[Predator] Move ``SingleFeatureScore`` to LLM.

appengine/findit/crash/loglinear/training.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 math
 import numpy as np
 # N.B., ``np.array`` can't take generators; you must pass explicit lists.
 import scipy.optimize as spo
 
 from crash.loglinear.model import LogLinearModel
@@ skipping 15 matching lines @@
         ``Y`` (it'll just take more computation time is all). This is
         needed for computing the partition function and expectation. N.B.,
         we do not actually need to know/enumerate of *all* of ``Y``,
         only the subsets for each ``x``.
       training_data (iterable): a collection of ``(x, y)`` pairs where
         ``y`` is the known-correct label for ``x``.
       feature_function: A function from ``X`` to ``Y`` to a list of
         ``float``. N.B., the length of the list must be the same for all
         ``x`` and ``y``, and must be the same as the length of the list
         of weights.
-      initial_weights (list of float): the pre-training coefficients
+      initial_weights (dict from str to float): the pre-training coefficients
         for how much we believe components of the feature vector. This
         provides the seed for training; this starting value shouldn't
         affect the final weights obtained by training (thanks to
         convexity), but will affect how long it takes for training
         to converge.
       epsilon (float): The absolute-error threshold for considering a
         weight to be "equal to zero". N.B., this should be a positive
         number, as we will compare it against the absolute value of
         each weight.
     """
     super(TrainableLogLinearModel, self).__init__(
         Y_given_X, feature_function, initial_weights, epsilon)
     self._training_data = training_data
+    self._feature_order = initial_weights.keys()

[wrengr, 2017/01/11 20:38:30]

You should explicitly convert this to a list (or tuple) since those have a
specified order, rather than leaving it as a set which doesn't.

[Sharu Jiang, 2017/01/12 01:41:38]

I think the ``.keys()`` will return a list, but explicitly converting to a list
should be safer. Done.

[wrengr, 2017/01/12 18:16:16]

So it does. Weird.

 
     self._observed_feature_vector = vsum([
         self.FeaturesAsNumPyArray(x)(y)
         for x, y in self._training_data])
 
   # Even though this is identical to the superclass definition, we must
   # re-provide it in order to define the setter.
   @property
   def weights(self):
     """The weight covector.
 
     At present we return the weights as an ``np.ndarray``, but in the
     future that may be replaced by a more general type which specifies
     the semantics rather than the implementation details.
     """
     return self._weights
 
   @weights.setter
-  def weights(self, new_weights): # pylint: disable=W0221
+  def weights(self, new_np_weights): # pylint: disable=W0221
     """Mutate the weight covector, and clear memos as necessary.
 
     This setter attempts to avoid clearing memos whenever possible,
     but errs on the side of caution/correctness when it needs to.
 
     Args:
-      new_weights (np.ndarray): the new weights to use. Must have the
-        same shape as the old ``np.ndarray``.
+      new_np_weights (np.ndarray): the new weights to use. It will be converted
+      to weights dict mapping feature_name to its weight.
     """
-    if new_weights is self._weights:
-      return
-
-    if not isinstance(new_weights, np.ndarray):
-      raise TypeError('Expected an np.ndarray but got %s instead'
-          % new_weights.__class__.__name__)
-
-    if new_weights.shape != self._weights.shape:
-      raise TypeError('Weight shape mismatch: %s != %s'
-          % (new_weights.shape, self._weights.shape))
-
+    new_weights = self.NumPyArrayToWeights(new_np_weights)
     self.ClearWeightBasedMemos()
     self._weights = new_weights

[wrengr, 2017/01/11 20:38:31]

To avoid excessive conversion back and forth, this class should store both the
dict and np.ndarray. That is, I suggest:

    def __init__(self,...):
        ...
        self._weights = initial_weights
        self._np_weights = self._DictToNpArray(initial_weights, 0.)

    @property
    def weights(self):
        return self._weights

    @weights.setter
    def weights(self, weights_dict):
        self.ClearWeightBasedMemos()
        self._weights = weights_dict
        self._np_weights = self._DictToNpArray(weights_dict, 0.)

    @property
    def np_weights(self):
        return self._np_weights

    @np_weights.setter
    def np_weights(self, weights_array):
        self.ClearWeightBasedMemos()
        self._weights = self._NpArrayToDict(weights_array, self._LogZeroish)
        self._np_weights = weights_array

    def _DictToNpArray(self, d, default=None):
        return np.array([d.get(i, default) for i in self._feature_order])

    def _NpArrayToDict(self, arr, predicate=None):
        if predicate is None:
          predicate = lambda _: True
        return {k: v for k, v in zip(self._feature_order, arr) if predicate(v)}

[Sharu Jiang, 2017/01/12 01:41:38]

Done.

 
   def FeaturesAsNumPyArray(self, x):
     """A variant of ``Features`` which returns a ``np.ndarray``.
 
+    Note, the features nparray should have the same order as in
+    self._feature_order to stay aligned with weights np array.
+
     For training we need to have the feature function return an
     ``np.ndarray(float)`` rather than the ``list(FeatureValue)`` used
     elsewhere. This function performes the necessary conversion.
 
