def forward(self, x, weights_row, weights_col):
r"""Solve the total variation problem and return the solution.
Arguments
---------
x: :class:`torch:torch.Tensor`
A tensor with shape ``(m, n)`` holding the input signal.
weights_row: :class:`torch:torch.Tensor`
The horizontal edge weights.
Tensor of shape ``(m, n - 1)``, or ``(1,)`` if all weights
are equal.
weights_col: :class:`torch:torch.Tensor`
The vertical edge weights.
Tensor of shape ``(m - 1, n)``, or ``(1,)`` if all weights
are equal.
Returns
-------
:class:`torch:torch.Tensor`
The solution to the total variation problem, of shape ``(m, n)``.
"""
opt = tv1w_2d(x.numpy(), weights_col.numpy(), weights_row.numpy(),
**self.tv_args)
if self.refine:
opt = self._refine(opt, x, weights_row, weights_col)
opt = torch.Tensor(opt).view_as(x)
self.save_for_backward(opt)
return opt
def solve_and_refine(x, w_col, w_row, refine=True, **tv_args):
opt = tv1w_2d(x, w_col, w_row, **tv_args)
if refine:
opt = TotalVariationBase._refine(opt, x, w_row, w_col)
return opt
def test_tv1w_2d_emengd():
r"""Issue reported by emengd
Make the solver fail due to missing checks on integer arguments
"""
a = -np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])/10.
sol1 = tv1w_2d(a, np.array([[1, 1, 1], [1, 1, 1]]),
np.array([[1, 1], [1, 1], [1, 1]]), max_iters=100)
sol2 = tv1_2d(a, 1)
assert np.allclose(sol1, sol2, atol=1e-3)
def test_tv1w_2d_uniform_weights():
for _ in range(20):
x = _generate2D()
rows = len(x)
cols = len(x[0])
w1 = np.random.rand()
w_rows = np.ones([rows - 1, cols]) * w1
w_cols = np.ones([rows, cols - 1]) * w1
solw = tv1w_2d(x, w_rows, w_cols, max_iters=5000)
solw1 = tv1_2d(x, w1, max_iters=5000)
assert np.allclose(solw, solw1, atol=1e-3)
def test_tv1w_2d_uniform_weights():
for _ in range(20):
rows = np.random.randint(1e1, 3e1)
cols = np.random.randint(1e1, 3e1)
x = 100*np.random.randn(rows, cols)
w1 = np.random.rand()
w_rows = np.ones([rows-1, cols]) * w1
w_cols = np.ones([rows, cols-1]) * w1
solw = tv1w_2d(x, w_rows, w_cols, max_iters=5000)
solw1 = tv1_2d(x, w1, max_iters=5000)
assert np.allclose(solw, solw1, atol=1e-3)
def test_tv1w_2d_uniform_weights():
for _ in range(20):
x = _generate2d()
rows = len(x)
cols = len(x[0])
w1 = np.random.rand()
w_rows = np.ones([rows-1, cols]) * w1
w_cols = np.ones([rows, cols-1]) * w1
solw = tv1w_2d(x, w_rows, w_cols, max_iters=5000)
solw1 = tv1_2d(x, w1, max_iters=5000)
assert np.allclose(solw, solw1, atol=1e-3)
def test_tv1w_2d_uniform_weights():
for _ in range(20):
rows = np.random.randint(1e1, 3e1)
cols = np.random.randint(1e1, 3e1)
x = 100 * np.random.randn(rows, cols)
w1 = np.random.rand()
w_rows = np.ones([rows - 1, cols]) * w1
w_cols = np.ones([rows, cols - 1]) * w1
solw = tv1w_2d(x, w_rows, w_cols, max_iters=5000)
solw1 = tv1_2d(x, w1, max_iters=5000)
assert np.allclose(solw, solw1, atol=1e-3)
def test_tv1_tv1w_2d():
"""Tests that 2D-TV1w == 2D-TV1 for unit weights"""
for _ in range(20):
x = _generate2D()
rows = len(x)
cols = len(x[0])
w = 20 * np.random.rand()
w_cols = w * np.ones((rows - 1, cols))
w_rows = w * np.ones((rows, cols - 1))
solution1 = tv1_2d(x, w, max_iters=5000)
solutionp = tv1w_2d(x, w_cols, w_rows, max_iters=5000)
assert np.allclose(solution1, solutionp, atol=1e-3)
def test_tv1w_2d_emengd():
r"""Issue reported by emengd
Make the solver fail due to missing checks on integer arguments
"""
a = -np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) / 10.
