Implementing loglinear classification (without training), for CL classification

appengine/findit/crash/test/loglinear_test.py

Index: appengine/findit/crash/test/loglinear_test.py
diff --git a/appengine/findit/crash/test/loglinear_test.py b/appengine/findit/crash/test/loglinear_test.py
new file mode 100644
index 0000000000000000000000000000000000000000..7b90d8f02e4a82ec4b437fdb182e8ebb99e9a299
--- /dev/null
+++ b/appengine/findit/crash/test/loglinear_test.py
@@ -0,0 +1,65 @@
+# Copyright 2016 The Chromium Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style license that can be
+# found in the LICENSE file.
+
+import math
+import numpy as np
+import random
+import unittest
+
+from crash.loglinear import ToFeatureFunction
+from crash.loglinear import LogLinearModel
+
+
+# Some arbitrary features.
+# We don't use double lambdas because gpylint complains about that.
+def feature0(x):
+  return lambda y: 1.0 if y == (x > 5) else 0.0
+
+
+def feature1(x):
+  return lambda y: 1.0 if y == ((x % 2) == 1) else 0.0
+
+
+def feature2(x):
+  return lambda y: 1.0 if y == (x <= 7) else 0.0
+
+
+features = [feature0, feature1, feature2]
+X = range(10)
+Y = [False, True]
+
+
+class LoglinearTest(unittest.TestCase):
+
+  def testToFeatureFunction(self):
+    """Test that ``ToFeatureFunction`` obeys the equality its docstring says."""
+    f = ToFeatureFunction(features)
+    for x in X:
+      for y in Y:
+        for i in xrange(len(features)):
+          self.assertEqual(features[i](x)(y), f(x)(y)[i])
+
+  # TODO(wrengr): how to make this test something reasonable?

[inferno, 2016/12/06 18:07:06]

rephrase this: Make this test reasonable in terms of accuracy because ...

+  def testLogLinearModel(self):
+    """An arbitrary test to get 100% code coverage."""
+    weights = [random.random() for _ in features]
+
+    model = LogLinearModel(Y, ToFeatureFunction(features), weights, 0.1)
+    model.ClearAllMemos()
+    model = LogLinearModel(Y, ToFeatureFunction(features), weights)
+    self.assertListEqual(weights, model.weights.tolist())
+    # TODO(wrengr): this may be flaky due to floating point fuzz.

[inferno, 2016/12/06 18:07:06]

Just one todo for the next three is better

# TODO(wrengr): In this block, condition matching might be flaky due to the
floating point fuzz.

+    self.assertEqual(math.sqrt(model.quadrance), model.l2)
+
+    for x in X:
+      # TODO(wrengr): this may be flaky due to floating point fuzz.
+      self.assertEqual(math.exp(model.logZ(x)), model.Z(x))
+      model.Expectation(x, lambda y: np.array([1.0]))
+      for y in Y:
+        model.Features(x)(y)
+        model.Score(x)(y)
+        # TODO(wrengr): this may be flaky due to floating point fuzz.
+        self.assertEqual(
+            math.exp(model.LogProbability(x)(y)),
+            model.Probability(x)(y))