def assert_array_almost_equal_diff(x,y,digits,err_msg=''):
x,y = numx.asarray(x), numx.asarray(y)
msg = '\nArrays are not almost equal'
assert 0 in [len(numx.shape(x)),len(numx.shape(y))] \
or (len(numx.shape(x))==len(numx.shape(y)) and \
numx.alltrue(numx.equal(numx.shape(x),numx.shape(y)))),\
msg + ' (shapes %s, %s mismatch):\n\t' \
% (numx.shape(x),numx.shape(y)) + err_msg
maxdiff = max(numx.ravel(abs(x-y)))/\
max(max(abs(numx.ravel(x))),max(abs(numx.ravel(y))))
if numx.iscomplexobj(x) or numx.iscomplexobj(y): maxdiff = maxdiff/2
cond = maxdiff< 10**(-digits)
msg = msg+'\n\t Relative maximum difference: %e'%(maxdiff)+'\n\t'+\
'Array1: '+str(x)+'\n\t'+\
'Array2: '+str(y)+'\n\t'+\
'Absolute Difference: '+str(abs(y-x))
assert cond, msg
def assert_array_almost_equal_diff(x, y, digits, err_msg=''):
x, y = numx.asarray(x), numx.asarray(y)
msg = '\nArrays are not almost equal'
assert 0 in [len(numx.shape(x)),len(numx.shape(y))] \
or (len(numx.shape(x))==len(numx.shape(y)) and \
numx.alltrue(numx.equal(numx.shape(x),numx.shape(y)))),\
msg + ' (shapes %s, %s mismatch):\n\t' \
% (numx.shape(x),numx.shape(y)) + err_msg
maxdiff = max(numx.ravel(abs(x-y)))/\
max(max(abs(numx.ravel(x))),max(abs(numx.ravel(y))))
if numx.iscomplexobj(x) or numx.iscomplexobj(y): maxdiff = maxdiff / 2
cond = maxdiff < 10**(-digits)
msg = msg+'\n\t Relative maximum difference: %e'%(maxdiff)+'\n\t'+\
'Array1: '+str(x)+'\n\t'+\
'Array2: '+str(y)+'\n\t'+\
'Absolute Difference: '+str(abs(y-x))
assert cond, msg
def _train(self, x, labels):
"""Add the sampel points to the classes.
labels -- Can be a list, tuple or array of labels (one for each data
point) or a single label, in which case all input data is assigned
to the same class (computationally this is more efficient).
"""
if isinstance(labels, (list, tuple, numx.ndarray)):
labels = numx.asarray(labels)
for label in set(labels):
x_label = numx.compress(labels == label, x, axis=0)
self._add_samples(x_label, label)
else:
self._add_samples(x, labels)
def _train(self, x, labels):
"""Add the sampel points to the classes.
labels -- Can be a list, tuple or array of labels (one for each data
point) or a single label, in which case all input data is assigned
to the same class (computationally this is more efficient).
"""
if isinstance(labels, (list, tuple, numx.ndarray)):
labels = numx.asarray(labels)
for label in set(labels):
x_label = numx.compress(labels == label, x, axis=0)
self._add_samples(x_label, label)
else:
self._add_samples(x, labels)
def _train(self, x, labels):
"""Update the mean information for the different classes.
:param x: The data.
:type x: numpy.ndarray
:param labels: Can be a list, tuple or array of labels (one for each data
point) or a single label, in which case all input data is assigned
to the same class (computationally this is more efficient).
"""
if isinstance(labels, (list, tuple, numx.ndarray)):
labels = numx.asarray(labels)
for label in set(labels):
x_label = numx.compress(labels == label, x, axis=0)
self._update_mean(x_label, label)
else:
self._update_mean(x, labels)
def _train(self, x, labels):
"""
:param x: Data
:type x: numpy.ndarray
:param labels: Can be a list, tuple or array of labels (one for each data point)
or a single label, in which case all input data is assigned to
the same class.
"""
# if labels is a number, all x's belong to the same class
if isinstance(labels, (list, tuple, numx.ndarray)):
labels_ = numx.asarray(labels)
# get all classes from cl
for lbl in set(labels_):
x_lbl = numx.compress(labels_ == lbl, x, axis=0)
self._update_covs(x_lbl, lbl)
else:
self._update_covs(x, labels)
def _train(self, x, labels):
"""
:Arguments:
x
data
labels
Can be a list, tuple or array of labels (one for each data point)
or a single label, in which case all input data is assigned to
the same class.
