[Predator] Fix weight flaky loglinear training test.

appengine/findit/crash/loglinear/weight.py

Index: appengine/findit/crash/loglinear/weight.py
diff --git a/appengine/findit/crash/loglinear/weight.py b/appengine/findit/crash/loglinear/weight.py
index a1fb4ef98894d86cec1ae393f6fabd6accc8b933..5b0f9be8192d087daddc7bbab936613245276fa8 100644
--- a/appengine/findit/crash/loglinear/weight.py
+++ b/appengine/findit/crash/loglinear/weight.py
@@ -69,24 +69,28 @@ class MetaWeight(MetaDict):
   """Dict-like class mapping features in ``Metafeature`` to their weights."""
 
   def IsZero(self, epsilon):
-    """A MetaWeight is zero only when all the sub weights are zeros."""
+    if not self._value:
+      return True
+
     return all(weight.IsZero(epsilon) for weight in self.itervalues())
 
   def DropZeroWeights(self, epsilon=0.):
     """Drops all zero weights."""
+    # Make all sub meta weights drop their zero weights.
+    for weight in self.itervalues():
+      if not weight.is_element:
+        weight.DropZeroWeights(epsilon=epsilon)
+
     self._value = {name: weight for name, weight in self.iteritems()
-                  if not weight.IsZero(epsilon)}
+                   if not weight.IsZero(epsilon)}
 
   def __len__(self):
     return len(self._value)
 
   def __mul__(self, meta_feature):
     """``MetaWeight`` can multiply with ``MetaFeatureValue``."""
-    assert len(self) == len(meta_feature), Exception(

[Sharu Jiang, 2017/03/14 05:06:13]

This assertion should be deleted. Because the ``DropZeroWeights`` will drop all
zeros in meta weight, which will make the length of the ``meta_weight``
different from the length of ``meta_feature``.

In
https://cs.chromium.org/chromium/infra/appengine/findit/crash/loglinear/test/...,
line 55.

If the trained weight happens to be 0 (the training set is randomly generated,
so the training results vary every time), the length of meta weight would be 0,
however the length of meta feature is still 1, the assertion will fail.

-        'MetaWeight can only multiply with ``MetaFeatureValue`` with the '
-        'same length')
-
-    # MetaWeight is a dense representation of a sparse array.
+    # MetaWeight is a dense representation of a sparse array. So MetaWeight and
+    # MetaFeature don't necessarily have the same length.
     return math.fsum(meta_feature[name] * weight
                      for name, weight in self.iteritems())