[Predator] Add MetaWeight and MetaFeatureValue to group multiple weights and features together.

appengine/findit/crash/loglinear/feature.py

Index: appengine/findit/crash/loglinear/feature.py
diff --git a/appengine/findit/crash/loglinear/feature.py b/appengine/findit/crash/loglinear/feature.py
index 4633c9f1869866e3e76dff39fcf75fdf05609a90..f1652373580b6450d75083b6dbed894e5c78268c 100644
--- a/appengine/findit/crash/loglinear/feature.py
+++ b/appengine/findit/crash/loglinear/feature.py
@@ -3,9 +3,13 @@
 # found in the LICENSE file.
 
 from collections import namedtuple
+import logging
 import math
 
 import libs.math.logarithms as lmath
+from libs.math.vectors import vsum
+from libs.meta_object import Element
+from libs.meta_object import MetaDict
 
 
 def LinearlyScaled(value, maximum):
@@ -79,9 +83,8 @@ class ChangedFile(namedtuple('ChangedFile',
         % (self.__class__.__name__, self.name, self.blame_url, self.reasons))
 
 
-class FeatureValue(namedtuple('FeatureValue',
-    ['name', 'value', 'reason', 'changed_files'])): # pragma: no cover
-  """The result of an individual feature.
+class FeatureValue(Element): # pragma: no cover
+  """The result of an feature.
 
   Attributes:
     name (str): the name of the feature producing this value.
@@ -95,17 +98,162 @@ class FeatureValue(namedtuple('FeatureValue',
   """
   __slots__ = ()
 
-  def __new__(cls, name, value, reason, changed_files):
-    return super(cls, FeatureValue).__new__(cls,
-        str(name), float(value), str(reason), changed_files)
+  def __init__(self, name, value, reason, changed_files):
+    self._value = float(value)
+    self._name = name
+    self._reason = reason
+    self._changed_files = changed_files
+
+  @property
+  def name(self):
+    return self._name
+
+  @property
+  def value(self):
+    return self._value
+
+  @property
+  def reason(self):
+    return self._reason
+
+  @property
+  def changed_files(self):
+    return self._changed_files
 
   def __str__(self):
     return ('%s(name = %s, value = %f, reason = %s, changed_files = %s)'
         % (self.__class__.__name__, self.name, self.value, self.reason,
            self.changed_files))
 
+  def __len__(self):
+    return 1
+
+  def __mul__(self, number):
+    return self._value * float(number)
+
+  __rmul__ = __mul__
+
+  def __add__(self, number):
+    return self._value + float(number)
+
+  __radd__ = __add__
+
+  def __float__(self):
+    return self._value
+
+  def __eq__(self, other):
+    return (self.name == other.name and self._value == other._value and
+            self.reason == other.reason and
+            self.changed_files == other.changed_files)
+
+  def __ne__(self, other):
+    return not self.__eq__(other)
+
+
+class MetaFeatureValue(MetaDict):
+  """The result of a meta feature which groups a list of ``FeatureValue``s.
+
+  N.B. ``MetaFeatureValue`` must have more than one ``FeatureValue``.
+
+  Attributes:
+
+  """
+  def __init__(self, name, feature_values):
+    """
+    Args:
+      feature_value (dict of FeatureValue/MetaFeatureValue):
+        All the sub features that this ``MetaFeatureValue`` contains.
+    """
+    super(MetaFeatureValue, self).__init__(feature_values)
+    self._name = name
+    self._reason = None
+    self._changed_files = None
+
+  @property
+  def name(self):
+    return self._name
+
+  @property
+  def reason(self):
+    """Collect and format a list of all ``FeatureValue.reason`` strings.
+
+    Returns:
+      A str of reasons, each line has a format
+      "feature_name: feature_value -- reason" triples; where the first string is
+      the feature name, the float is some numeric representation of how
+      much influence this feature exerts on the ``Suspect`` being blamed,
+      and the final string is the ``FeatureValue.reason``. The list is
+      sorted by feature name, just to ensure that it comes out in some
+      canonical order.
+
+      At present, the float is the log-domain score of the feature
+      value. However, this isn't the best thing for UX reasons. In the
+      future it might be replaced by the normal-domain score, or by
+      the probability.
+    """
+    if self._reason:
+      return self._reason
+
+    formatted_reasons = []
+    for feature in self.itervalues():
+      if feature.reason:
+        formatted_reasons.append('%s: %f -- %s' % (feature.name,
+                                                   feature.value,
+                                                   feature.reason))
+
+    formatted_reasons.sort()
+    self._reason = '\n'.join(formatted_reasons)
+    return self._reason
+
+  @property
+  def changed_files(self):
+    """Merge multiple``FeatureValue.changed_files`` lists into one.
+
+    Returns:
+      A list of ``ChangedFile`` objects sorted by file name. The sorting
+      is not essential, but is provided to ease testing by ensuring the
+      output is in some canonical order.
+
+    Raises:
+      ``ValueError`` if any file name is given inconsistent ``blame_url``s.
+    """
+    if self._changed_files:
+      return self._changed_files
+
+    all_changed_files = {}
+    for feature in self.itervalues():
+      if not feature.changed_files:
+        continue
+
+      for changed_file in feature.changed_files or []:
+        accumulated_changed_file = all_changed_files.get(changed_file.name)
+        if accumulated_changed_file is None:
+          all_changed_files[changed_file.name] = changed_file
+          continue
+
+        if (accumulated_changed_file.blame_url !=
+            changed_file.blame_url): # pragma: no cover
+          raise ValueError('Blame URLs do not match: %s != %s'
+              % (accumulated_changed_file.blame_url, changed_file.blame_url))
+        accumulated_changed_file.reasons.extend(changed_file.reasons or [])
 
