions"""
for _ in range(20):
dimension = np.random.randint(1e1, 3e1)
x = 100 * np.random.randn(dimension)
w = 20 * np.random.rand()
specific = tv1_1d(x, w)
general = tvgen(x, [w], [1], [1])
assert np.allclose(specific, general, atol=1e-3)
def test_tvgen_1d():
"""Tests that the general solver returns correct 1d solutions"""
for _ in range(20):
dimension = np.random.randint(1e1, 3e1)
x = 100*np.random.randn(dimension)
w = 20*np.random.rand()
specific = tv1_1d(x, w)
general = tvgen(x, [w], [1], [1])
assert np.allclose(specific, general, atol=1e-3)
def test_tvgen_multireg():
"""Test applying several regularizers on same dimension"""
for _ in range(20):
x = _generate2D()
w = 20 * np.random.rand()
specific = tv1_2d(x, w, max_iters=1000)
general = tvgen(x, [w / 2., w / 2., w / 3., w / 3., w / 3.],
[1, 1, 2, 2, 2], [1, 1, 1, 1, 1],
max_iters=1000)
print("Max diff: " + str((specific - general).max()))
assert np.allclose(specific, general, atol=1e-2)
def prox_FL(a,
beta,
lamda,
p=1,
symmetric=False,
use_matlab=False,
optimize=True):
"""Fused Lasso prox.
It is calculated as the Total variation prox + soft thresholding
on the solution, as in
http://ieeexplore.ieee.org/abstract/document/6579659/
"""
# if any([any(np.diag(x) < 0) for x in a]):
# for a_i in a:
# np.fill_diagonal(a_i, np.sum(np.abs(a_i), axis=1))
if optimize:
Y = tvgen(a, [beta], [1], [p], n_threads=32, max_iters=30)
else:
Y = np.empty_like(a)
# if use_matlab:
# from regain.wrapper.tv_condat import wrapper
# func = wrapper.total_variation_condat
# else:
func = tv1_1d if p == 1 else partial(tvp_1d, p=p)
if symmetric:
x, y = np.triu_indices_from(a[0])
b = np.vstack(a.transpose(1, 2, 0))
upper_ind = x * a.shape[1] + y
Z = np.zeros_like(b)
Z[upper_ind] = [func(row, beta) for row in b[upper_ind]]
e = np.array(np.split(Z, a.shape[1], axis=0)).transpose(2, 0, 1)
Y = (e + e.transpose(0, 2, 1)) / (
np.array([np.eye(a.shape[1]) for i in range(a.shape[0])]) + 1)
else:
for i in range(np.power(a.shape[1], 2)):
solution = func(a.flat[i::np.power(a.shape[1], 2)], beta)
Y.flat[i::np.power(a.shape[1], 2)] = solution
# fused-lasso (soft-thresholding on the solution)
Y_soft = soft_thresholding(Y, lamda)
# restore diagonal
for y_s, y_fv in zip(Y_soft, Y):
np.fill_diagonal(y_s, np.diag(y_fv))
return Y_soft
def test_tvgen_multireg():
"""Test applying several regularizers on same dimension"""
for _ in range(20):
x = _generate2d()
w = 20*np.random.rand()
specific = tv1_2d(x, w, max_iters=1000)
general = tvgen(
x,
[w/2., w/2., w/3., w/3., w/3.],
[1, 1, 2, 2, 2],
[1, 1, 1, 1, 1],
max_iters=1000
)
print("Max diff: " + str((specific-general).max()))
assert np.allclose(specific, general, atol=1e-2)
def trendFilter3D(tensor, lamb, k=0):
if isinstance(lamb, tuple):
lamb_x, lamb_y, lamb_z = lamb
else:
lamb_x, lamb_y, lamb_z = lamb, lamb, lamb
if k == 0:
filtered_block = ptv.tvgen(tensor, [lamb_x, lamb_y, lamb_z], [1, 2, 3],
[1, 1, 1])
elif k == 1:
pass
elif k == 2:
pass
else:
raise NotImplementedError
return filtered_block
def _prox_s(x, reg):
return tvgen(x, [reg, reg, reg], [1, 2, 3], [1, 1, 1])
from skimage import io, util
import numpy as np
# Load color image (3 dimensions: length, width and color)
X = io.imread('colors.png')
X = ski.img_as_float(X)
# Introduce noise
noiseLevel = 0.2
N = util.random_noise(X, mode='gaussian', var=noiseLevel)
# Filter using 3D TV-L1: we only want to penalize X and Y dims (not color)
lam = 100. / 255.
print('Filtering image...')
start = time.time()
F = ptv.tvgen(N, np.array([lam, lam]), [1, 2], np.array([1, 1]))
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
end = time.time()
print('Elapsed time ' + str(end - start))
# Any dimension can be penalized under any norm. By also penalizing the color dimension under TV-L2 we get a "decolored" image
lam2 = 50. / 255.
print('Color filtering...')
start = time.time()
F2 = ptv.tvgen(N, np.array([lam, lam, lam2]), [1, 2, 3], np.array([1, 1, 2]))
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
end = time.time()
print('Elapsed time ' + str(end - start))
# Plot results
plt.subplot(2, 2, 1)
from skimage import io, util
import numpy as np
# Load color image (3 dimensions: length, width and color)
X = io.imread('colors.png')
X = ski.img_as_float(X)
# Introduce noise
noiseLevel = 0.2
N = util.random_noise(X, mode='gaussian', var=noiseLevel)
# Filter using 3D TV-L1: we only want to penalize X and Y dims (not color)
lam=100./255.;
print('Filtering image...')