"""
# if labels is a number, all x's belong to the same class
if isinstance(labels, (list, tuple, numx.ndarray)):
labels_ = numx.asarray(labels)
# get all classes from cl
for lbl in set(labels_):
x_lbl = numx.compress(labels_ == lbl, x, axis=0)
self._update_covs(x_lbl, lbl)
else:
self._update_covs(x, labels)
def test_switchboard_gradient2(self):
"""Test gradient for a larger switchboard."""
dim = 100
connections = [int(i) for i in numx.random.random((dim,)) * (dim-1)]
sboard = mdp.hinet.Switchboard(input_dim=dim, connections=connections)
x = numx.random.random((10, dim))
# assume a 5-dimensional gradient at this stage
grad = numx.random.random((10, dim, 5))
# original reference implementation
def _switchboard_grad(self, x):
grad = numx.zeros((self.output_dim, self.input_dim))
grad[range(self.output_dim), self.connections] = 1
return numx.tile(grad, (len(x), 1, 1))
with mdp.extension("gradient"):
result = sboard._gradient(x, grad)
ext_grad = result[1]["grad"]
tmp_grad = _switchboard_grad(sboard, x)
ref_grad = numx.asarray([numx.dot(tmp_grad[i], grad[i])
for i in range(len(tmp_grad))])
assert numx.all(ext_grad == ref_grad)
def test_switchboard_gradient2(self):
"""Test gradient for a larger switchboard."""
dim = 100
connections = [int(i) for i in numx.random.random((dim, )) * (dim - 1)]
sboard = mdp.hinet.Switchboard(input_dim=dim, connections=connections)
x = numx.random.random((10, dim))
# assume a 5-dimensional gradient at this stage
grad = numx.random.random((10, dim, 5))
# original reference implementation
def _switchboard_grad(self, x):
grad = numx.zeros((self.output_dim, self.input_dim))
grad[range(self.output_dim), self.connections] = 1
return numx.tile(grad, (len(x), 1, 1))
with mdp.extension("gradient"):
result = sboard._gradient(x, grad)
ext_grad = result[1]["grad"]
tmp_grad = _switchboard_grad(sboard, x)
ref_grad = numx.asarray(
[numx.dot(tmp_grad[i], grad[i]) for i in range(len(tmp_grad))])
assert numx.all(ext_grad == ref_grad)
def __init__(self,
in_channels_xy,
field_channels_xy,
in_channel_dim=1,
out_channels=1,
field_dstr='uniform'):
"""Calculate the connections.
Keyword arguments:
in_channels_xy -- 2-Tuple with number of input channels in the x- and
y-direction (or a single number for both). This has to be
specified, since the actual input is only one 1d array.
field_channels_xy -- 2-Tuple with number of channels in each field in
the x- and y-direction (or a single number for both).
field_dstr -- Distribution for random sampling of fields. (default: uniform).
Currently supports only uniform random distribution.
in_channel_dim -- Number of connections per input channel.
out_channels -- Number of output channels (default: 1)
"""
in_channels_xy = to_2tuple(in_channels_xy)
field_channels_xy = to_2tuple(field_channels_xy)
self.field_dstr = field_dstr
self.in_channels_xy = in_channels_xy
self.field_channels_xy = field_channels_xy
self.out_channels = out_channels
out_channel_dim = (in_channel_dim * field_channels_xy[0] *
field_channels_xy[1])
# check parameters for inconsistencies
for i, name in enumerate(["x", "y"]):
if field_channels_xy[i] > in_channels_xy[i]:
err = ("Number of field channels exceeds the number of "
"input channels in %s-direction. "
"This would lead to an empty connection list." % name)
raise RandomChannelSwitchboardException(err)
in_trans = CoordinateTranslator(*in_channels_xy)