sol1 = tv1w_2d(a,
np.array([[1, 1, 1], [1, 1, 1]]),
np.array([[1, 1], [1, 1], [1, 1]]),
max_iters=100)
sol2 = tv1_2d(a, 1)
assert np.allclose(sol1, sol2, atol=1e-3)
def test_tv1_tv1w_2d():
"""Tests that 2D-TV1w == 2D-TV1 for unit weights"""
for _ in range(20):
x = _generate2d()
rows = len(x)
cols = len(x[0])
w = 20*np.random.rand()
w_cols = w * np.ones((rows-1, cols))
w_rows = w * np.ones((rows, cols-1))
solution1 = tv1_2d(x, w, max_iters=5000)
solutionp = tv1w_2d(x, w_cols, w_rows, max_iters=5000)
assert np.allclose(solution1, solutionp, atol=1e-3)
def test_tv1_2d():
methods = ('yang', 'condat', 'chambolle-pock')
for _ in range(20):
rows = np.random.randint(1e1, 3e1)
cols = np.random.randint(1e1, 3e1)
x = 100*np.random.randn(rows, cols)
w = 20*np.random.rand()
solutions = [tv1_2d(x, w, method=method, max_iters=5000)
for method in methods]
solutions.append([tvp_2d(x, w, w, 1, 1, max_iters=5000)])
w_cols = w * np.ones((rows-1, cols))
w_rows = w * np.ones((rows, cols-1))
solutions.append(tv1w_2d(x, w_cols, w_rows, max_iters=5000))
for i in range(1, len(solutions)):
assert np.allclose(solutions[i], solutions[0], atol=1e-3)
def test_tv1_2d():
methods = ('yang', 'condat', 'chambolle-pock')
for _ in range(20):
rows = np.random.randint(1e1, 3e1)
cols = np.random.randint(1e1, 3e1)
x = 100 * np.random.randn(rows, cols)
w = 20 * np.random.rand()
solutions = [
tv1_2d(x, w, method=method, max_iters=5000) for method in methods
]
solutions.append([tvp_2d(x, w, w, 1, 1, max_iters=5000)])
w_cols = w * np.ones((rows - 1, cols))
w_rows = w * np.ones((rows, cols - 1))
solutions.append(tv1w_2d(x, w_cols, w_rows, max_iters=5000))
for i in range(1, len(solutions)):
assert np.allclose(solutions[i], solutions[0], atol=1e-3)
-1 * (np.sum(grady_l**2, axis=2) + np.sum(grady_b**2, axis=2) +
np.sum(grady_a**2, axis=2)) / 2.0)
weight_x = 10**exp_factor * np.exp(
-1 * (np.sum(gradx_l**2, axis=2) + np.sum(gradx_b**2, axis=2) +
np.sum(gradx_a**2, axis=2)) / 2.0)
#weight_y = 10 ** -4 * ( 1 - (np.sum(grady_l**2, axis=2) + np.sum(grady_b**2, axis=2) + np.sum(grady_a**2, axis=2) ) ) **2
#weight_x = 10 ** -4 * ( 1 - (np.sum(gradx_l**2, axis=2) + np.sum(gradx_b**2, axis=2) + np.sum(gradx_a**2, axis=2) ) ) **2
#print np.mean(weight_y.ravel())
#print np.mean(weight_x.ravel())
print('TV optimization: Iterration ' + str(iterations_ - iterations + 1))
#F = ptv.tvgen(X, [weight_y, weight_x], [1,2], np.array([1,1]));
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
F[:, :, 0] = ptv.tv1w_2d(im[:, :, 0].reshape(X_gray.shape),
weight_y[1:, :], weight_x[:, 1:])
F[:, :, 1] = ptv.tv1w_2d(im[:, :, 1].reshape(X_gray.shape),
weight_y[1:, :], weight_x[:, 1:])
F[:, :, 2] = ptv.tv1w_2d(im[:, :, 2].reshape(X_gray.shape),
weight_y[1:, :], weight_x[:, 1:])
im = F
im_lab = color.rgb2lab(F)
iterations -= 1
#F = (F- np.min(F.ravel()) ) / ( np.max(F.ravel()) - np.min(F.ravel()))
print np.min(F.ravel())
print np.max(F.ravel())
io.imsave('result_' + str(iterations_) + '.png', F)
#io.imsave('result_'+str(iterations_)+'.png', F.clip(0.0, 1.0));
# Load image
X = io.imread('colors.png')
X = ski.img_as_float(X)
X = color.rgb2gray(X)
# Introduce noise
noiseLevel = 0.01
N = util.random_noise(X, mode='speckle', var=noiseLevel)
# Gradient in columns
W1 = 0.01 * np.cumsum(np.ones((X.shape[0]-1, X.shape[1])), 1)
W2 = 0.01 * np.ones((X.shape[0], X.shape[1]-1))
print('Solving 2D weighted TV...' )
start = time.time()
FW = ptv.tv1w_2d(N, W1, W2)
end = time.time()
print('Elapsed time ' + str(end-start))
plt.subplot(3, 4, 1); io.imshow(W1); plt.title('Weights along columns');
plt.subplot(3, 4, 5); io.imshow(W2); plt.title('Weights along rows');
plt.subplot(3, 4, 9); io.imshow(FW); plt.title('Filter result');
# Gradient in rows
W1 = 0.01 * np.ones((X.shape[0]-1, X.shape[1]))
W2 = 0.01 * np.cumsum(np.ones((X.shape[0], X.shape[1]-1)), 0)
print('Solving 2D weighted TV...')
