def __reverse_single_level(self, wp):
# local bindings
level_paths = self.__level_paths
# define wavelet packet to use
WP = pywt.WaveletPacket(
data=None, wavelet=self._wavelet,
mode=self._mode, maxlevel=self.__level)
# prepare storage
signal_shape = wp.shape[:1] + self._inshape[1:]
signal = np.zeros(signal_shape)
Ntime_points = self._intimepoints
for indexes in _get_indexes(signal_shape,
self._dim):
if __debug__:
debug('MAP_', " %s" % (indexes,), lf=False, cr=True)
for path, level_data in zip(level_paths, wp[indexes]):
WP[path] = level_data
signal[indexes] = WP.reconstruct(True)[:Ntime_points]
return signal
def _hdf_list_to_objarray(hdf, memo):
if not ('shape' in hdf.attrs):
if __debug__:
debug('HDF5', "Enountered objarray stored without shape (due to a bug "
"in post 2.1 release). Some nested structures etc might not be "
"loaded incorrectly")
# yoh: we have possibly a problematic case due to my fix earlier
# resolve to old logic: nested referencing might not work :-/
obj = _hdf_list_to_obj(hdf, memo)
# need to handle special case of arrays of objects
if np.isscalar(obj):
obj = np.array(obj, dtype=np.object)
else:
obj = asobjarray(obj)
else:
shape = tuple(hdf.attrs['shape'])
# reserve space first
if len(shape):
obj = np.empty(np.prod(shape), dtype=object)
else:
# scalar
obj = np.array(None, dtype=object)
# now load the items from the list, noting existence of this
# container
obj_items = _hdf_list_to_obj(hdf, memo, target_container=obj)
# assign to the object array
for i, v in enumerate(obj_items):
obj[i] = v
if len(shape) and shape != obj.shape:
obj = obj.reshape(shape)
return obj
def _call(self, dataset):
sensitivities = []
for ind, analyzer in enumerate(self.__analyzers):
if __debug__:
debug("SA", "Computing sensitivity for SA#%d:%s" %
(ind, analyzer))
sensitivity = analyzer(dataset)
sensitivities.append(sensitivity)
if __debug__:
debug("SA",
"Returning %d sensitivities from %s" %
(len(sensitivities), self.__class__.__name__))
sa_attr = self._sa_attr
if isinstance(sensitivities[0], AttrDataset):
smerged = None
for i, s in enumerate(sensitivities):
s.sa[sa_attr] = np.repeat(i, len(s))
if smerged is None:
smerged = s
else:
smerged.append(s)
sensitivities = smerged
else:
sensitivities = \
Dataset(sensitivities,
sa={sa_attr: np.arange(len(sensitivities))})
self.ca.sensitivities = sensitivities
return sensitivities
def _get_selected_ids(self, dataset):
"""Given a dataset actually select the features
Returns
-------
indexes of the selected features
"""
# optionally train the analyzer first
if self.__train_analyzer:
self.__sensitivity_analyzer.train(dataset)
sensitivity = self.__sensitivity_analyzer(dataset)
"""Compute the sensitivity map."""
self.ca.sensitivity = sensitivity
# Select features to preserve
selected_ids = self.__feature_selector(sensitivity)
if __debug__:
debug("FS_", "Sensitivity: %s Selected ids: %s" %
(sensitivity, selected_ids))
# XXX not sure if it really has to be sorted
selected_ids.sort()
return selected_ids
def _suppress_scipy_warnings():
# Infiltrate warnings if necessary
numpy_ver = versions['numpy']
scipy_ver = versions['scipy']
# There is way too much deprecation warnings spit out onto the
# user. Lets assume that they should be fixed by scipy 0.7.0 time
if not __debug__ or (__debug__ and 'PY' not in debug.active):
filter_lines = []
if "0.6.0" <= scipy_ver and scipy_ver < "0.7.0" \
and numpy_ver > "1.1.0":
if __debug__:
debug('EXT', "Setting up filters for numpy DeprecationWarnings "
"regarding scipy < 0.7.0")
filter_lines += [
('NumpyTest will be removed in the next release.*',
DeprecationWarning),
('PyArray_FromDims: use PyArray_SimpleNew.',
DeprecationWarning),
('PyArray_FromDimsAndDataAndDescr: use PyArray_NewFromDescr.',
DeprecationWarning),
# Trick re.match, since in warnings absent re.DOTALL in re.compile
('[\na-z \t0-9]*The original semantics of histogram is scheduled to be.*'
'[\na-z \t0-9]*', Warning) ]
if scipy_ver >= "0.15":
filter_lines += [("`scipy.weave` is deprecated, use `weave` instead!",
DeprecationWarning)]
if scipy_ver >= "0.16":
# scipy deprecated it but statsmodels still import it for now
filter_lines += [("`scipy.linalg.calc_lwork` is deprecated!",
DeprecationWarning)]
for f, w in filter_lines:
warnings.filterwarnings('ignore', f, w)
def newfunc(*arg, **kwargs):
nfailed, i = 0, 0 # define i just in case
for i in xrange(niter):
try:
ret = func(*arg, **kwargs)
if i + 1 - nfailed >= niter - nfailures:
# so we know already that we wouldn't go over
# nfailures
break
except AssertionError, e:
nfailed += 1
if __debug__:
debug("TEST", "Upon %i-th run, test %s failed with %s", (i, func.__name__, e))
if nfailed > nfailures:
if __debug__:
debug(
"TEST",
"Ran %s %i times. Got %d failures, "
"while was allowed %d "
"-- re-throwing the last failure %s",
(func.__name__, i + 1, nfailed, nfailures, e),
)
exc_info = sys.exc_info()
raise exc_info[1], None, exc_info[2]
def _train(self, samples):
"""Perform network training.
Parameters
----------
samples : array-like
Used for unsupervised training of the SOM.
Notes
-----
It is assumed that prior to calling this method the _pretrain method
was called with the same argument.
"""
# ensure that dqd was set properly
dqd = self._dqd
if dqd is None:
raise ValueError("This should not happen - was _pretrain called?")
