def test_tv2_prox():
"""
similar test, but for 2D total variation penalty.
"""
np.random.seed(0)
n_rows, n_cols = 6, 8
n_features = n_rows * n_cols
gamma = np.random.rand()
epsilon = 0.1 # account for some numerical errors
def tv_norm(x, n_rows, n_cols):
X = x.reshape((n_rows, n_cols))
return np.sum(np.abs(np.diff(X, 0))) + np.sum(np.abs(np.diff(X, 1)))
for nrun in range(20):
x = np.random.randn(n_features)
x_next = tv_prox.prox_tv2d(x,
gamma,
n_rows,
n_cols,
tol=1e-10,
max_iter=10000)
diff_obj = tv_norm(x, n_rows, n_cols) - tv_norm(x_next, n_rows, n_cols)
testing.assert_array_less(((x - x_next)**2).sum() / gamma,
(1 + epsilon) * diff_obj)
def prox(self, x, step_size):
# here to avoid circular imports
from copt import tv_prox
return tv_prox.prox_tv2d(
x,
step_size * self.alpha,
self.n_rows,
self.n_cols,
max_iter=self.max_iter,
tol=self.tol,
)