start = time.time()
FW = ptv.tv1w_2d(N, W1, W2)
end = time.time()
print('Elapsed time ' + str(end-start))
# Load image
X = io.imread('colors.png')
X = ski.img_as_float(X)
X = color.rgb2gray(X)
# Introduce noise
noiseLevel = 0.01
N = util.random_noise(X, mode='speckle', var=noiseLevel)
# Gradient in columns
W1 = 0.01 * np.cumsum(np.ones((X.shape[0] - 1, X.shape[1])), 1)
W2 = 0.01 * np.ones((X.shape[0], X.shape[1] - 1))
print('Solving 2D weighted TV...')
start = time.time()
FW = ptv.tv1w_2d(N, W1, W2)
end = time.time()
print('Elapsed time ' + str(end - start))
plt.subplot(3, 4, 1)
io.imshow(W1)
plt.title('Weights along columns')
plt.subplot(3, 4, 5)
io.imshow(W2)
plt.title('Weights along rows')
plt.subplot(3, 4, 9)
io.imshow(FW)
plt.title('Filter result')
# Gradient in rows
W1 = 0.01 * np.ones((X.shape[0] - 1, X.shape[1]))
def forward(self, x, weights_row, weights_col):
r"""Solve the total variation problem and return the solution.
Arguments
---------
x: :class:`torch:torch.Tensor`
A tensor with shape ``(m, n)`` holding the input signal.
weights_row: :class:`torch:torch.Tensor`
The horizontal edge weights.
Tensor of shape ``(m, n - 1)``, or ``(1,)`` if all weights
are equal.
weights_col: :class:`torch:torch.Tensor`
The vertical edge weights.
Tensor of shape ``(m - 1, n)``, or ``(1,)`` if all weights
are equal.
Returns
-------
:class:`torch:torch.Tensor`
The solution to the total variation problem, of shape ``(m, n)``.
"""
self.refine = True
self.tv_args = {}
self.average_connected = True
def _linearize(y, weights_row, weights_col):
"""Compute a linearization of the graph-cut function at the given point.
Arguments
---------
y : numpy.ndarray
The point where the linearization is computed, shape ``(m, n)``.
weights_row : numpy.ndarray
The non-negative row weights, with shape ``(m, n - 1)``.
y : numpy.ndarray
The non-negative column weights, with shape ``(m - 1, n)``.
Returns
-------
numpy.ndarray
The linearization of the graph-cut function at ``y``."""
diffs_col = np.sign(y[1:, :] - y[:-1, :])
diffs_row = np.sign(y[:, 1:] - y[:, :-1])
f = np.zeros_like(y) # The linearization.
f[:, 1:] += diffs_row * weights_row
f[:, :-1] -= diffs_row * weights_row
f[1:, :] += diffs_col * weights_col
f[:-1, :] -= diffs_col * weights_col
return f
def _refine(opt, x, weights_row, weights_col):
"""Refine the solution by solving an isotonic regression.
The weights can either be two-dimensional tensors, or of shape (1,)."""
idx = np.argsort(opt.ravel()) # Will pick an arbitrary order cone.
