def gen_InvertibleLinearTransform(shape):
assert len(shape) == 1
dim = shape[0]
invertible = False
while not invertible:
mat = randn(dim, dim)
invertible = (logDet(mat) > float('-inf'))
return xf.LinearTransform(mat).withTag(randTag())
def gen_InvertibleLinearTransform(shape):
assert len(shape) == 1
dim = shape[0]
invertible = False
while not invertible:
mat = randn(dim, dim)
invertible = (logDet(mat) > float('-inf'))
return xf.LinearTransform(mat).withTag(randTag())
def test_logDet(self, numPoints=200):
for pointIndex in range(numPoints):
n = randint(0, 20)
if randint(0, 10) == 0:
A = np.zeros([n, n])
else:
A = randn(n, n)
trueDet = la.det(A) if n > 0 else 1.0
assert_allclose(np.exp(mathhelp.logDet(A)), abs(trueDet))
def computeLogJac(transform, x):
shapeOut = np.shape(x)
deriv = transform.deriv(x)
if len(shapeOut) == 0:
return math.log(abs(deriv))
elif len(shapeOut) == 1:
return logDet(deriv)
else:
raise RuntimeError('log-Jacobian computation not implemented for output of rank >= 2')
def test_logDet(self, numPoints = 200):
for pointIndex in range(numPoints):
n = randint(0, 20)
if randint(0, 10) == 0:
A = np.zeros([n, n])
else:
A = randn(n, n)
trueDet = la.det(A) if n > 0 else 1.0
assert_allclose(np.exp(mathhelp.logDet(A)), abs(trueDet))
def computeLogJac(transform, x):
shapeOut = np.shape(x)
deriv = transform.deriv(x)
if len(shapeOut) == 0:
return math.log(abs(deriv))
elif len(shapeOut) == 1:
return logDet(deriv)
else:
raise RuntimeError(
'log-Jacobian computation not implemented for output of rank >= 2')
def logDetMat(self):
return logDet(self.mat)