# units weight vector deltas for batch training
# (height x width x #features)
unit_deltas = np.zeros(self._K.shape, dtype='float')
# for all iterations
for it in xrange(1, self.niter + 1):
# compute the neighborhood impact kernel for this iteration
# has to be recomputed since kernel shrinks over time
k = self._compute_influence_kernel(it, dqd)
# for all training vectors
for s in samples:
# determine closest unit (as element coordinate)
b = self._get_bmu(s)
# train all units at once by unfolding the kernel (from the
# single quadrant that is precomputed), cutting it to the
# right shape and simply multiply it to the difference of target
# and all unit weights....
infl = np.vstack((
np.hstack((
# upper left
k[b[0]:0:-1, b[1]:0:-1],
# upper right
k[b[0]:0:-1, :self.kshape[1] - b[1]])),
np.hstack((
# lower left
k[:self.kshape[0] - b[0], b[1]:0:-1],
# lower right
k[:self.kshape[0] - b[0], :self.kshape[1] - b[1]]))
))
unit_deltas += infl[:, :, np.newaxis] * (s - self._K)
# apply cumulative unit deltas
self._K += unit_deltas
if __debug__:
debug("SOM", "Iteration %d/%d done: ||unit_deltas||=%g" %
(it, self.niter, np.sqrt(np.sum(unit_deltas ** 2))))
# reset unit deltas
unit_deltas.fill(0.)
def _forward_dataset(self, dataset):
# invoke super class _forward_dataset, this calls, _forward_dataset
# and this calls _forward_data in this class
mds = super(FlattenMapper, self)._forward_dataset(dataset)
# attribute collection needs to have a new length check
mds.fa.set_length_check(mds.nfeatures)
# we need to duplicate all existing feature attribute, as each original
# feature is now spread across the new feature axis
# take all "additional" axes after the actual feature axis and count
# elements a sample -- if not axis exists this will be 1
for k in dataset.fa:
if __debug__:
debug("MAP_", "Forward-mapping fa '%s'." % k)
attr = dataset.fa[k].value
# the maximmum number of axis to flatten in the attr
if not self.__maxdims is None:
maxdim = min(len(self.__origshape), self.__maxdims)
else:
maxdim = len(self.__origshape)
multiplier = mds.nfeatures / np.prod(attr.shape[:maxdim])
if __debug__:
debug("MAP_", "Broadcasting fa '%s' %s %d times" % (k, attr.shape, multiplier))
# broadcast as many times as necessary to get 'matching dimensions'
bced = np.repeat(attr, multiplier, axis=0)
# now reshape as many dimensions as the mapper knows about
mds.fa[k] = bced.reshape((-1,) + bced.shape[maxdim:])
# if there is no inspace return immediately
if self.get_space() is None:
return mds
# otherwise create the coordinates as feature attributes
else:
mds.fa[self.get_space()] = list(np.ndindex(dataset.samples[0].shape))
return mds
def _postcall(self, ds, result):
"""Postprocessing of results.
By default, does nothing.
Parameters
----------
ds: Dataset
Original input dataset.
result: Dataset
Preliminary result dataset (as produced by ``_call()``).
Returns
-------
Dataset
"""
if not self.__postproc is None:
if __debug__:
debug("NO",
"Applying post-processing node %s", (self.__postproc,))
self.ca.raw_results = result
result = self.__postproc(result)
return result
def _train(self, samples):
"""Determine the projection matrix onto the SVD components from
a 2D samples x feature data matrix.
"""
X = np.asmatrix(samples)
X = self._demean_data(X)
# singular value decomposition
U, SV, Vh = np.linalg.svd(X, full_matrices=0)
# store the final matrix with the new basis vectors to project the
# features onto the SVD components. And store its .H right away to
# avoid computing it in forward()
self._proj = Vh.H
# also store singular values of all components
self._sv = SV
if __debug__:
debug("MAP", "SVD was done on %s and obtained %d SVs " %
(samples, len(SV)) + " (%d non-0, max=%f)" %
(len(SV.nonzero()), SV[0]))
# .norm might be somewhat expensive to compute
if "MAP_" in debug.active:
debug("MAP_", "Mixing matrix has %s shape and norm=%f" %
(self._proj.shape, np.linalg.norm(self._proj)))
def __init__(self, space=None, pass_attr=None, postproc=None, **kwargs):
"""
Parameters
----------
space : str, optional
Name of the 'processing space'. The actual meaning of this argument
heavily depends on the sub-class implementation. In general, this is
a trigger that tells the node to compute and store information about
the input data that is "interesting" in the context of the
corresponding processing in the output dataset.
pass_attr : str, list of str, optional
What attribute(s) (from sa, fa, a collections, see
:meth:`Dataset.get_attr`) to pass from original dataset
provided to __call__ (before applying postproc), or from
'ca' collection of this instance (use 'ca.' prefix)
into the resultant dataset.
postproc : Node instance, optional
Node to perform post-processing of results. This node is applied
in `__call__()` to perform a final processing step on the to be
result dataset. If None, nothing is done.
"""
ClassWithCollections.__init__(self, **kwargs)
if __debug__:
debug("NO",
"Init node '%s' (space: '%s', postproc: '%s')",
(self.__class__.__name__, space, str(postproc)))
self.set_space(space)
self.set_postproc(postproc)
if isinstance(pass_attr, basestring):
pass_attr = (pass_attr,)
self.__pass_attr = pass_attr
def _untrain(self):
if __debug__:
debug("FS_", "Untraining Iterative FS: %s" % self)
self._fmeasure.untrain()
self._pmeasure.untrain()
# ask base class to do its untrain
super(IterativeFeatureSelection, self)._untrain()
def _prepredict(self, dataset):
"""Functionality prior prediction
"""
if not ('notrain2predict' in self.__tags__):
# check if classifier was trained if that is needed
if not self.trained:
raise FailedToPredictError(
"Classifier %s wasn't yet trained, therefore can't "
"predict" % self)
nfeatures = dataset.nfeatures #data.shape[1]
# check if number of features is the same as in the data
# it was trained on
if nfeatures != self.__trainednfeatures:
raise ValueError, \
"Classifier %s was trained on data with %d features, " % \
(self, self.__trainednfeatures) + \
"thus can't predict for %d features" % nfeatures
if self.params.retrainable:
if not self.__changedData_isset:
self.__reset_changed_data()
_changedData = self._changedData
data = np.asanyarray(dataset.samples)
_changedData['testdata'] = \
self.__was_data_changed('testdata', data)
if __debug__:
debug('CLF_', "prepredict: Obtained _changedData is %s",
(_changedData,))
def _set(self, val):
if __debug__ and __mvpadebug__:
# Since this call is quite often, don't convert
# values to strings here, rely on passing them
# withing msgargs
debug("COL", "Setting %s to %s ", (self, val))
self._value = val
def __was_data_changed(self, key, entry, update=True):
"""Check if given entry was changed from what known prior.
If so -- store only the ones needed for retrainable beastie
"""
idhash_ = idhash(entry)
__idhashes = self.__idhashes
changed = __idhashes[key] != idhash_
if __debug__ and 'CHECK_RETRAIN' in debug.active:
__trained = self.__trained
changed2 = entry != __trained[key]
if isinstance(changed2, np.ndarray):
changed2 = changed2.any()
if changed != changed2 and not changed:
raise RuntimeError, \
'idhash found to be weak for %s. Though hashid %s!=%s %s, '\
'estimates %s!=%s %s' % \
(key, idhash_, __idhashes[key], changed,
entry, __trained[key], changed2)
if update:
__trained[key] = entry
if __debug__ and changed:
debug('CLF_', "Changed %s from %s to %s.%s",
(key, __idhashes[key], idhash_,
('','updated')[int(update)]))
if update:
__idhashes[key] = idhash_
return changed
def label_voxel(self, c, levels = None):
if self.__referenceLevel is None:
warning("You did not provide what level to use "
"for reference. Assigning 0th level -- '%s'"
% (self._levels[0],))
self.set_reference_level(0)
# return self.__referenceAtlas.label_voxel(c, levels)
c = self._check_range(c)
# obtain coordinates of the closest voxel
cref = self._data[ self.__referenceLevel.indexes, c[0], c[1], c[2] ]
dist = norm( (cref - c) * self.voxdim )
if __debug__:
debug('ATL__', "Closest referenced point for %r is "
"%r at distance %3.2f" % (c, cref, dist))
if (self.distance - dist) >= 1e-3: # neglect everything smaller
result = self.__referenceAtlas.label_voxel(cref, levels)
result['voxel_referenced'] = c
result['distance'] = dist
else:
result = self.__referenceAtlas.label_voxel(c, levels)
if __debug__:
debug('ATL__', "Closest referenced point is "
"further than desired distance %.2f" % self.distance)
result['voxel_referenced'] = None
result['distance'] = 0
return result
def _train(self, samples):
"""Perform network training.