ordered_vec = np.zeros_like(idx, dtype=np.float)
ordered_vec[idx] = np.arange(np.size(opt))
f = _linearize(ordered_vec.reshape(opt.shape),
weights_row.cpu().detach().numpy(),
weights_col.cpu().detach().numpy())
opt_idx = isotonic(
(x.view(-1).cpu().detach().numpy() - f.ravel())[idx])
opt = np.zeros_like(opt_idx)
opt[idx] = opt_idx
return opt
opt = tv1w_2d(x.cpu().detach().numpy(),
weights_col.cpu().detach().numpy(),
weights_row.cpu().detach().numpy(), **self.tv_args)
if self.refine:
# opt = self._refine(opt, x, weights_row, weights_col)
opt = _refine(opt, x, weights_row, weights_col)
opt = torch.Tensor(opt).view_as(x)
self.save_for_backward(opt)
return opt.to(device)
def flatten_color(im, output_folder=None, iterations=4, exp_factor=10):
import time
import numpy as np
import prox_tv as ptv
from skimage import color
import json
im = np.asfarray(im)
row, col, ch = im.shape
shape_2d = (row, col)
im_lab = np.asfarray(color.rgb2lab(im))
# Hyperparameters
#iterations = 1 # default 2 or 4
#exp_factor = 11 # Degree of flattening (changes according to image type)
h = 1 # neighbourhood (best 1)
start = time.time()
iterations_ = iterations
# Mem alocate
grady_l = np.zeros((row, col, h))
grady_a = np.zeros((row, col, h))
grady_b = np.zeros((row, col, h))
gradx_l = np.zeros((row, col, h))
gradx_a = np.zeros((row, col, h))
gradx_b = np.zeros((row, col, h))
F = np.zeros((row, col, ch))
while iterations > 0:
#for i in range(h):
i = 0 # only immediate neighbourhood
grady_l[:, :, i] = im_lab[:, :, 0] - np.roll(
im_lab[:, :, 0].reshape(shape_2d), i + 1, axis=0)
grady_a[:, :, i] = im_lab[:, :, 1] - np.roll(
im_lab[:, :, 1].reshape(shape_2d), i + 1, axis=0)
grady_b[:, :, i] = im_lab[:, :, 2] - np.roll(
im_lab[:, :, 2].reshape(shape_2d), i + 1, axis=0)
gradx_l[:, :, i] = im_lab[:, :, 0] - np.roll(
im_lab[:, :, 0].reshape(shape_2d), i + 1, axis=1)
gradx_a[:, :, i] = im_lab[:, :, 1] - np.roll(
im_lab[:, :, 1].reshape(shape_2d), i + 1, axis=1)
gradx_b[:, :, i] = im_lab[:, :, 2] - np.roll(
im_lab[:, :, 2].reshape(shape_2d), i + 1, axis=1)
weight_y = 10**exp_factor * np.exp(
-1 * (np.sum(grady_l**2, axis=2) + np.sum(grady_b**2, axis=2) +
np.sum(grady_a**2, axis=2)) / 2.0)
weight_x = 10**exp_factor * np.exp(
-1 * (np.sum(gradx_l**2, axis=2) + np.sum(gradx_b**2, axis=2) +
np.sum(gradx_a**2, axis=2)) / 2.0)
#weight_y = 10 ** -4 * ( 1 - (np.sum(grady_l**2, axis=2) + np.sum(grady_b**2, axis=2) + np.sum(grady_a**2, axis=2) ) ) **2
#weight_x = 10 ** -4 * ( 1 - (np.sum(gradx_l**2, axis=2) + np.sum(gradx_b**2, axis=2) + np.sum(gradx_a**2, axis=2) ) ) **2
#print np.mean(weight_y.ravel())
#print np.mean(weight_x.ravel())
print('TV-l1 flattening by proximal algo: Iterration ' +
str(iterations_ - iterations + 1))
#F = ptv.tvgen(X, [weight_y, weight_x], [1,2], np.array([1,1]));
# Image | Penalty in each dimension | Dimensions to penalize | Norms to use
F[:, :, 0] = ptv.tv1w_2d(im[:, :, 0].reshape(shape_2d),
weight_y[1:, :], weight_x[:, 1:])
F[:, :, 1] = ptv.tv1w_2d(im[:, :, 1].reshape(shape_2d),
weight_y[1:, :], weight_x[:, 1:])
F[:, :, 2] = ptv.tv1w_2d(im[:, :, 2].reshape(shape_2d),
weight_y[1:, :], weight_x[:, 1:])
# update
im = F
im_lab = color.rgb2lab(F)
iterations -= 1
print np.min(F.ravel())
print np.max(F.ravel())
# Notify if exp_factor too much
if np.max(F.ravel()) >= 1.2:
with open(output_folder + 'error_log.json', 'w') as outfile:
json.dump('Overflattening: Decrease exp_factor hyperparameter!',
outfile)
assert np.max(F.ravel(
)) < 1.2, 'Overflattening: Decrease exp_factor hyperparameter!'
#io.imsave('result_'+str(iterations_)+'.png', F);
#io.imsave('result_'+str(iterations_)+'.png', F.clip(0.0, 1.0));
end = time.time()
print('Elapsed time ' + str(end - start))
return F