# setup a look up table for accessing connected grid points.
# This incurs a one-time expensive computation when the node is created but
# keeps the execute call efficient computationally.
self._lut_conn = numx.zeros([
self.in_channels_xy[1] - self.field_channels_xy[1] + 1,
self.in_channels_xy[0] - self.field_channels_xy[0] + 1,
out_channel_dim
],
dtype=numx.int32)
x_fields_origin = range(self.in_channels_xy[0] -
self.field_channels_xy[0] + 1)
y_fields_origin = range(self.in_channels_xy[1] -
self.field_channels_xy[1] + 1)
tot_conns = self._lut_conn.shape[0] * self._lut_conn.shape[1]
origins = numx.asarray(numx.meshgrid(x_fields_origin,
y_fields_origin)).reshape(
2, tot_conns).T
for i in xrange(tot_conns):
# inner loop over field
x_start_chan, y_start_chan = origins[i]
connections = numx.zeros([out_channel_dim], dtype=numx.int32)
first_out_con = 0
for y_in_chan in range(y_start_chan,
y_start_chan + field_channels_xy[1]):
for x_in_chan in range(x_start_chan,
x_start_chan + field_channels_xy[0]):
first_in_con = (
in_trans.image_to_index(x_in_chan, y_in_chan) *
in_channel_dim)
connections[first_out_con:first_out_con +
in_channel_dim] = range(
first_in_con,
first_in_con + in_channel_dim)
first_out_con += in_channel_dim
self._lut_conn[y_start_chan, x_start_chan] = connections
connections = self._new_connections()
super(RandomChannelSwitchboard, self).__init__(
input_dim=(in_channel_dim * in_channels_xy[0] * in_channels_xy[1]),
connections=connections,
out_channel_dim=out_channel_dim,
in_channel_dim=in_channel_dim)
def _adjust_output_dim(self):
# this function is called if we need to compute the number of
# output dimensions automatically; some quantities that are
# useful later are pre-calculated to spare precious time
if self.verbose:
print ' - adjusting output dim:'
#otherwise, we need to compute output_dim
# from desired_variance
M = self.data
k = self.k
N, d_in = M.shape
m_est_array = []
Qs = numx.zeros((N, k, k))
sig2s = numx.zeros((N, d_in))
nbrss = numx.zeros((N, k), dtype='i')
for row in range(N):
#-----------------------------------------------
# find k nearest neighbors
#-----------------------------------------------
M_Mi = M - M[row]
nbrs = numx.argsort((M_Mi**2).sum(1))[1:k + 1]
M_Mi = M_Mi[nbrs]
# compute covariance matrix of distances
Qs[row, :, :] = mult(M_Mi, M_Mi.T)
nbrss[row, :] = nbrs
#-----------------------------------------------
# singular values of M_Mi give the variance:
# use this to compute intrinsic dimensionality
# at this point
#-----------------------------------------------
sig2 = (svd(M_Mi, compute_uv=0))**2
sig2s[row, :sig2.shape[0]] = sig2
#-----------------------------------------------
# use sig2 to compute intrinsic dimensionality of the
# data at this neighborhood. The dimensionality is the
# number of eigenvalues needed to sum to the total
# desired variance
#-----------------------------------------------
sig2 /= sig2.sum()
S = sig2.cumsum()
m_est = S.searchsorted(self.desired_variance)
if m_est > 0:
m_est += (self.desired_variance - S[m_est - 1]) / sig2[m_est]
else:
m_est = self.desired_variance / sig2[m_est]
m_est_array.append(m_est)
m_est_array = numx.asarray(m_est_array)
self.output_dim = int(numx.ceil(numx.median(m_est_array)))
if self.verbose:
msg = (' output_dim = %i'
' for variance of %.2f' %
(self.output_dim, self.desired_variance))
print msg
return Qs, sig2s, nbrss
def _adjust_output_dim(self):
# this function is called if we need to compute the number of
# output dimensions automatically; some quantities that are
# useful later are pre-calculated to spare precious time
if self.verbose:
print ' - adjusting output dim:'
#otherwise, we need to compute output_dim
# from desired_variance
M = self.data
k = self.k
N, d_in = M.shape
m_est_array = []
Qs = numx.zeros((N, k, k))
sig2s = numx.zeros((N, d_in))
nbrss = numx.zeros((N, k), dtype='i')
for row in range(N):
#-----------------------------------------------
# find k nearest neighbors
#-----------------------------------------------
M_Mi = M-M[row]
nbrs = numx.argsort((M_Mi**2).sum(1))[1:k+1]
M_Mi = M_Mi[nbrs]
# compute covariance matrix of distances
Qs[row, :, :] = mult(M_Mi, M_Mi.T)
nbrss[row, :] = nbrs
#-----------------------------------------------
# singular values of M_Mi give the variance:
# use this to compute intrinsic dimensionality
# at this point
#-----------------------------------------------
sig2 = (svd(M_Mi, compute_uv=0))**2
sig2s[row, :sig2.shape[0]] = sig2
#-----------------------------------------------
# use sig2 to compute intrinsic dimensionality of the
# data at this neighborhood. The dimensionality is the
# number of eigenvalues needed to sum to the total
# desired variance
#-----------------------------------------------
sig2 /= sig2.sum()
S = sig2.cumsum()
m_est = S.searchsorted(self.desired_variance)
if m_est > 0:
m_est += (self.desired_variance-S[m_est-1])/sig2[m_est]
else:
m_est = self.desired_variance/sig2[m_est]
m_est_array.append(m_est)
m_est_array = numx.asarray(m_est_array)
self.output_dim = int( numx.ceil( numx.median(m_est_array) ) )
if self.verbose:
msg = (' output_dim = %i'
' for variance of %.2f' % (self.output_dim,
self.desired_variance))
print msg
return Qs, sig2s, nbrss
