def fit(self,
chips,
specs=None,
n_samples=0,
counts=None,
a=0,
bounds=None):
"""Primary execution point. Run either standard KDE or class-membership
based KDE. If any of the class-membership based KDE arguments are set,
it will be run instead of standard KDE.
Parameters
----------
chips : list
A list of chip model objects.
n_samples : int
The number of samples to generate.
specs : qikify.models.Specs, optional
If using partitioned sampling, boundaries defining pass / critical /
fail subspaces must be provided.
counts : dict, optional
If using partitioned sampling, counts dictionary must be provided,
with three keys: nGood, nCritical, nFail.
"""
X = pandas.DataFrame([chip.LCT for chip in chips])
self.n, self.d = X.shape
self.specs = specs
self.columns = getattr(X, 'columns', None)
# Normalize data/bounds
self.scale_factors, self.Xn = standardize(X)
self.bounds = standardize(np.array([X.min(0), X.max(0)]), \
self.scale_factors)
if bounds is not None:
self.bounds = standardize(bounds, self.scale_factors)
# Select bandwidth for Epanechnikov kernel (Rule of Thumb, see
# Silverman, p.86)
self.bandwidth = 0.8 # Magic number, default bandwidth scaling factor
self.c_d = 2.0 * pow( np.pi, (self.d/2.0) ) / \
( self.d * gamma(self.d/2) )
self.h = self._compute_h(self.n, self.d, self.c_d, self.bandwidth)
self._set_bandwith_factors(a)
# Generate samples
if counts is None:
return self._gen_samples(n_samples)
else:
print 'KDE: Running on dataset of size n: %d d: %d and \
generating %d samples.' % (self.n, self.d, sum(
counts.values()))
self._gen_spec_limits(X)
return self._gen_partitioned_samples(counts)
def fit(self,
chips,
specs = None,
n_samples = 0,
counts = None,
a = 0,
bounds = None):
"""Primary execution point. Run either standard KDE or class-membership
based KDE. If any of the class-membership based KDE arguments are set,
it will be run instead of standard KDE.
Parameters
----------
chips : list
A list of chip model objects.
n_samples : int
The number of samples to generate.
specs : qikify.models.Specs, optional
If using partitioned sampling, boundaries defining pass / critical /
fail subspaces must be provided.
counts : dict, optional
If using partitioned sampling, counts dictionary must be provided,
with three keys: nGood, nCritical, nFail.
"""
X = pandas.DataFrame([chip.LCT for chip in chips])
self.n, self.d = X.shape
self.specs = specs
self.columns = getattr(X, 'columns', None)
# Normalize data/bounds
self.scale_factors, self.Xn = standardize(X)
self.bounds = standardize(np.array([X.min(0), X.max(0)]), \
self.scale_factors)
if bounds is not None:
self.bounds = standardize(bounds, self.scale_factors)
# Select bandwidth for Epanechnikov kernel (Rule of Thumb, see
# Silverman, p.86)
self.bandwidth = 0.8 # Magic number, default bandwidth scaling factor
self.c_d = 2.0 * pow( np.pi, (self.d/2.0) ) / \
( self.d * gamma(self.d/2) )
self.h = self._compute_h(self.n, self.d, self.c_d, self.bandwidth)
self._set_bandwith_factors(a)
# Generate samples
if counts is None:
return self._gen_samples(n_samples)
else:
print 'KDE: Running on dataset of size n: %d d: %d and \
generating %d samples.' % (self.n, self.d, sum(counts.values()))
self._gen_spec_limits(X)
return self._gen_partitioned_samples(counts)
def fit(self, chips):
"""Run Laplacian Score Feature Selection.
.. note:: Eventually, it'd be nice to maintain col names with Xin so
that we can add a plot method to plot scores vs. column names.
Notes
-----
This code is based on the definition from the paper [1]_:
.. \\frac{\sum_{ij} (f_r^i - f_r^j) * S_{ij}}{sigma_2}
.. [1] He, X. and Cai, D. and Niyogi, P., "Laplacian Score for Feature
Selection", NIPS 2005.
Parameters
----------
chips : list
A list of chip objects
"""
X = np.array([chip.LCT.values() for chip in chips])
gnd = np.array([chip.gnd for chip in chips])
assert X.shape[0] == len(gnd), \
"Data and gnd do not have matching sizes"
_, X = standardize(X)
# Per LSFS paper, S_ij = exp(-||x_i - x_j||^2 / t). I've found that
# t = ncol(X) to be a suitable choice; anything on that order should
# work just fine.
S = self._construct_w(X, gnd, t=X.shape[1])
D = sum(S, 1)
dot_d_x = np.dot(D, X)
z = (dot_d_x * dot_d_x) / sum(D)
dprime = sum(np.dot(X.T, np.diag(D)).T * X, 0) - z
lprime = sum(np.dot(X.T, S).T * X, 1) - z
# Remove trivial solutions
dprime[dprime < 1e-12] = np.inf
# Compute and retain Laplacian scores and rankings
self.scores = (lprime/dprime).T
self.ranking = np.argsort(-self.scores)
del S # Clean up to save memory
return self
def fit(self, chips):
"""Run Laplacian Score Feature Selection.
