Python attrmap.AttributeMap类代码示例

def test_attrmap_conflicts():
    am_n = AttributeMap({'a':1, 'b':2, 'c':1})
    am_t = AttributeMap({'a':1, 'b':2, 'c':1}, collisions_resolution='tuple')
    am_l = AttributeMap({'a':1, 'b':2, 'c':1}, collisions_resolution='lucky')
    q_f = ['a', 'b', 'a', 'c']
    # should have no effect on forward mapping
    ok_(np.all(am_n.to_numeric(q_f) == am_t.to_numeric(q_f)))
    ok_(np.all(am_t.to_numeric(q_f) == am_l.to_numeric(q_f)))

    assert_raises(ValueError, am_n.to_literal, [2])
    r_t = am_t.to_literal([2, 1])
    r_l = am_l.to_literal([2, 1])

    def _call(self, dataset):
        sens = super(self.__class__, self)._call(dataset)
        clf = self.clf
        targets_attr = clf.get_space()
        if targets_attr in sens.sa:
            # if labels are present -- transform them into meaningful tuples
            # (or not if just a single beast)
            am = AttributeMap(dict([(l, -1) for l in clf.neglabels] + [(l, +1) for l in clf.poslabels]))

            # XXX here we still can get a sensitivity per each label
            # (e.g. with SMLR as the slave clf), so I guess we should
            # tune up Multiclass...Analyzer to add an additional sa
            # And here we might need to check if asobjarray call is necessary
            # and should be actually done
            # asobjarray(
            sens.sa[targets_attr] = am.to_literal(sens.sa[targets_attr].value, recurse=True)
        return sens

    def _test_gpr_model_selection(self):  # pragma: no cover
        """Smoke test for running model selection while getting GPRWeights

        TODO: DISABLED because setting of hyperparameters was not adopted for 0.6 (yet)
        """
        if not externals.exists('openopt'):
            return
        amap = AttributeMap()           # we would need to pass numbers into the GPR
        dataset = datasets['uni2small'].copy() #data_generators.linear1d_gaussian_noise()
        dataset.targets = amap.to_numeric(dataset.targets).astype(float)
        k = GeneralizedLinearKernel()
        clf = GPR(k, enable_ca=['log_marginal_likelihood'])
        sa = clf.get_sensitivity_analyzer() # should be regular weights
        sa_ms = clf.get_sensitivity_analyzer(flavor='model_select') # with model selection
        def prints():
            print clf.ca.log_marginal_likelihood, clf.kernel.Sigma_p, clf.kernel.sigma_0

        sa(dataset)
        lml = clf.ca.log_marginal_likelihood

        sa_ms(dataset)
        lml_ms = clf.ca.log_marginal_likelihood

        self.assertTrue(lml_ms > lml)

    def __init__(self, space=None, **kwargs):
        # by default we want classifiers to use the 'targets' sample attribute
        # for training/testing
        if space is None:
            space = 'targets'
        Learner.__init__(self, space=space, **kwargs)

        # XXX
        # the place to map literal to numerical labels (and back)
        # this needs to be in the base class, since some classifiers also
        # have this nasty 'regression' mode, and the code in this class
        # needs to deal with converting the regression output into discrete
        # labels
        # however, preferably the mapping should be kept in the respective
        # low-level implementations that need it
        self._attrmap = AttributeMap()

        self.__trainednfeatures = 0
        """Stores number of features for which classifier was trained.
        If 0 -- it wasn't trained at all"""

        self._set_retrainable(self.params.retrainable, force=True)

class Classifier(Learner):
    """Abstract classifier class to be inherited by all classifiers
    """

    # Kept separate from doc to don't pollute help(clf), especially if
    # we including help for the parent class
    _DEV__doc__ = """
    Required behavior:

    For every classifier is has to be possible to be instantiated without
    having to specify the training pattern.

    Repeated calls to the train() method with different training data have to
    result in a valid classifier, trained for the particular dataset.