Parameter
---------
samples : array-like
Used for unsupervised training of the SOM.
"""
# XXX initialize with clever default, e.g. plain of first two PCA
# components
self._K = np.random.standard_normal(tuple(self.kshape) + (samples.shape[1],))
# units weight vector deltas for batch training
# (height x width x #features)
unit_deltas = np.zeros(self._K.shape, dtype='float')
# precompute distance kernel between elements in the Kohonen layer
# that will remain constant throughout the training
# (just compute one quadrant, as the distances are symmetric)
# XXX maybe do other than squared Euclidean?
dqd = np.fromfunction(lambda x, y: (x**2 + y**2)**0.5,
self.kshape, dtype='float')
# for all iterations
for it in xrange(1, self.niter + 1):
# compute the neighborhood impact kernel for this iteration
# has to be recomputed since kernel shrinks over time
k = self._compute_influence_kernel(it, dqd)
# for all training vectors
for s in samples:
# determine closest unit (as element coordinate)
b = self._get_bmu(s)
# train all units at once by unfolding the kernel (from the
# single quadrant that is precomputed), cutting it to the
# right shape and simply multiply it to the difference of target
# and all unit weights....
infl = np.vstack((
np.hstack((
# upper left
k[b[0]:0:-1, b[1]:0:-1],
# upper right
k[b[0]:0:-1, :self.kshape[1] - b[1]])),
np.hstack((
# lower left
k[:self.kshape[0] - b[0], b[1]:0:-1],
# lower right
k[:self.kshape[0] - b[0], :self.kshape[1] - b[1]]))
))
unit_deltas += infl[:,:,np.newaxis] * (s - self._K)
# apply cumulative unit deltas
self._K += unit_deltas
if __debug__:
debug("SOM", "Iteration %d/%d done: ||unit_deltas||=%g" %
(it, self.niter, np.sqrt(np.sum(unit_deltas **2))))
# reset unit deltas
unit_deltas.fill(0.)
def _binary_data_bytecount(niml):
'''helper function that returns how many bytes a NIML binary data
element should have'''
niform = niml['ni_form']
if not 'binary' in niform:
raise ValueError('Illegal niform %s' % niform)
tps = niml['vec_typ']
onetype = types.findonetype(tps)
if onetype is None:
debug('NIML', 'Not unique type: %r', tps)
return None
# numeric, either int or float
ncols = niml['vec_num']
nrows = niml['vec_len']
tp = types.code2numpy_type(onetype)
bytes_per_elem = types.numpy_type2bytecount(tp)
if bytes_per_elem is None:
raise ValueError("Type not supported: %r" % onetype)
nb = ncols * nrows * bytes_per_elem
debug('NIML', 'Number of bytes for %s: %d x %d with %d bytes / element',
(niform, ncols, nrows, bytes_per_elem))
return nb
def _level3(self, datasets):
params = self.params # for quicker access ;)
# create a mapper per dataset
mappers = [deepcopy(params.alignment) for ds in datasets]
# key different from level-2; the common space is uniform
#temp_commonspace = commonspace
residuals = None
if self.ca['residual_errors'].enabled:
residuals = np.zeros((1, len(datasets)))
self.ca.residual_errors = Dataset(samples=residuals)
# start from original input datasets again
for i, (m, ds_new) in enumerate(zip(mappers, datasets)):
if __debug__:
debug('HPAL_', "Level 3: ds #%i" % i)
# retrain mapper on final common space
ds_new.sa[m.get_space()] = self.commonspace
m.train(ds_new)
# remove common space attribute again to save on memory
del ds_new.sa[m.get_space()]
if residuals is not None:
# obtain final projection
data_mapped = m.forward(ds_new.samples)
residuals[0, i] = np.linalg.norm(data_mapped - self.commonspace)
return mappers
def _untrain(self):
if __debug__:
debug("FS_", "Untraining combined FS: %s" % self)
for fs in self.__selectors:
fs.untrain()
# ask base class to do its untrain
super(CombinedFeatureSelection, self)._untrain()
def __init__(self, index=None, *args, **kwargs):
"""
Parameters
----------
value : arbitrary (see derived implementations)
The actual value of this attribute.
**kwargs
Passed to `Collectable`
"""
if index is None:
IndexedCollectable._instance_index += 1
index = IndexedCollectable._instance_index
else:
# TODO: there can be collision between custom provided indexes
# and the ones automagically assigned.
# Check might be due
pass
self._instance_index = index
self._isset = False
self.reset()
Collectable.__init__(self, *args, **kwargs)
if __debug__ and 'COL' in debug.active:
debug("COL", "Initialized new IndexedCollectable #%d:%s %r",
(index, self.name, self))
def _train(self, dataset):
"""Select the most important features
Parameters
----------
dataset : Dataset
used to compute sensitivity maps
"""
# optionally train the analyzer first
if self.__train_analyzer:
self.__sensitivity_analyzer.train(dataset)
sensitivity = self.__sensitivity_analyzer(dataset)
"""Compute the sensitivity map."""
self.ca.sensitivity = sensitivity
# Select features to preserve
selected_ids = self.__feature_selector(sensitivity)
if __debug__:
debug("FS_", "Sensitivity: %s Selected ids: %s" %
(sensitivity, selected_ids))
# XXX not sure if it really has to be sorted
selected_ids.sort()
# announce desired features to the underlying slice mapper
self._safe_assign_slicearg(selected_ids)
# and perform its own training
super(SensitivityBasedFeatureSelection, self)._train(dataset)
def _train(self, samples):
"""Train PrototypeMapper
"""
self._proj = np.hstack([similarity.computed(samples, self.prototypes) for similarity in self.similarities])
if __debug__:
debug("MAP", "projected data of shape %s: %s " % (self._proj.shape, self._proj))
def get_selected_indexes(self, n_cfgs):
"""A naive selection of indexes according to strategy and count
Parameters
----------
n_cfgs: int
Total number of configurations to select from
"""
strategy = self.selection_strategy
count = self.count
if strategy == 'first':
indexes = slice(0, count)
elif strategy in ['equidistant', 'random']:
if strategy == 'equidistant':
# figure out what step is needed to
# accommodate the `count` number
step = float(n_cfgs) / count
assert (step >= 1.0)
indexes = [int(round(step * i)) for i in xrange(count)]
elif strategy == 'random':
indexes = np.random.permutation(range(n_cfgs))[:count]
# doesn't matter much but lets keep them in the original
# order at least
indexes.sort()
else:
# who said that I am paranoid?
raise RuntimeError("Really should not happen")
if __debug__:
debug("SPL", "For %s selection strategy selected %s "
"partition specs from %d total",
(strategy, indexes, n_cfgs))
return indexes
def _call(self, dataset):
# OPT: local bindings
clfclf = self.clf.clf
analyzer = self.__analyzer
if analyzer is None:
analyzer = clfclf.get_sensitivity_analyzer(
**(self._slave_kwargs))
if analyzer is None:
raise ValueError, \
"Wasn't able to figure basic analyzer for clf %s" % \
`clfclf`
if __debug__:
debug("SA", "Selected analyzer %s for clf %s" % \
(analyzer, clfclf))