-class Feature(object):
+    self._changed_files = all_changed_files.values()
+    self._changed_files.sort(key=lambda changed_file: changed_file.name)
+    return self._changed_files
+
+  def __len__(self):
+    return len(self._value)
+
+  def __eq__(self, other):
+    return (self.name == other.name and self._value == other._value and
+            self.reason == other.reason and
+            self.changed_files == other.changed_files)
+
+  def __ne__(self, other):
+    return not self.__eq__(other)
+
+
+class Feature(Element):
   """Abstract base class for features use by loglinear models."""
 
   @property
@@ -120,8 +268,8 @@ class Feature(object):
     ``X`` and ``Y``, as described in the documentation there. As an
     example: for the CL classifier, ``X`` is ``CrashReport`` and ``Y`` is
     ``Suspect``. Given those two types, this method is a curried function
-    of type ``X -> Y -> FeatureValue``. That is, given some ``x`` of type
-    ``X``, we return a function of type ``Y -> FeatureValue``, where
+    of type ``X -> Y -> FeatureValue``. That is, given some ``x`` of
+    type ``X``, we return a function of type ``Y -> FeatureValue``, where
     the final result for each ``y`` of type ``Y`` is the value of that
     ``y`` given that ``x``.
 
@@ -140,25 +288,74 @@ class Feature(object):
     raise NotImplementedError()
 
 
-class FeatureFunction(object):
-  """Given an iterable of scalar-valued functions, return an dict function.
+class MetaFeature(MetaDict):
+  """Abstract base class for meta features use by loglinear models.
+
+  MetaFeature is a dict of (Meta)Features.
+  """
+
+  @property
+  def name(self):
+    """The name of this feature."""
+    raise NotImplementedError()
+
+  def __call__(self, x):
+    """Returns a value for a ``y`` given some ``x``.
+
+    The loglinear model this feature is used in will specify some types
+    ``X`` and ``Y``, as described in the documentation there. As an
+    example: for the CL classifier, ``X`` is ``CrashReport`` and ``Y`` is
+    ``Suspect``. Given those two types, this method is a curried function
+    of type ``X -> Y -> FeatureValue``. That is, given some ``x`` of
+    type ``X``, we return a function of type ``Y -> FeatureValue``, where
+    the final result for each ``y`` of type ``Y`` is the value of that
+    ``y`` given that ``x``.
+
+    Values closer to zero indicate this feature has less to say about
+    whether the ``y`` is to be blamed. Values further from zero indicate
+    that this feature has more to say about it. (Whether this feature
+    thinks the ``y`` should be blamed or should not be depends on the sign
+    of the value and the sign of the weight given to this feature.) As
+    special cases, a value of negative infinity means "do not blame this
+    ``y`` no matter what any other features say", and a value of positive
+    infinity means "definitely blame this ``y`` no matter what any other
+    features say". Both of those special values should be used sparingly,
+    since they override the model's ability to combine multiple sources of
+    information and decide the cuplrit based on all the evidence together.
+    """
+    raise NotImplementedError()
+
+
+class WrapperMetaFeature(MetaFeature):
+  """Given a dict of scalar-valued functions, return an dict-valued function.
+
+  Note, the features that get wrapped should be independent to each other, which
+  means their feature values can be computed independently.
+
+  Either wrap single Feature or wrap features whose final results are computed
+  independently.
 
   Properties:
-    fs (iterable of functions): A collection of curried functions
-    ``X -> Y -> FeatureValue``. That is, given a particular ``x`` they
-    return a function ``Y -> dict(FeatureValue)``. N.B. each function should
-    have a name property.
+    fs (Feature of iterable of (Meta)Features): A collection of curried
+    functions ``X -> Y -> (Meta)FeatureValue``. That is, given a particular
+    ``x`` they return a function ``Y -> dict(FeatureValue)``. N.B. each function
+    should have a name property.
   """
   def __init__(self, fs):
-    self._fs = fs
+    super(WrapperMetaFeature, self).__init__({f.name: f for f in fs or []})
+
+  @property
+  def name(self):
+    return 'WrapperFeature'
 
   def __call__(self, x):
     """Fuction mapping ``X -> Y -> dict(FeatureValue.name to FeatureValue).
 
     Returns:
-      A function ``X -> Y -> dict(FeatureValue.name to FeatureValue)`` where for 
+      A function ``X -> Y -> dict(FeatureValue.name to FeatureValue)`` where for
       all ``x``, ``y``, and for a feature f in fs, we have
       ``FeatureFunction(fs)(x)(y)[f.name] == f(x)(y)``.
     """
-    name_to_fx = {f.name: f(x) for f in self._fs}
-    return lambda y: {name: fx(y) for name, fx in name_to_fx.iteritems()}
+    fxs = {name: f(x) for name, f in self.iteritems()}
+    return lambda y: MetaFeatureValue(
+        self.name, {name: fx(y) for name, fx in fxs.iteritems()})