    It must be possible to specify all classifier parameters as keyword
    arguments to the constructor.

    Recommended behavior:

    Derived classifiers should provide access to *estimates* -- i.e. that
    information that is finally used to determine the predicted class label.

    Michael: Maybe it works well if each classifier provides a 'estimates'
             state member. This variable is a list as long as and in same order
             as Dataset.uniquetargets (training data). Each item in the list
             corresponds to the likelyhood of a sample to belong to the
             respective class. However the semantics might differ between
             classifiers, e.g. kNN would probably store distances to class-
             neighbors, where PLR would store the raw function value of the
             logistic function. So in the case of kNN low is predictive and for
             PLR high is predictive. Don't know if there is the need to unify
             that.

             As the storage and/or computation of this information might be
             demanding its collection should be switchable and off be default.

    Nomenclature
     * predictions  : result of the last call to .predict()
     * estimates : might be different from predictions if a classifier's predict()
                   makes a decision based on some internal value such as
                   probability or a distance.
    """
    # Dict that contains the parameters of a classifier.
    # This shall provide an interface to plug generic parameter optimizer
    # on all classifiers (e.g. grid- or line-search optimizer)
    # A dictionary is used because Michael thinks that access by name is nicer.
    # Additionally Michael thinks ATM that additional information might be
    # necessary in some situations (e.g. reasonably predefined parameter range,
    # minimal iteration stepsize, ...), therefore the value to each key should
    # also be a dict or we should use mvpa2.base.param.Parameter'...

    training_stats = ConditionalAttribute(enabled=False,
        doc="Confusion matrix of learning performance")

    predictions = ConditionalAttribute(enabled=True,
        doc="Most recent set of predictions")

    estimates = ConditionalAttribute(enabled=True,
        doc="Internal classifier estimates the most recent " +
            "predictions are based on")

    predicting_time = ConditionalAttribute(enabled=True,
        doc="Time (in seconds) which took classifier to predict")

    __tags__ = []
    """Describes some specifics about the classifier -- is that it is
    doing regression for instance...."""

    # TODO: make it available only for actually retrainable classifiers
    retrainable = Parameter(False, allowedtype='bool',
        doc="""Either to enable retraining for 'retrainable' classifier.""",
        index=1002)


    def __init__(self, space=None, **kwargs):
        # by default we want classifiers to use the 'targets' sample attribute
        # for training/testing
        if space is None:
            space = 'targets'
        Learner.__init__(self, space=space, **kwargs)

        # XXX
        # the place to map literal to numerical labels (and back)
        # this needs to be in the base class, since some classifiers also
        # have this nasty 'regression' mode, and the code in this class
        # needs to deal with converting the regression output into discrete
        # labels
        # however, preferably the mapping should be kept in the respective
        # low-level implementations that need it
        self._attrmap = AttributeMap()

        self.__trainednfeatures = 0
        """Stores number of features for which classifier was trained.
        If 0 -- it wasn't trained at all"""

        self._set_retrainable(self.params.retrainable, force=True)

        # deprecate
        #self.__trainedidhash = None
#.........这里部分代码省略.........

    def _train(self, dataset):
        """Train SVM
        """

        # XXX watchout
        # self.untrain()
        newkernel, newsvm = False, False
        # local bindings for faster lookup
        params = self.params
        retrainable = self.params.retrainable

        targets_sa_name = self.get_space()    # name of targets sa
        targets_sa = dataset.sa[targets_sa_name] # actual targets sa

        if retrainable:
            _changedData = self._changedData

        # LABELS
        ul = None
        self.__traindataset = dataset


        # OK -- we have to map labels since
        #  binary ones expect -1/+1
        #  Multiclass expect labels starting with 0, otherwise they puke
        #   when ran from ipython... yikes
        if __debug__:
            debug("SG_", "Creating labels instance")

        if self.__is_regression__:
            labels_ = np.asarray(targets_sa.value, dtype='double')
        else:
            ul = targets_sa.unique
            # ul.sort()

            if len(ul) == 2:
                # assure that we have -1/+1
                _labels_dict = {ul[0]:-1.0, ul[1]:+1.0}
            elif len(ul) < 2:
                raise FailedToTrainError, \
                      "We do not have 1-class SVM brought into SG yet"
            else:
                # can't use plain enumerate since we need them swapped
                _labels_dict = dict([ (ul[i], i) for i in range(len(ul))])