# bind to the instance finally
self.__analyzer = analyzer
# TODO "remove" unnecessary things below on each call...
# assign corresponding classifier
analyzer.clf = clfclf
# if clf was trained already - don't train again
if clfclf.trained:
analyzer._force_train = False
result = analyzer._call(dataset)
self.ca.clf_sensitivities = result
return result
def __init__(self, value=None, name=None, doc=None):
"""
Parameters
----------
value : arbitrary (see derived implementations)
The actual value of this attribute.
name : str
Name of the collectable under which it should be available in its
respective collection.
doc : str
Documentation about the purpose of this collectable.
"""
if doc is not None:
# to prevent newlines in the docstring
try:
doc = re.sub('[\n ]+', ' ', doc)
except TypeError:
# catch some old datasets stored in HDF5
doc = re.sub('[\n ]+', ' ', np.asscalar(doc))
self.__doc__ = doc
self.__name = name
self._value = None
if value is not None:
self._set(value)
if __debug__ and __mvpadebug__:
debug("COL", "Initialized %r", (self,))
def _forward_dataset(self, dataset):
"""Forward-map a dataset.
This is a private method that can be reimplemented in derived
classes. The default implementation forward-maps the dataset samples
and returns a new dataset that is a shallow copy of the input with
the mapped samples.
Parameters
----------
dataset : Dataset-like
"""
if __debug__:
debug('MAP_', "Forward-map %s-shaped samples in dataset with '%s'."
% (dataset.samples.shape, self))
msamples = self._forward_data(dataset.samples)
if __debug__:
debug('MAP_', "Make shallow copy of to-be-forward-mapped dataset "
"and assigned forward-mapped samples ({sf}a_filters: "
"%s, %s, %s)." % (self._sa_filter, self._fa_filter,
self._a_filter))
mds = dataset.copy(deep=False,
sa=self._sa_filter,
fa=self._fa_filter,
a=self._a_filter)
mds.samples = msamples
return mds
def _call(self, dataset=None):
"""Extract weights from SMLR classifier.
SMLR always has weights available, so nothing has to be computed here.
"""
clf = self.clf
# transpose to have the number of features on the second axis
# (as usual)
weights = clf.weights.T
if __debug__:
debug('SMLR',
"Extracting weights for %d-class SMLR" %
(len(weights) + 1) +
"Result: min=%f max=%f" %\
(np.min(weights), np.max(weights)))
# limit the labels to the number of sensitivity sets, to deal
# with the case of `fit_all_weights=False`
ds = Dataset(weights,
sa={clf.get_space(): clf._ulabels[:len(weights)]})
if clf.params.has_bias:
ds.sa['biases'] = clf.biases
return ds
def _call(self, dataset):
analyzers = []
# create analyzers
for clf in self.clf.clfs:
if self.__analyzer is None:
analyzer = clf.get_sensitivity_analyzer(**(self._slave_kwargs))
if analyzer is None:
raise ValueError, \
"Wasn't able to figure basic analyzer for clf %r" % \
(clf,)
if __debug__:
debug("SA", "Selected analyzer %r for clf %r" % \
(analyzer, clf))
else:
# XXX shallow copy should be enough...
analyzer = copy.copy(self.__analyzer)
# assign corresponding classifier
analyzer.clf = clf
# if clf was trained already - don't train again
if clf.trained:
analyzer._force_train = False
analyzers.append(analyzer)
self.__combined_analyzer.analyzers = analyzers
# XXX not sure if we don't want to call directly ._call(dataset) to avoid
# double application of transformers/combiners, after all we are just
# 'proxying' here to combined_analyzer...
# YOH: decided -- lets call ._call
return self.__combined_analyzer._call(dataset)
def forward(self, data):
"""Map data from input to output space.
Parameters
----------
data : Dataset-like, (at least 2D)-array-like
Typically this is a `Dataset`, but it might also be a plain data
array, or even something completely different(TM) that is supported
by a subclass' implementation. If such an object is Dataset-like it
is handled by a dedicated method that also transforms dataset
attributes if necessary. If an array-like is passed, it has to be
at least two-dimensional, with the first axis separating samples
or observations. For single samples `forward1()` might be more
appropriate.
"""
if is_datasetlike(data):
if __debug__:
debug('MAP', "Forward-map %s-shaped dataset through '%s'."
% (data.shape, self))
return self._forward_dataset(data)
else:
if hasattr(data, 'ndim') and data.ndim < 2:
raise ValueError(
'Mapper.forward() only support mapping of data with '
'at least two dimensions, where the first axis '
'separates samples/observations. Consider using '
'Mapper.forward1() instead.')
if __debug__:
debug('MAP', "Forward-map data through '%s'." % (self))
return self._forward_data(data)
def _wm_reverse(self, data):
if __debug__:
debug('MAP', "Converting signal back using DWP")
if self.__level is None:
raise NotImplementedError
else:
if not externals.exists('pywt wp reconstruct'):
raise NotImplementedError, \
"Reconstruction for a single level for versions of " \
"pywt < 0.1.7 (revision 103) is not supported"
if not externals.exists('pywt wp reconstruct fixed'):
warning(
"%s: Reverse mapping with this version of 'pywt' might "
"result in incorrect data in the tails of the signal. "
"Please check for an update of 'pywt', or be careful "
"when interpreting the edges of the reverse mapped "
"data." % self.__class__.__name__)
return self.__reverse_single_level(data)
def __init__(self, kernel=None, **kwargs):
"""Initialize a GPR regression analysis.
Parameters
----------
kernel : Kernel
a kernel object defining the covariance between instances.
