[Predator] Implement training for loglinear models

appengine/findit/crash/loglinear/test/training_test.py

+# 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 unittest
+import random
+import numpy as np
+
+import crash.loglinear.test.model_test as loglinear_test
+from crash.loglinear.model import ToFeatureFunction
+from crash.loglinear.training import TrainableLogLinearModel
+
+
+training_data = [(x, x == 7) for x in xrange(10)]
+
+
+class TrainableLogLinearModelTest(unittest.TestCase):
+
+  def setUp(self):
+    super(TrainableLogLinearModelTest, self).setUp()
+    feature_function = ToFeatureFunction(loglinear_test.features)
+    initial_weights = [random.random() for _ in loglinear_test.features]
+    self.model = TrainableLogLinearModel(
+        loglinear_test.Y, training_data, feature_function, initial_weights)
+
+  def testWeightsSetterNotAnNdarray(self):
+    def _WeightSettingExpression():
+      """Wrap the ``self.model.weights = stuff`` expression.
+
+      The ``assertRaises`` method expects a callable object, so we need
+      to wrap the expression in a def. If we didn't wrap it in a def
+      then we'd throw the exception too early, and ``assertRaises``
+      would never get called in order to see it. Normally we'd use a
+      lambda for wrapping the expression up, but because the expression
+      we want to check is actually a statement it can't be in a lambda
+      but rather must be in a def.
+      """
+      self.model.weights = 'this is not an np.ndarray'
+
+    self.assertRaises(TypeError, _WeightSettingExpression)
+
+  def testWeightsSetterShapeMismatch(self):
+    def _WeightSettingExpression():
+      """Wrap the ``self.model.weights = stuff`` expression."""
+      # This np.ndarray has the wrong shape.
+      self.model.weights = np.array([[1,2], [3,4]])
+
+    self.assertRaises(TypeError, _WeightSettingExpression)
+
+  def testTrainWeights(self):
+    """Tests that ``TrainWeights`` actually improves the loglikelihood.
+
+    Actually, this is more of a test that we're calling SciPy's BFGS
+    implementation correctly. But any bugs we find about that will show
+    up in trying to run this rest rather than in the assertaion failing
+    per se.
+    """
+    initial_loglikelihood = self.model.LogLikelihood()
+    self.model.TrainWeights(0.5)
+    trained_loglikelihood = self.model.LogLikelihood()
+    self.assertTrue(trained_loglikelihood >= initial_loglikelihood,
+        'Training reduced the loglikelihood from %f to %f,'
+        ' when it should have increased it!'
+        % (initial_loglikelihood, trained_loglikelihood))