            # Create SG-customized attrmap to assure -1 / +1 if necessary
            self._attrmap = AttributeMap(_labels_dict, mapnumeric=True)

            if __debug__:
                debug("SG__", "Mapping labels using dict %s" % _labels_dict)
            labels_ = self._attrmap.to_numeric(targets_sa.value).astype(float)

        labels = shogun.Features.Labels(labels_)
        _setdebug(labels, 'Labels')


        # KERNEL

        # XXX cruel fix for now... whole retraining business needs to
        # be rethought
        if retrainable:
            _changedData['kernel_params'] = _changedData.get('kernel_params', False)

        # TODO: big RF to move non-kernel classifiers away
        if 'kernel-based' in self.__tags__ and (not retrainable
               or _changedData['traindata'] or _changedData['kernel_params']):
            # If needed compute or just collect arguments for SVM and for
            # the kernel

            if retrainable and __debug__:
                if _changedData['traindata']:
                    debug("SG",
                          "Re-Creating kernel since training data has changed")

                if _changedData['kernel_params']:
                    debug("SG",
                          "Re-Creating kernel since params %s has changed" %
                          _changedData['kernel_params'])


            k = self.params.kernel
            k.compute(dataset)
            self.__kernel = kernel = k.as_raw_sg()

            newkernel = True
            self.kernel_params.reset()  # mark them as not-changed
            #_setdebug(kernel, 'Kernels')

            #self.__condition_kernel(kernel)
            if retrainable:
                if __debug__:
                    debug("SG_", "Resetting test kernel for retrainable SVM")
                self.__kernel_test = None

        # TODO -- handle _changedData['params'] correctly, ie without recreating
        # whole SVM
        Cs = None
        if not retrainable or self.__svm is None or _changedData['params']:
            # SVM
            if self.params.has_key('C'):
#.........这里部分代码省略.........

#.........这里部分代码省略.........

        if retrainable:
            _changedData = self._changedData

        # LABELS
        ul = None
        self.__traindataset = dataset


        # OK -- we have to map labels since
        #  binary ones expect -1/+1
        #  Multiclass expect labels starting with 0, otherwise they puke
        #   when ran from ipython... yikes
        if __debug__:
            debug("SG_", "Creating labels instance")

        if self.__is_regression__:
            labels_ = np.asarray(targets_sa.value, dtype='double')
        else:
            ul = targets_sa.unique
            # ul.sort()

            if len(ul) == 2:
                # assure that we have -1/+1
                _labels_dict = {ul[0]:-1.0, ul[1]:+1.0}
            elif len(ul) < 2:
                raise FailedToTrainError, \
                      "We do not have 1-class SVM brought into SG yet"
            else:
                # can't use plain enumerate since we need them swapped
                _labels_dict = dict([ (ul[i], i) for i in range(len(ul))])

            # Create SG-customized attrmap to assure -1 / +1 if necessary
            self._attrmap = AttributeMap(_labels_dict, mapnumeric=True)

            if __debug__:
                debug("SG__", "Mapping labels using dict %s" % _labels_dict)
            labels_ = self._attrmap.to_numeric(targets_sa.value).astype(float)

        labels = shogun.Features.Labels(labels_)
        _setdebug(labels, 'Labels')


        # KERNEL

        # XXX cruel fix for now... whole retraining business needs to
        # be rethought
        if retrainable:
            _changedData['kernel_params'] = _changedData.get('kernel_params', False)