(Defaults to SquaredExponentialKernel if None in arguments)
"""
# init base class first
Classifier.__init__(self, **kwargs)
# It does not make sense to calculate a confusion matrix for a GPR
# XXX it does ;) it will be a RegressionStatistics actually ;-)
# So if someone desires -- let him have it
# self.ca.enable('training_stats', False)
# set kernel:
if kernel is None:
kernel = SquaredExponentialKernel()
debug("GPR",
"No kernel was provided, falling back to default: %s"
% kernel)
self.__kernel = kernel
# append proper clf_internal depending on the kernel
# TODO: add "__tags__" to kernels since the check
# below does not scale
if isinstance(kernel, GeneralizedLinearKernel) or \
isinstance(kernel, LinearKernel):
self.__tags__ += ['linear']
else:
self.__tags__ += ['non-linear']
if externals.exists('openopt') \
and not 'has_sensitivity' in self.__tags__:
self.__tags__ += ['has_sensitivity']
# No need to initialize conditional attributes. Unless they got set
# they would raise an exception self.predicted_variances =
# None self.log_marginal_likelihood = None
self._init_internals()
pass
def __check(name, a='__version__'):
exec "import %s" % name
# it might be lxml.etree, so take only first module
topmodname = name.split('.')[0]
try:
v = getattr(sys.modules[name], '__version__')
except Exception as e:
# we can't assign version but it is there
if __debug__:
debug('EXT',
'Failed to acquire a version of %(name)s: %(e)s' % locals())
# if module is present but does not bear __version__
try:
import pkg_resources
v = pkg_resources.get_distribution(topmodname).version
except Exception as e:
# and if all that failed -- just assign '0'
v = '0'
versions[topmodname] = SmartVersion(v)
return True # we did manage to import it -- so it is there
def _get_cvec(self, data):
"""Estimate default and return scaled by it negative user's C values
"""
if not 'C' in self.params:#svm_type in [_svm.svmc.C_SVC]:
raise RuntimeError("Requested estimation of default C whenever C was not set")
C = self.params.C
if not is_sequence_type(C):
# we were not given a tuple for balancing between classes
C = [C]
Cs = list(C[:]) # copy
for i in range(len(Cs)):
if Cs[i] < 0:
Cs[i] = self._get_default_c(data.samples)*abs(Cs[i])
if __debug__:
debug("SVM", "Default C for %s was computed to be %s" %
(C[i], Cs[i]))
return Cs
def test_debug(self):
verbose.handlers = [] # so debug doesn't spoil it
debug.active = ['1', '2', 'SLC']
debug.metrics = debug._known_metrics.keys()
# do not offset for this test
debug('SLC', self.msg, lf=False)
self.assertRaises(ValueError, debug, 3, 'bugga')
#Should complain about unknown debug id
svalue = self.sout.getvalue()
regexp = "\[SLC\] DBG(?:{.*})?: %s" % self.msg
rematch = re.match(regexp, svalue)
self.assertTrue(rematch, msg="Cannot match %s with regexp %s" %
(svalue, regexp))
# find metrics
self.assertTrue('RSS/VMS:' in svalue,
msg="Cannot find vmem metric in " + svalue)
self.assertTrue('>test_verbosity:' in svalue,
msg="Cannot find tbc metric in " + svalue)
self.assertTrue(' sec' in svalue,
msg="Cannot find tbc metric in " + svalue)
def _recon_customobj_customrecon(hdf, memo):
"""Reconstruct a custom object from HDF using a custom recontructor"""
# we found something that has some special idea about how it wants
# to be reconstructed
mod_name = hdf.attrs['module']
recon_name = hdf.attrs['recon']
if __debug__:
debug(
'HDF5', "Load from custom reconstructor '%s.%s' [%s]" %
(mod_name, recon_name, hdf.name))
# turn names into definitions
try:
mod = __import__(mod_name, fromlist=[recon_name])
except ImportError, e:
if mod_name.startswith('mvpa') and not mod_name.startswith('mvpa2'):
# try to be gentle on data that got stored with PyMVPA 0.5 or 0.6
mod_name = mod_name.replace('mvpa', 'mvpa2', 1)
mod = __import__(mod_name, fromlist=[recon_name])
else:
raise e
def __call__(self, ds):
"""
.. note:
Will raise KeyError if lookup for sample_ids fails, or ds has not
been mapped at all
"""
if (not 'magic_id' in ds.a) or ds.a.magic_id != self._orig_ds_id:
raise KeyError, \
'Dataset %s is not indexed by %s' % (ds, self)
_map = self._map
_origids = ds.sa.origids
res = np.array([_map[i] for i in _origids])
if __debug__:
debug('SAL',
"Successful lookup: %(inst)s on %(ds)s having "
"origids=%(origids)s resulted in %(res)s",
msgargs=dict(inst=self, ds=ds, origids=_origids, res=res))
return res
def wr1996(size=200):
"""Generate '6d robot arm' dataset (Williams and Rasmussen 1996)
Was originally created in order to test the correctness of the
implementation of kernel ARD. For full details see:
http://www.gaussianprocess.org/gpml/code/matlab/doc/regression.html#ard
x_1 picked randomly in [-1.932, -0.453]
x_2 picked randomly in [0.534, 3.142]
r_1 = 2.0
r_2 = 1.3
f(x_1,x_2) = r_1 cos (x_1) + r_2 cos(x_1 + x_2) + N(0,0.0025)
etc.
Expected relevances:
ell_1 1.804377
ell_2 1.963956
ell_3 8.884361
ell_4 34.417657
ell_5 1081.610451
ell_6 375.445823
sigma_f 2.379139
sigma_n 0.050835
"""
intervals = np.array([[-1.932, -0.453], [0.534, 3.142]])
r = np.array([2.0, 1.3])
x = np.random.rand(size, 2)
x *= np.array(intervals[:, 1] - intervals[:, 0])
x += np.array(intervals[:, 0])
if __debug__:
for i in xrange(2):
debug(
'DG', '%d columnt Min: %g Max: %g' %
(i, x[:, i].min(), x[:, i].max()))
y = r[0] * np.cos(x[:, 0] + r[1] * np.cos(x.sum(1))) + \
np.random.randn(size) * np.sqrt(0.0025)
y -= y.mean()
x34 = x + np.random.randn(size, 2) * 0.02
x56 = np.random.randn(size, 2)
x = np.hstack([x, x34, x56])
return dataset_wizard(samples=x, targets=y)
def reverse(self, data):
"""Reverse-map data from output back into input space.
Parameters
----------
data : Dataset-like, anything
Typically this is a `Dataset`, but it might also be a plain data
array, or even something completely different(TM) that is supported
by a subclass' implementation. If such an object is Dataset-like it
is handled by a dedicated method that also transforms dataset
attributes if necessary.
"""
if is_datasetlike(data):
if __debug__:
debug('MAP', "Reverse-map %s-shaped dataset through '%s'."
% (data.shape, self))
return self._reverse_dataset(data)
else:
if __debug__:
debug('MAP', "Reverse-map data through '%s'." % (self))
return self._reverse_data(data)
def _set_matplotlib_backend():
"""Check if we have custom backend to set and it is different
from current one
"""
backend = cfg.get('matplotlib', 'backend')
if backend:
import matplotlib as mpl
mpl_backend = mpl.get_backend().lower()
if mpl_backend != backend.lower():
if __debug__:
debug('EXT_', "Trying to set matplotlib backend to %s" % backend)
mpl.use(backend)
import warnings
# And disable useless warning from matplotlib in the future
warnings.filterwarnings(
'ignore', 'This call to matplotlib.use() has no effect.*',
UserWarning)
elif __debug__:
debug('EXT_',
"Not trying to set matplotlib backend to %s since it was "
"already set" % backend)
def reverse1(self, data):
"""Wrapper method to map single samples.