start = time.time()
F = ptv.tvgen(N, np.array([lam, lam]), [1, 2], np.array([1, 1]))
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
end = time.time()
print('Elapsed time ' + str(end-start))
# Any dimension can be penalized under any norm. By also penalizing the color dimension under TV-L2 we get a "decolored" image
lam2=50./255.;
print('Color filtering...')
start = time.time()
F2 = ptv.tvgen(N, np.array([lam, lam, lam2]), [1, 2, 3], np.array([1, 1, 2]));
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
end = time.time()
print('Elapsed time ' + str(end-start))
# Plot results
plt.subplot(2, 2, 1)
import skimage as ski
from skimage import data, io, filters, color, util
# Load color image (3 dimensions: length, width and color)
X = io.imread('colors.png')
X = ski.img_as_float(X)
# Introduce noise
noiseLevel = 0.2
N = util.random_noise(X, mode='gaussian', var=noiseLevel)
# Filter using 3D TV-L1: we only want to penalize X and Y dims (not color)
lam = 100. / 255.
print('Filtering image...')
start = time.time()
F = ptv.tvgen(N, [lam, lam], [1, 2], [1, 1])
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
end = time.time()
print('Elapsed time ' + str(end - start))
# Any dimension can be penalized under any norm. By also penalizing the color dimension under TV-L2 we get a "decolored" image
lam2 = 50. / 255.
print('Color filtering...')
start = time.time()
F2 = ptv.tvgen(N, [lam, lam, lam2], [1, 2, 3], [1, 1, 2])
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
end = time.time()
print('Elapsed time ' + str(end - start))
# Plot results
plt.subplot(2, 2, 1)
input_filename = sys.argv[1] basename = os.path.basename(input_filename) filename = os.path.splitext(basename)[0] output_folder = sys.argv[2] output_extension = "nrrd" lam = float(sys.argv[3]) output_filename = os.path.join(output_folder, filename + "_tv" + str(lam) + "." + output_extension) # PixelType = itk.F # PixelType = itk.UC # Dimension = 3 # ImageType = itk.Image[PixelType, Dimension] # reader = itk.ImageFileReader[ImageType].New(FileName=input_filename) reader = itk.ImageFileReader.New(FileName=input_filename) reader.Update() image = reader.GetOutput() image_array = itk.GetArrayViewFromImage(image) norm = 1 dimensions_to_penalyze = [1,2,3] tv_array = ptv.tvgen(image_array, np.array([lam,lam,lam]), dimensions_to_penalyze, np.array([norm,norm,norm])) tv_array = np.ascontiguousarray(tv_array, dtype=image_array.dtype) # modifiedImage = itk.PyBuffer[ImageType].GetImageViewFromArray(tv_array) modifiedImage = itk.GetImageViewFromArray(tv_array) itk.ImageFileWriter.New(Input=modifiedImage, FileName=output_filename).Update() # from subprocess import Popen # testDriver = "/home/phc/tmp/IsotropicWaveletsTestDriver" # pin = Popen([testDriver, 'runViewImage', input_filename, 'input']) # pout = Popen([testDriver, 'runViewImage', output_filename, 'TV'])
uft = UFT(samp0, axes=ax, scale=True)
forward = uft.forward_ortho
inverse = uft.inverse_ortho
y = forward(x)
# view(y, log=True)
imspace_u = inverse(y)
# from mr_utils.cs import SpatioTemporalTVSB
# recon, err = SpatioTemporalTVSB(
# samp0, y, betaxy=1/4, betat=1, mu=1, lam=1, gamma=1/2,
# nInner=1, niter=3, x=x)
# view(recon)
w = 50
recon_l1 = ptv.tvgen(
np.abs(imspace_u), np.array([w]), [3], np.array([1]))
recon_l2 = ptv.tvgen(
np.abs(imspace_u), np.array([w]), [3], np.array([2]))
recon_sort = np.zeros_like(imspace_u)
roi_cnt = 0
for idx in tqdm(np.ndindex((sx, sy)), total=sx*sy, leave=False):
ii, jj = idx[:]
# TODO:
# Mask out small ROIs to do run ordinator in
if roi[ii, jj]:
# sord = ordinator1d(
# np.abs(imspace_true[ii, jj, :]), k=10,
# forward=S.forward_wvlt, inverse=S.inverse_wvlt,
# chunksize=10, pdf=None, pdf_metric=None,
import skimage as ski
from skimage import data, io, filters, color, util
# Load color image (3 dimensions: length, width and color)
X = io.imread('colors.png')
X = ski.img_as_float(X)
# Introduce noise
noiseLevel = 0.2
N = util.random_noise(X, mode='gaussian', var=noiseLevel)
# Filter using 3D TV-L1: we only want to penalize X and Y dims (not color)
lam=100./255.;
print('Filtering image...')
start = time.time()
F = ptv.tvgen(N, [lam, lam], [1, 2], [1, 1])
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
end = time.time()
print('Elapsed time ' + str(end-start))
# Any dimension can be penalized under any norm. By also penalizing the color dimension under TV-L2 we get a "decolored" image
lam2=50./255.;
print('Color filtering...')
start = time.time()
F2 = ptv.tvgen(N, [lam, lam, lam2], [1, 2, 3], [1, 1, 2]);
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
end = time.time()
print('Elapsed time ' + str(end-start))
# Plot results
plt.subplot(2, 2, 1)