        # TODO: big RF to move non-kernel classifiers away
        if 'kernel-based' in self.__tags__ and (not retrainable
               or _changedData['traindata'] or _changedData['kernel_params']):
            # If needed compute or just collect arguments for SVM and for
            # the kernel

            if retrainable and __debug__:
                if _changedData['traindata']:
                    debug("SG",
                          "Re-Creating kernel since training data has changed")

                if _changedData['kernel_params']:
                    debug("SG",
                          "Re-Creating kernel since params %s has changed" %
                          _changedData['kernel_params'])

def plot_decision_boundary_2d(dataset, clf=None,
                              targets=None, regions=None, maps=None,
                              maps_res=50, vals=None,
                              data_callback=None):
    """Plot a scatter of a classifier's decision boundary and data points

    Assumes data is 2d (no way to visualize otherwise!!)

    Parameters
    ----------
    dataset : `Dataset`
      Data points to visualize (might be the data `clf` was train on, or
      any novel data).
    clf : `Classifier`, optional
      Trained classifier
    targets : string, optional
      What samples attributes to use for targets.  If None and clf is
      provided, then `clf.params.targets_attr` is used.
    regions : string, optional
      Plot regions (polygons) around groups of samples with the same
      attribute (and target attribute) values. E.g. chunks.
    maps : string in {'targets', 'estimates'}, optional
      Either plot underlying colored maps, such as clf predictions
      within the spanned regions, or estimates from the classifier
      (might not work for some).
    maps_res : int, optional
      Number of points in each direction to evaluate.
      Points are between axis limits, which are set automatically by
      matplotlib.  Higher number will yield smoother decision lines but come
      at the cost of O^2 classifying time/memory.
    vals : array of floats, optional
      Where to draw the contour lines if maps='estimates'
    data_callback : callable, optional
      Callable object to preprocess the new data points.
      Classified points of the form samples = data_callback(xysamples).
      I.e. this can be a function to normalize them, or cache them
      before they are classified.
    """
    if vals is None:
        vals = [-1, 0, 1]

    if False:
        ## from mvpa2.misc.data_generators import *
        ## from mvpa2.clfs.svm import *
        ## from mvpa2.clfs.knn import *
        ## ds = dumb_feature_binary_dataset()
        dataset = normal_feature_dataset(nfeatures=2, nchunks=5,
                                         snr=10, nlabels=4, means=[ [0,1], [1,0], [1,1], [0,0] ])
        dataset.samples += dataset.sa.chunks[:, None]*0.1 # slight shifts for chunks ;)
        #dataset = normal_feature_dataset(nfeatures=2, nlabels=3, means=[ [0,1], [1,0], [1,1] ])
        #dataset = normal_feature_dataset(nfeatures=2, nlabels=2, means=[ [0,1], [1,0] ])
        #clf = LinearCSVMC(C=-1)
        clf = kNN(4)#LinearCSVMC(C=-1)
        clf.train(dataset)
        #clf = None
        #plot_decision_boundary_2d(ds, clf)
        targets = 'targets'
        regions = 'chunks'
        #maps = 'estimates'
        maps = 'targets'
        #maps = None #'targets'
        res = 50
        vals = [-1, 0, 1]
        data_callback=None
        pl.clf()

    if dataset.nfeatures != 2:
        raise ValueError('Can only plot a decision boundary in 2D')

    Pioff()
    a = pl.gca() # f.add_subplot(1,1,1)

    attrmap = None
    if clf:
        estimates_were_enabled = clf.ca.is_enabled('estimates')
        clf.ca.enable('estimates')

        if targets is None:
            targets = clf.get_space()
        # Lets reuse classifiers attrmap if it is good enough
        attrmap = clf._attrmap
        predictions = clf.predict(dataset)

    targets_sa_name = targets           # bad Yarik -- will rebind targets to actual values
    targets_lit = dataset.sa[targets_sa_name].value
    utargets_lit = dataset.sa[targets_sa_name].unique

    if not (attrmap is not None
            and len(attrmap)
            and set(clf._attrmap.keys()).issuperset(utargets_lit)):
        # create our own
        attrmap = AttributeMap(mapnumeric=True)

    targets = attrmap.to_numeric(targets_lit)
    utargets = attrmap.to_numeric(utargets_lit)

    vmin = min(utargets)
    vmax = max(utargets)
    cmap = pl.cm.RdYlGn                  # argument