.. note:: Eventually, it'd be nice to maintain col names with Xin so
that we can add a plot method to plot scores vs. column names.
Notes
-----
This code is based on the definition from the paper [1]_:
.. \\frac{\sum_{ij} (f_r^i - f_r^j) * S_{ij}}{sigma_2}
.. [1] He, X. and Cai, D. and Niyogi, P., "Laplacian Score for Feature
Selection", NIPS 2005.
Parameters
----------
chips : list
A list of chip objects
"""
X = np.array([chip.LCT.values() for chip in chips])
gnd = np.array([chip.gnd for chip in chips])
assert X.shape[0] == len(gnd), \
"Data and gnd do not have matching sizes"
_, X = standardize(X)
# Per LSFS paper, S_ij = exp(-||x_i - x_j||^2 / t). I've found that
# t = ncol(X) to be a suitable choice; anything on that order should
# work just fine.
S = self._construct_w(X, gnd, t=X.shape[1])
D = sum(S, 1)
dot_d_x = np.dot(D, X)
z = (dot_d_x * dot_d_x) / sum(D)
dprime = sum(np.dot(X.T, np.diag(D)).T * X, 0) - z
lprime = sum(np.dot(X.T, S).T * X, 1) - z
# Remove trivial solutions
dprime[dprime < 1e-12] = np.inf
# Compute and retain Laplacian scores and rankings
self.scores = (lprime / dprime).T
self.ranking = np.argsort(-self.scores)
del S # Clean up to save memory
return self
def _gen_partitioned_samples(self, counts):
"""Generates nCritical critical devices, nGood good devices, nFail
failing devices, with each region defined by specs.inner /
specs.outer.
"""
# Initialize arrays for speed
Sg, Sc, Sf = np.zeros((counts['nGood'], self.d)), \
np.zeros((counts['nCritical'], self.d)), \
np.zeros((counts['nFail'], self.d))
ng, nc, nf = 0, 0, 0
thresh = 0.02
while (ng + nc + nf < sum(counts.values())):
sample = standardize(self._gen_sample(), \
self.scale_factors, reverse = True)
if self._is_good(sample) and ng < counts['nGood']:
Sg[ng, :] = sample
ng += 1
if self._is_failing(sample) and nf < counts['nFail']:
Sf[nf, :] = sample
nf += 1
if self._is_critical(sample) and nc < counts['nCritical']:
Sc[nc, :] = sample
nc += 1
# Prints # generated in each category so we can monitor progress,
# since this can take a while :)
if float(ng + nc + nf) / sum(counts.values()) > thresh:
print 'Ng:%i/%i Nc:%i/%i Nf:%i/%i' % \
(ng, counts['nGood'], \
nc, counts['nCritical'], \
nf, counts['nFail'])
thresh += 0.02
print 'Non-parametric density estimation sampling complete.'
return pandas.DataFrame(np.vstack((Sc, Sg, Sf)), columns=self.columns)
def _gen_partitioned_samples(self, counts):
"""Generates nCritical critical devices, nGood good devices, nFail
failing devices, with each region defined by specs.inner /
specs.outer.
"""
# Initialize arrays for speed
Sg, Sc, Sf = np.zeros((counts['nGood'], self.d)), \
np.zeros((counts['nCritical'], self.d)), \
np.zeros((counts['nFail'], self.d))
ng, nc, nf = 0, 0, 0
thresh = 0.02
while ( ng+nc+nf < sum(counts.values()) ):
sample = standardize(self._gen_sample(), \
self.scale_factors, reverse = True)
if self._is_good(sample) and ng < counts['nGood']:
Sg[ng, :] = sample
ng += 1
if self._is_failing(sample) and nf < counts['nFail']:
Sf[nf, :] = sample
nf += 1
if self._is_critical(sample) and nc < counts['nCritical']:
Sc[nc, :] = sample
nc += 1
# Prints # generated in each category so we can monitor progress,
# since this can take a while :)
if float(ng+nc+nf) / sum(counts.values()) > thresh:
print 'Ng:%i/%i Nc:%i/%i Nf:%i/%i' % \
(ng, counts['nGood'], \
nc, counts['nCritical'], \
nf, counts['nFail'])
thresh += 0.02
print 'Non-parametric density estimation sampling complete.'
return pandas.DataFrame(np.vstack((Sc, Sg, Sf)), columns=self.columns)
def _gen_samples(self, n_samples):
"""Generate KDE samples.
"""
Sn = np.vstack([self._gen_sample() for _ in xrange(n_samples)])
sample = standardize(Sn, self.scale_factors, reverse=True)
return pandas.DataFrame(sample, columns=self.columns)
def _gen_samples(self, n_samples):
"""Generate KDE samples.
"""
Sn = np.vstack([ self._gen_sample() for _ in xrange(n_samples) ])
sample = standardize(Sn, self.scale_factors, reverse = True)
return pandas.DataFrame(sample, columns=self.columns)
