def test_sfa2_gradient2(self):
def _alt_sfa2_grad(self, x):
"""Reference grad method based on quadratic forms."""
# note that the H and f arrays are cached in the node and remain even
# after the extension has been deactivated
if not hasattr(self, "__gradient_Hs"):
quad_forms = [
self.get_quadratic_form(i) for i in range(self.output_dim)
]
self.__gradient_Hs = numx.vstack(
(quad_form.H[numx.newaxis] for quad_form in quad_forms))
self.__gradient_fs = numx.vstack(
(quad_form.f[numx.newaxis] for quad_form in quad_forms))
grad = (numx.dot(x, self.__gradient_Hs) + numx.repeat(
self.__gradient_fs[numx.newaxis, :, :], len(x), axis=0))
return grad
sfa2_node = bimdp.nodes.SFA2BiNode(output_dim=3)
x = numx_rand.random((300, 6))
sfa2_node.train(x)
sfa2_node.stop_training()
x = numx_rand.random((2, 6))
mdp.activate_extension("gradient")
try:
result1 = sfa2_node.execute(x, {"method": "gradient"})
grad1 = result1[1]["grad"]
grad2 = _alt_sfa2_grad(sfa2_node, x)
assert numx.amax(abs(grad1 - grad2)) < 1E-9
finally:
mdp.deactivate_extension("gradient")
def test_sfa2_gradient2(self):
def _alt_sfa2_grad(self, x):
"""Reference grad method based on quadratic forms."""
# note that the H and f arrays are cached in the node and remain even
# after the extension has been deactivated
if not hasattr(self, "__gradient_Hs"):
quad_forms = [self.get_quadratic_form(i)
for i in range(self.output_dim)]
self.__gradient_Hs = numx.vstack((quad_form.H[numx.newaxis]
for quad_form in quad_forms))
self.__gradient_fs = numx.vstack((quad_form.f[numx.newaxis]
for quad_form in quad_forms))
grad = (numx.dot(x, self.__gradient_Hs) +
numx.repeat(self.__gradient_fs[numx.newaxis,:,:],
len(x), axis=0))
return grad
sfa2_node = bimdp.nodes.SFA2BiNode(output_dim=3)
x = numx_rand.random((300, 6))
sfa2_node.train(x)
sfa2_node.stop_training()
x = numx_rand.random((2, 6))
mdp.activate_extension("gradient")
try:
result1 = sfa2_node.execute(x, {"method": "gradient"})
grad1 = result1[1]["grad"]
grad2 = _alt_sfa2_grad(sfa2_node, x)
assert numx.amax(abs(grad1 - grad2)) < 1E-9
finally:
mdp.deactivate_extension("gradient")
def verify_ICANode(icanode, rand_func = uniform, vars=3, N=8000, prec=3):
dim = (N, vars)
mat,mix,inp = get_random_mix(rand_func=rand_func,mat_dim=dim)
icanode.train(inp)
act_mat = icanode.execute(inp)
cov = mdp.utils.cov2((mat-mean(mat,axis=0))/std(mat,axis=0), act_mat)
maxima = numx.amax(abs(cov), axis=0)
assert_array_almost_equal(maxima,numx.ones(vars), prec)
def verify_ICANode(icanode, rand_func=uniform, vars=3, N=8000, prec=3):
dim = (N, vars)
mat, mix, inp = get_random_mix(rand_func=rand_func, mat_dim=dim)
icanode.train(inp)
act_mat = icanode.execute(inp)
cov = mdp.utils.cov2((mat - mean(mat, axis=0)) / std(mat, axis=0), act_mat)
maxima = numx.amax(abs(cov), axis=0)
assert_array_almost_equal(maxima, numx.ones(vars), prec)
def _train(self, x):
"""Determine coordinatewise and absolute maxima and minima.
:param x: Chuck of data to be used for training. Observations/samples
must be along the first axis, variables along the second.
:type x: numpy.ndarray
The values are used to generate a transformation to the valid domain
for the data.
"""
if self.amin is None:
self.amaxcolumn = np.amax(x, axis=0)
self.amincolumn = np.amin(x, axis=0)
self.amax = np.amax(self.amaxcolumn)
self.amin = np.amin(self.amincolumn)
else:
self.amaxcolumn = np.maximum(self.amaxcolumn, np.amax(x, axis=0))
self.amincolumn = np.minimum(self.amincolumn, np.amin(x, axis=0))
self.amax = np.amax(self.amaxcolumn)
self.amin = np.amin(self.amincolumn)
def check_domain(self, x, prec=1e-6):
"""Checks for compliance of the data x with the domain on which
the function sequence selected is defined or orthogonal.
:param x: The data to be expanded. Observations/samples must
be along the first axis, variables along the second.
:type x: numpy.ndarray
:param prec: (Numerical) tolerance when checking validity.
:type prec: float
:raise mdp.NodeException: If one or more values lie outside of the function
specific domain.
"""
xmax = np.amax(x)-prec
xmin = np.amin(x)+prec
if (self.upper < xmax) or (self.lower > xmin):
raise mdp.NodeException(
"One or more values lie outside of the function specific domain.")
def test_gradient_product(self):
"""Test that the product of gradients is calculated correctly."""
sfa_node1 = bimdp.nodes.SFABiNode(output_dim=5)
sfa_node2 = bimdp.nodes.SFABiNode(output_dim=3)
flow = sfa_node1 + sfa_node2
x = numx_rand.random((300, 10))
flow.train(x)
mdp.activate_extension("gradient")
try:
x1 = numx_rand.random((2, 10))
x2, msg = sfa_node1.execute(x1, {"method": "gradient"})
grad1 = msg["grad"]
_, msg = sfa_node2.execute(x2, {"method": "gradient"})
grad2 = msg["grad"]
grad12 = flow.execute(x1, {"method": "gradient"})[1]["grad"]
# use a different way to calculate the product of the gradients,
# this method is too memory intensive for large data
ref_grad = numx.sum(grad2[:,:,numx.newaxis,:] *
numx.transpose(grad1[:,numx.newaxis,:,:], (0,1,3,2)),
axis=3)
assert numx.amax(abs(ref_grad - grad12)) < 1E-9
finally:
mdp.deactivate_extension("gradient")
def test_gradient_product(self):
"""Test that the product of gradients is calculated correctly."""
sfa_node1 = bimdp.nodes.SFABiNode(output_dim=5)
sfa_node2 = bimdp.nodes.SFABiNode(output_dim=3)
flow = sfa_node1 + sfa_node2
x = numx_rand.random((300, 10))
flow.train(x)
mdp.activate_extension("gradient")
try:
x1 = numx_rand.random((2, 10))
x2, msg = sfa_node1.execute(x1, {"method": "gradient"})
grad1 = msg["grad"]
_, msg = sfa_node2.execute(x2, {"method": "gradient"})
grad2 = msg["grad"]
grad12 = flow.execute(x1, {"method": "gradient"})[1]["grad"]
# use a different way to calculate the product of the gradients,
# this method is too memory intensive for large data
ref_grad = numx.sum(grad2[:, :, numx.newaxis, :] *
numx.transpose(grad1[:, numx.newaxis, :, :],
(0, 1, 3, 2)),
axis=3)
assert numx.amax(abs(ref_grad - grad12)) < 1E-9
finally:
mdp.deactivate_extension("gradient")