#.........这里部分代码省略.........

        except ValueError, e:
            print "Sorry - plotting of estimates isn't full supported for %s. " \
                  "Got exception %s" % (clf, e)
    elif maps == 'targets':
        map_values = attrmap.to_numeric(predictions_new).reshape(x.shape)
        a.imshow(map_values.T, **imshow_kwargs)
        #CS = a.contour(x, y, map_values, vals, zorder=6,
        #               linestyles=linestyles, extent=extent, colors='k')

    # Plot regions belonging to the same pair of attribute given
    # (e.g. chunks) and targets attribute
    if regions:
        chunks_sa = dataset.sa[regions]
        chunks_lit = chunks_sa.value
        uchunks_lit = chunks_sa.value
        chunks_attrmap = AttributeMap(mapnumeric=True)
        chunks = chunks_attrmap.to_numeric(chunks_lit)
        uchunks = chunks_attrmap.to_numeric(uchunks_lit)

        from matplotlib.delaunay.triangulate import Triangulation
        from matplotlib.patches import Polygon
        # Lets figure out convex halls for each chunk/label pair
        for target in utargets:
            t_mask = targets == target
            for chunk in uchunks:
                tc_mask = np.logical_and(t_mask,
                                        chunk == chunks)
                tc_samples = dataset.samples[tc_mask]
                tr = Triangulation(tc_samples[:, 0],
                                   tc_samples[:, 1])
                poly = pl.fill(tc_samples[tr.hull, 0],

def to_lightsvm_format(dataset, out, targets_attr='targets',
                       domain=None, am=None):
    """Export dataset into LightSVM format

    Parameters
    ----------
    dataset : Dataset
    out
      Anything understanding .write(string), such as `File`
    targets_attr : string, optional
      Name of the samples attribute to be output
    domain : {None, 'regression', 'binary', 'multiclass'}, optional
      What domain dataset belongs to.  If `None`, it would be deduced
      depending on the datatype ('regression' if float, classification
      in case of int or string, with 'binary'/'multiclass' depending on
      the number of unique targets)
    am : `AttributeMap` or None, optional
      Which mapping to use for storing the non-conformant targets. If
      None was provided, new one would be automagically generated
      depending on the given/deduced domain.

    Returns
    -------
    am

    LightSVM format is an ASCII representation with a single sample per
    each line::

      output featureIndex:featureValue ... featureIndex:featureValue

    where ``output`` is specific for a given domain:

    regression
      float number
    binary
      integer labels from {-1, 1}
    multiclass
      integer labels from {1..ds.targets_attr.nunique}

    """
    targets_a = dataset.sa[targets_attr]
    targets = targets_a.value

    # XXX this all below
    #  * might become cleaner
    #  * might be RF to become more generic to be used may be elsewhere as well

    if domain is None:
        if targets.dtype.kind in ['S', 'i']:
            if len(targets_a.unique) == 2:
                domain = 'binary'
            else:
                domain = 'multiclass'
        else:
            domain = 'regression'

    if domain in ['multiclass', 'binary']:
        # check if labels are appropriate and provide mapping if necessary
        utargets = targets_a.unique
        if domain == 'binary' and set(utargets) != set([-1, 1]):
            # need mapping
            if len(utargets) != 2:
                raise ValueError, \
                      "We need 2 unique targets in %s of %s. Got targets " \
                      "from set %s" % (targets_attr, dataset, utargets)
            if am is None:
                am = AttributeMap(dict(zip(utargets, [-1, 1])))
            elif set(am.keys()) != set([-1, 1]):
                raise ValueError, \
                      "Provided %s doesn't map into binary " \
                      "labels -1,+1" % (am,)
        elif domain == 'multiclass' \
                 and set(utargets) != set(range(1, len(utargets)+1)):
            if am is None:
                am = AttributeMap(dict(zip(utargets,
                                           range(1, len(utargets) + 1))))
            elif set(am.keys()) != set([-1, 1]):
                raise ValueError, \
                      "Provided %s doesn't map into multiclass " \
                      "range 1..N" % (am, )