It is basically identical to `reverse()`, but accepts one-dimensional
arguments. To map whole dataset this method cannot be used. but
`reverse()` handles them.
"""
if isinstance(data, np.ndarray):
data = data[np.newaxis]
else:
data = np.array([data])
if __debug__:
debug(
'MAP', "Reverse-map single %s-shaped sample through '%s'." %
(data.shape[1:], self))
mapped = self.reverse(data)[0]
if __debug__:
debug(
'MAP', "Mapped single %s-shaped sample to %s." %
(data.shape[1:], mapped.shape))
return mapped
def __init__(self, space=None, postproc=None, **kwargs):
"""
Parameters
----------
space: str, optional
Name of the 'processing space'. The actual meaning of this argument
heavily depends on the sub-class implementation. In general, this is
a trigger that tells the node to compute and store information about
the input data that is "interesting" in the context of the
corresponding processing in the output dataset.
postproc : Node instance, optional
Node to perform post-processing of results. This node is applied
in `__call__()` to perform a final processing step on the to be
result dataset. If None, nothing is done.
"""
ClassWithCollections.__init__(self, **kwargs)
if __debug__:
debug("NO", "Init node '%s' (space: '%s', postproc: '%s')",
(self.__class__.__name__, space, str(postproc)))
self.set_space(space)
self.set_postproc(postproc)
def _cache(self, ds1, ds2=None):
"""Initializes internal lookups + _kfull via caching the kernel matrix
"""
if __debug__ and 'KRN' in debug.active:
debug('KRN', "Caching %(inst)s for ds1=%(ds1)s, ds2=%(ds1)s"
% dict(inst=self, ds1=ds1, ds2=ds2))
self._lhsids = SamplesLookup(ds1)
if (ds2 is None) or (ds2 is ds1):
self._rhsids = self._lhsids
else:
self._rhsids = SamplesLookup(ds2)
ckernel = self._kernel
ckernel.compute(ds1, ds2)
self._kfull = ckernel.as_raw_np()
ckernel.cleanup()
self._k = self._kfull
self._recomputed = True
self.params.reset()
def _concat_results(sl=None, dataset=None, roi_ids=None, results=None):
"""The simplest implementation for collecting the results --
just put them into a list
This this implementation simply collects them into a list and
uses only sl. for assigning conditional attributes. But
custom implementation might make use of more/less of them.
Implemented as @staticmethod just to emphasize that in
principle it is independent of the actual searchlight instance
"""
# collect results
results = sum(results, [])
if __debug__ and 'SLC' in debug.active:
debug('SLC', '') # just newline
resshape = len(results) and np.asanyarray(results[0]).shape or 'N/A'
debug('SLC', ' hstacking %d results of shape %s'
% (len(results), resshape))
# but be careful: this call also serves as conversion from parallel maps
# to regular lists!
# this uses the Dataset-hstack
result_ds = hstack(results)
if __debug__:
debug('SLC', " hstacked shape %s" % (result_ds.shape,))
if sl.ca.is_enabled('roi_feature_ids'):
sl.ca.roi_feature_ids = [r.a.roi_feature_ids for r in results]
if sl.ca.is_enabled('roi_sizes'):
sl.ca.roi_sizes = [r.a.roi_sizes for r in results]
if sl.ca.is_enabled('roi_center_ids'):
sl.ca.roi_center_ids = [r.a.roi_center_ids for r in results]
if 'mapper' in dataset.a:
# since we know the space we can stick the original mapper into the
# results as well
if roi_ids is None:
result_ds.a['mapper'] = copy.copy(dataset.a.mapper)
else:
# there is an additional selection step that needs to be
# expressed by another mapper
mapper = copy.copy(dataset.a.mapper)
# NNO if the orignal mapper has no append (because it's not a
# chainmapper, for example), we make our own chainmapper.
feat_sel_mapper = StaticFeatureSelection(
roi_ids, dshape=dataset.shape[1:])
if hasattr(mapper, 'append'):
mapper.append(feat_sel_mapper)
else:
mapper = ChainMapper([dataset.a.mapper,
feat_sel_mapper])
result_ds.a['mapper'] = mapper
# store the center ids as a feature attribute
result_ds.fa['center_ids'] = roi_ids
return result_ds
def newfunc(*arg, **kwargs):
nfailed, i = 0, 0 # define i just in case
for i in range(niter):
try:
ret = func(*arg, **kwargs)
if i + 1 - nfailed >= niter - nfailures:
# so we know already that we wouldn't go over
# nfailures
break
except AssertionError as e:
nfailed += 1
if __debug__:
debug('TEST', "Upon %i-th run, test %s failed with %s",
(i, func.__name__, e))
if nfailed > nfailures:
if __debug__:
debug(
'TEST', "Ran %s %i times. Got %d failures, "
"while was allowed %d "
"-- re-throwing the last failure %s",
(func.__name__, i + 1, nfailed, nfailures, e))
exc_info = sys.exc_info()
raise exc_info[1].with_traceback(exc_info[2])
if __debug__:
debug('TEST', "Ran %s %i times. Got %d failures.",
(func.__name__, i + 1, nfailed))
return ret
def __smart_apply_along_axis(self, data):
# because apply_along_axis could be very much slower than a
# direct invocation of native functions capable of operating
# along specific axis, let's make it smarter for those we know
# could do that.
fx = None
naxis = {'samples': 0, 'features': 1}[self.__axis]
try:
# if first argument is 'axis' -- just proceed with a native call
if inspect.getargs(self.__fx.__code__).args[1] == 'axis':
fx = self.__fx
elif __debug__:
debug('FX', "Will apply %s via apply_along_axis",
(self.__fx))
except Exception as e:
if __debug__:
debug('FX',
"Failed to deduce either %s has 'axis' argument: %s",
(self.__fx, repr(e)))
pass
if fx is not None:
if __debug__:
debug('FX', "Applying %s directly to data giving axis=%d",
(self.__fx, naxis))
mdata = fx(data, naxis, *self.__fxargs)
else:
# either failed to deduce signature or just didn't
# have 'axis' second
# apply fx along naxis for each sample/feature
mdata = np.apply_along_axis(self.__fx, naxis, data, *self.__fxargs)
assert(mdata.ndim in (data.ndim, data.ndim-1))
return mdata
def reverse1(self, data):
"""Reverse-maps data or datasets through the chain (backwards).
See `Mapper` for more information.
"""
mp = data
for i, m in enumerate(reversed(self)):
# we ignore mapper that do not have reverse mapping implemented
# (e.g. detrending). That might cause problems if ignoring the
# mapper make the data incompatible input for the next mapper in
# the chain. If that pops up, we have to think about a proper
# solution.
try:
if __debug__:
debug('MAP',
"Reversing single %s-shaped input though chain node '%s'."
% (mp.shape, str(m)))
mp = m.reverse1(mp)
except NotImplementedError:
if __debug__:
debug('MAP', "Ignoring %s on reverse mapping." % m)
except ValueError:
if __debug__:
debug('MAP',
"Failed to reverse-map through chain at '%s'. Maybe "
"previous mapper return multiple samples. Trying to "
"switch to reverse() for the remainder of the chain."
