def test_norm2(self):
shape = [3, 4]
x_ = randn(shape)
x = tf.constant(x_)
with self.test_session():
self.assertAllClose(
np.linalg.norm(x_.ravel())**2,
util.norm2(x).eval())
def add(self, next_f):
"""Inform the solver of a new function value."""
next_f = _n.reshape(_n.array(next_f, copy=True), (self.n, 1))
# Update the Jacobian, if applicable
if not self.first_step:
self.__update_jacobian(self.next_x, next_f)
# Check if improvement is found
if (_util.norm2(next_f) < self.skip_treshold*_util.norm2(self.F)
or self.first_step):
self.x = self.next_x
self.F = next_f
self.last_reset = False
F = next_f
else:
new_H = _n.identity(self.n) / self.initial_scale
if self.last_reset:
warnings.warn("%% Broyden reset II", BroydenWarning)
# also last step failed...
self.next_x = self.next_x - next_f
self.F = next_f
else:
warnings.warn("%% Broyden reset", BroydenWarning)
self.next_x = self.x - self.F
self.last_reset = True
self.H = new_H
self.iteration += 1
return
if self.first_step:
self.first_norm = self.residual_norm()
self.first_step = False
step = _n.dot(self.H, F)
factor = min(1., self.max_step*_util.norm2(self.x)/_util.norm2(step))
step *= factor
self.next_x = self.x - step
self.iteration += 1
def norm2_matrix(mat):
for k in mat:
mat[k] = norm2(mat[k])
return mat.to_sparse(0)
def normeuclid(vec1, vec2):
n1 = norm2(vec1)
n2 = norm2(vec2)
return euclid(n1, n2)
def cosine(vec1, vec2):
return norm2(vec1).dot(norm2(vec2))
def normeuclid(vec1, vec2):
n1 = norm2(vec1)
n2 = norm2(vec2)
return euclid(n1, n2)
def cosine(vec1, vec2):
return norm2(vec1).dot(norm2(vec2))
def norm2_matrix(mat):
for k in mat:
mat[k] = norm2(mat[k])
return mat.to_sparse(0)