    if am is not None:
        # map the targets
        targets = am.to_numeric(targets)

    for t, s in zip(targets, dataset.samples):
        out.write('%g %s\n'
                  % (t,
                     ' '.join(
                         '%i:%.8g' % (i, v)
                         for i,v in zip(range(1, dataset.nfeatures+1), s))))

    out.flush()                # push it out
    return am

def test_attrmap():
    map_default = {'eins': 0, 'zwei': 2, 'sieben': 1}
    map_custom = {'eins': 11, 'zwei': 22, 'sieben': 33}
    literal = ['eins', 'zwei', 'sieben', 'eins', 'sieben', 'eins']
    literal_nonmatching = ['uno', 'dos', 'tres']
    num_default = [0, 2, 1, 0, 1, 0]
    num_custom = [11, 22, 33, 11, 33, 11]

    # no custom mapping given
    am = AttributeMap()
    assert_false(am)
    ok_(len(am) == 0)
    assert_array_equal(am.to_numeric(literal), num_default)
    assert_array_equal(am.to_literal(num_default), literal)
    ok_(am)
    ok_(len(am) == 3)

    #
    # Tests for recursive mapping + preserving datatype
    class myarray(np.ndarray):
        pass

    assert_raises(KeyError, am.to_literal, [(1, 2), 2, 0])
    literal_fancy = [(1, 2), 2, [0], np.array([0, 1]).view(myarray)]
    literal_fancy_tuple = tuple(literal_fancy)
    literal_fancy_array = np.array(literal_fancy, dtype=object)

    for l in (literal_fancy, literal_fancy_tuple,
              literal_fancy_array):
        res = am.to_literal(l, recurse=True)
        assert_equal(res[0], ('sieben', 'zwei'))
        assert_equal(res[1], 'zwei')
        assert_equal(res[2], ['eins'])
        assert_array_equal(res[3], ['eins', 'sieben'])

        # types of result and subsequences should be preserved
        ok_(isinstance(res, l.__class__))
        ok_(isinstance(res[0], tuple))
        ok_(isinstance(res[1], str))
        ok_(isinstance(res[2], list))
        ok_(isinstance(res[3], myarray))

    # yet another example
    a = np.empty(1, dtype=object)
    a[0] = (0, 1)
    res = am.to_literal(a, recurse=True)
    ok_(isinstance(res[0], tuple))

    #
    # with custom mapping
    am = AttributeMap(map=map_custom)
    assert_array_equal(am.to_numeric(literal), num_custom)
    assert_array_equal(am.to_literal(num_custom), literal)

    # if not numeric nothing is mapped
    assert_array_equal(am.to_numeric(num_custom), num_custom)
    # even if the map doesn't fit
    assert_array_equal(am.to_numeric(num_default), num_default)

    # need to_numeric first
    am = AttributeMap()
    assert_raises(RuntimeError, am.to_literal, [1,2,3])
    # stupid args
    assert_raises(ValueError, AttributeMap, map=num_custom)

    # map mismatch
    am = AttributeMap(map=map_custom)
    if __debug__:
        # checked only in __debug__
        assert_raises(KeyError, am.to_numeric, literal_nonmatching)
    # needs reset and should work afterwards
    am.clear()
    assert_array_equal(am.to_numeric(literal_nonmatching), [2, 0, 1])
    # and now reverse
    am = AttributeMap(map=map_custom)
    assert_raises(KeyError, am.to_literal, num_default)

    # dict-like interface
    am = AttributeMap()

    ok_([(k, v) for k, v in am.iteritems()] == [])