% str(m))
mp = self[:-1 * i].reverse(mp)
return mp
return mp
def _forward_dataset(self, dataset):
"""Forward-map a dataset.
This is a private method that can be reimplemented in derived
classes. The default implementation forward-maps the dataset samples
and returns a new dataset that is a shallow copy of the input with
the mapped samples.
Parameters
----------
dataset : Dataset-like
"""
if __debug__:
debug('MAP_', "Forward-map %s-shaped samples in dataset with '%s'."
% (dataset.samples.shape, self))
msamples = self._forward_data(dataset.samples)
if __debug__:
debug('MAP_', "Make shallow copy of to-be-forward-mapped dataset "
"and assigned forward-mapped samples ({sf}a_filters: "
"%s, %s, %s)." % (self._sa_filter, self._fa_filter,
self._a_filter))
mds = dataset.copy(deep=False,
sa=self._sa_filter,
fa=self._fa_filter,
a=self._a_filter)
mds.samples = msamples
_assure_consistent_a(mds, dataset.shape)
if __debug__:
debug('MAP_', "Return forward-mapped dataset.")
return mds
def _concat_results(sl=None, dataset=None, roi_ids=None, results=None):
"""The simplest implementation for collecting the results --
just put them into a list
This this implementation simply collects them into a list and
uses only sl. for assigning conditional attributes. But
custom implementation might make use of more/less of them.
Implemented as @staticmethod just to emphasize that in
principle it is independent of the actual searchlight instance
"""
# collect results
results = sum(results, [])
if __debug__ and 'SLC' in debug.active:
debug('SLC', '') # just newline
resshape = len(results) and np.asanyarray(
results[0]).shape or 'N/A'
debug(
'SLC',
' hstacking %d results of shape %s' % (len(results), resshape))
# but be careful: this call also serves as conversion from parallel maps
# to regular lists!
# this uses the Dataset-hstack
result_ds = hstack(results)
if __debug__:
debug('SLC', " hstacked shape %s" % (result_ds.shape, ))
if sl.ca.is_enabled('roi_feature_ids'):
sl.ca.roi_feature_ids = [r.a.roi_feature_ids for r in results]
if sl.ca.is_enabled('roi_sizes'):
sl.ca.roi_sizes = [r.a.roi_sizes for r in results]
return result_ds
def __call__(self, ds):
# overwrite __call__ to perform a rigorous check whether the learner was
# trained before use and auto-train
if self.is_trained:
# already trained
if self.force_train:
if __debug__:
debug('LRN', "Forcing training of %s on %s",
(self, ds))
# but retraining is enforced
self.train(ds)
elif __debug__:
debug('LRN', "Skipping training of already trained %s on %s",
(self, ds))
else:
# not trained
if self.auto_train:
# auto training requested
if __debug__:
debug('LRN', "Auto-training %s on %s",
(self, ds))
self.train(ds)
else:
# we always have to have trained before using a learner
raise RuntimeError("%s needs to be trained before it can be "
"used and auto training is disabled."
% str(self))
return super(Learner, self).__call__(ds)
def __init__(self, startpoints, boxlength, offset=0, **kwargs):
"""
Parameters
----------
startpoints : sequence
Index values along the first axis of 'data'.
boxlength : int
The number of elements after 'startpoint' along the first axis of
'data' to be considered for the boxcar.
offset : int
The offset between the provided starting point and the actual start
of the boxcar.
"""
Mapper.__init__(self, **kwargs)
self._outshape = None
startpoints = np.asanyarray(startpoints)
if np.issubdtype(startpoints.dtype, 'i'):
self.startpoints = startpoints
else:
if __debug__:
debug(
'MAP', "Boxcar: obtained startpoints are not of int type."
" Rounding and changing dtype")
self.startpoints = np.asanyarray(np.round(startpoints), dtype='i')
# Sanity checks
if boxlength < 1:
raise ValueError, "Boxlength lower than 1 makes no sense."
if boxlength - int(boxlength) != 0:
raise ValueError, "boxlength must be an integer value."
self.boxlength = int(boxlength)
self.offset = offset
self.__selectors = None
# build a list of list where each sublist contains the indexes of to be
# averaged data elements
self.__selectors = [ slice(i + offset, i + offset + boxlength) \
for i in startpoints ]
def mask2slice(mask):
"""Convert a boolean mask vector into an equivalent slice (if possible).
Parameters
----------
mask: boolean array
The mask.
Returns
-------
slice or boolean array
If possible the boolean mask is converted into a `slice`. If this is not
possible the unmodified boolean mask is returned.
"""
# the filter should be a boolean array
# TODO Could be easily extended to also accept index arrays
if not len(mask):
raise ValueError("Got an empty mask.")
# get indices of non-zero filter elements
idx = mask.nonzero()[0]
if not len(idx):
return slice(0)
idx_start = idx[0]
idx_end = idx[-1] + 1
idx_step = None
if len(idx) > 1:
# we need to figure out if there is a regular step-size
# between elements
stepsizes = np.unique(idx[1:] - idx[:-1])
if len(stepsizes) > 1:
# multiple step-sizes -> slicing is not possible -> return
# orginal filter
return mask
else:
idx_step = stepsizes[0]
sl = slice(idx_start, idx_end, idx_step)
if __debug__:
debug("SPL", "Boolean mask conversion to slice is possible (%s)." % sl)
return sl
def df(x):
"""
Proxy to the log_marginal_likelihood first
derivative. Necessary for OpenOpt when using derivatives.
"""
self.hyp_running_guess[self.freeHypers] = x
# REMOVE print "df guess:",self.hyp_running_guess,x
# XXX EO: Most of the following lines can be skipped if
# df() is computed just after f() with the same
# hyperparameters. The partial results obtained during f()
# are what is needed for df(). For now, in order to avoid
# bugs difficult to trace, we keep this redunundancy. A
# deep check with how OpenOpt works or using memoization
# should solve this issue.
try:
if self.logscale:
self.parametric_model.set_hyperparameters(
np.exp(self.hyp_running_guess))
else:
self.parametric_model.set_hyperparameters(
self.hyp_running_guess)
pass
except InvalidHyperparameterError:
if __debug__:
debug("MOD_SEL", "WARNING: invalid hyperparameters!")
return -np.inf
# Check if it is possible to avoid useless computations
# already done in f(). According to tests and information
# collected from OpenOpt people, it is sufficiently
# unexpected that the following test succeed:
if np.any(x != self.f_last_x):
if __debug__:
debug(
"MOD_SEL",
"UNEXPECTED: recomputing train+log_marginal_likelihood."
)
try:
self.parametric_model.train(self.dataset)
except (np.linalg.linalg.LinAlgError, SL.basic.LinAlgError,
ValueError):
if __debug__:
debug(
"MOD_SEL",
"WARNING: Cholesky failed! Invalid hyperparameters!"
)
# XXX EO: which value for the gradient to return to
# OpenOpt when hyperparameters are wrong?
return np.zeros(x.size)
log_marginal_likelihood = self.parametric_model.compute_log_marginal_likelihood(
) # recompute what's needed (to be safe) REMOVE IN FUTURE!
pass
if self.logscale:
gradient_log_marginal_likelihood = self.parametric_model.compute_gradient_log_marginal_likelihood_logscale(
)
else:
gradient_log_marginal_likelihood = self.parametric_model.compute_gradient_log_marginal_likelihood(
)
pass
# REMOVE print "grad:",gradient_log_marginal_likelihood
return gradient_log_marginal_likelihood[self.freeHypers]
def _recon_customobj_defaultrecon(hdf, memo):
"""Reconstruct a custom object from HDF using the default recontructor"""
cls_name = hdf.attrs['class']
mod_name = hdf.attrs['module']
if __debug__:
debug('HDF5', "Load class instance '%s.%s' instance [%s]"
% (mod_name, cls_name, hdf.name))
mod, cls = _import_from_thin_air(mod_name, cls_name)
# create the object
# use specialized __new__ if necessary or beneficial
pcls, = _get_subclass_entry(cls, ((dict,), (list,), (object,)),
"Do not know how to create instance of %(cls)s")
obj = pcls.__new__(cls)
# insert any stored object state
_update_obj_state_from_hdf(obj, hdf, memo)
# do we process a container?
if 'items' in hdf:
# charge the items -- handling depends on the parent class
pcls, umeth, cfunc = _get_subclass_entry(
cls,
((dict, 'update', _hdf_dict_to_obj),
(list, 'extend', _hdf_list_to_obj)),
"Unhandled container type (got: '%(cls)s').")
if __debug__:
debug('HDF5', "Populating %s object." % pcls)
getattr(obj, umeth)(cfunc(hdf, memo))
if __debug__:
debug('HDF5', "Loaded %i items." % len(obj))
return obj
def _set(self, val, init=False):
if self.constraints is not None:
# for c in self.constraints:
# val = c(val)
# #val = c.validate(val)
val = self.constraints(val)
different_value = self._value != val
isarray = isinstance(different_value, np.ndarray)
if self._ro and not init:
raise RuntimeError("Attempt to set read-only parameter %s to %s" \
% (self.name, val))
if (isarray and np.any(different_value)) or \
((not isarray) and different_value):
if __debug__:
debug("COL", "Parameter: setting %s to %s " % (str(self), val))
self._value = val
# Set 'isset' only if not called from initialization routine
self._isset = not init #True
elif __debug__:
debug("COL",
"Parameter: not setting %s since value is the same" \
% (str(self)))
def check_all_dependencies(force=False, verbosity=1):
"""
Test for all known dependencies.
Parameters
----------
force : boolean
Whether to force the test even if it has already been
performed.
"""
# loop over all known dependencies
for dep in _KNOWN:
if not exists(dep, force):
if verbosity:
warning("%s is not available." % dep)
if __debug__:
debug('EXT', 'The following optional externals are present: %s'
% [k[5:] for k in cfg.options('externals')
if k.startswith('have')
and cfg.getboolean('externals', k)])
def reverse(self, data):
"""Reverse-maps data or datasets through the chain (backwards).
See `Mapper` for more information.
"""
mp = data
for m in reversed(self):
# we ignore mapper that do not have reverse mapping implemented
# (e.g. detrending). That might cause problems if ignoring the
# mapper make the data incompatible input for the next mapper in
# the chain. If that pops up, we have to think about a proper
# solution.
try:
if __debug__:
debug('MAP',
"Reversing %s-shaped input though '%s'."
% (mp.shape, str(m)))
mp = m.reverse(mp)
except NotImplementedError:
if __debug__:
debug('MAP', "Ignoring %s on reverse mapping." % m)
return mp
def __init__(self, value=None, name=None, doc=None):
"""
Parameters
----------
value : arbitrary (see derived implementations)
The actual value of this attribute.
name : str
Name of the collectable under which it should be available in its
respective collection.
doc : str
Documentation about the purpose of this collectable.
"""
if doc is not None:
# to prevent newlines in the docstring
doc = re.sub('[\n ]+', ' ', doc)
self.__doc__ = doc
self.__name = name
self._value = None
if not value is None:
self._set(value)
if __debug__ and __mvpadebug__:
debug("COL", "Initialized %r", (self, ))
def solve(self, problem=None):
"""Solve the maximization problem, check outcome and collect results.
"""
# XXX: this method can be made more abstract in future in the
# sense that it could work not only for
# log_marginal_likelihood but other measures as well
# (e.g. cross-valideted error).
if np.all(self.freeHypers==False): # no optimization needed
self.hyperparameters_best = self.hyp_initial_guess.copy()
try:
self.parametric_model.set_hyperparameters(self.hyperparameters_best)
except InvalidHyperparameterError:
if __debug__: debug("MOD_SEL", "WARNING: invalid hyperparameters!")
self.log_marginal_likelihood_best = -np.inf
return self.log_marginal_likelihood_best
self.parametric_model.train(self.dataset)
self.log_marginal_likelihood_best = self.parametric_model.compute_log_marginal_likelihood()
return self.log_marginal_likelihood_best
result = self.problem.solve(self.optimization_algorithm) # perform optimization!
if result.stopcase == -1:
# XXX: should we use debug() for the following messages?
# If so, how can we track the missing convergence to a
# solution?
print "Unable to find a maximum to log_marginal_likelihood"
elif result.stopcase == 0:
print "Limits exceeded"
elif result.stopcase == 1:
self.hyperparameters_best = self.hyp_initial_guess.copy()
if self.logscale:
self.hyperparameters_best[self.freeHypers] = np.exp(result.xf) # best hyperparameters found # NOTE is it better to return a copy?
else:
self.hyperparameters_best[self.freeHypers] = result.xf
pass
self.log_marginal_likelihood_best = result.ff # actual best vuale of log_marginal_likelihood
pass
self.stopcase = result.stopcase
return self.log_marginal_likelihood_best
def _get_default_c(self, data):
"""Compute default C
TODO: for non-linear SVMs
"""
if self.params.kernel.__kernel_name__ == 'linear':
# TODO: move into a function wrapper for
# np.linalg.norm
if np.issubdtype(data.dtype, np.integer):
# we are dealing with integers and overflows are
# possible, so assure working with floats
def sq_func(x):
y = x.astype(float) # copy as float
y *= y # in-place square
return y
else:
sq_func = np.square
# perform it per each sample so we do not double memory
# with calling sq_func on full data
# Having a list of norms here automagically resolves issue
# with memmapped operations on which return
# in turn another memmap
datasetnorm = np.mean([np.sqrt(np.sum(sq_func(s)))
for s in data])
if datasetnorm == 0:
warning("Obtained degenerate data with zero norm for training "
"of %s. Scaling of C cannot be done." % self)
return 1.0
value = 1.0/(datasetnorm**2)
if __debug__:
debug("SVM", "Default C computed to be %f" % value)
else:
warning("TODO: Computation of default C is not yet implemented" +
" for non-linear SVMs. Assigning 1.0")
value = 1.0
return value