     N.B., at present we do not memoize this function. The underlying
     ``Features`` method is memoized, so we won't re-compute the features
     each time; but we will repeatedly copy the floats into newly allocated
     ``np.ndarray`` objects. If that turns out to be a performance
     bottleneck, we can add the extra layer of memoization to avoid that.
     """
     fx = self.Features(x)
-    return lambda y: np.array([fxy.value for fxy in fx(y)])
+    def FeaturesAsNumPyArrayGivenX(y):
+      fxys = fx(y)
+      return np.array([fxys[feature_name].value if feature_name in fxys else 0.

[wrengr, 2017/01/11 20:38:30]

is clearer to use ``fxys.get(feature_name, 0.)``

[Sharu Jiang, 2017/01/12 01:41:38]

I have to do this, because I need ``fxys[feature_name].value`` instead of
``fxys[feature_name]``

[wrengr, 2017/01/12 18:16:16]

hrm... that does make it tricky.

This sort of thing is why I had FeatureValue.__float__ previously; with that we'
could've ``float(fxys.get(feature_name, 0.))`` without needing to wrap ``0.`` up
in something that provides a ``.value`` property. I really wish Python had
let-bindings in its comprehension syntax. One alternative is to build an
intermediate generator, thus: ``return np.array([0. if fxyn is None else
fxyn.value for fxyn in (fxys.get(n) for n in self._feature_order)])``. Not sure
that that's any cleaner though :(

+                       for feature_name in self._feature_order])
+
+    return FeaturesAsNumPyArrayGivenX
+
+  def WeightsAsNumPyArray(self):
+    """Converts dict (mapping feature name to weight) to numpy array.
+
+    Note, this conversion is needed because model uses weights dict to organize
+    weights for features, however SciPy trainning (e.g. BFGS) needs numpy array
+    to do computaion.
+    """
+    return np.array([self._weights.get(feature_name, 0.)
+                     for feature_name in self._feature_order])

[wrengr, 2017/01/11 20:38:31]

Again, to avoid excessive conversion, we can do this once in the ``weights``
setter and then this method simply returns the cached np.array

[Sharu Jiang, 2017/01/12 01:41:38]

Done.

+
+  def NumPyArrayToWeights(self, np_weights):
+    """Converts numpy array to dict (mapping feature name to weight).
+
+    Note, this conversion is needed because model uses weights dict to organize
+    weights for features, however SciPy trainning (e.g. BFGS) needs numpy array
+    to do computaion.
+
+    Args:
+      np_weights (np.ndarray): Weights which have the same order of
+        self._feature_order. Note, featuer np array should also be serialized by
+        the same order as self._feature_order to match.
+
+    Returns:
+      A dict mapping feature name to weight.
+    """
+    if not isinstance(np_weights, np.ndarray):
+      raise TypeError('Expected an np.ndarray but got %s instead'
+          % np_weights.__class__.__name__)
+
+    return {feature_name: weight
+            for feature_name, weight in zip(self._feature_order, np_weights)}
 
   def LogLikelihood(self):
     """The conditional log-likelihood of the training data.
 
     The conditional likelihood of the training data is the product
     of ``Pr(y|x)`` for each ``(x, y)`` pair in the training data; so
     the conditional log-likelihood is the log of that. This is called
     "likelihood" because it is thought of as a function of the weight
     covector, with the training data held fixed.
 
     This is the ideal objective function for training the weights, as it
     will give us the MLE weight covector for the training data. However,
     in practice, we want to do regularization to ensure we don't overfit
     the training data and to reduce classification time by ensuring that
     the weight vector is sparse. Thus, the actual objective function
     will be the log-likelihood plus some penalty terms for regularization.
     """
     observed_zeta = math.fsum(self.LogZ(x) for x, _ in self._training_data)
-    observed_score = self.weights.dot(self._observed_feature_vector)
+    observed_score = self.WeightsAsNumPyArray().dot(
+        self._observed_feature_vector)
     return observed_score - observed_zeta
 
   def LogLikelihoodGradient(self):
     """The gradient (aka Jacobian) of ``LogLikelihood``."""
     expected_feature_vector = vsum([
         self.Expectation(x, self.FeaturesAsNumPyArray(x))
         for x, _ in self._training_data])
     return self._observed_feature_vector - expected_feature_vector
 
   def TrainWeights(self, l2_penalty):
     """Optimize the weight covector based on the training data.
 
     Args:
       l2_penalty (float): the hyperparameter for how much to penalize
         weight covectors far from zero.
 
     Returns:
       Nothing, but has the side effect of mutating the stored weights.
     """
-    initial_weights = self.weights
+    initial_weights = self.WeightsAsNumPyArray()
 
     # We want to minimize the number of times we reset the weights since
     # that clears our memos. One might think we could do that in the
     # between-iterations callback; but actually, in a single iteration,
     # BFGS calls the objective function and gradient more than once with
     # different arguments; so, alas, we must reset the weights in both.
     # This is why the ``weights`` setter tries to avoid clearing memos
     # when possible.
 
     def objective_function(new_weights):
       self.weights = new_weights
       return -self.LogLikelihood() + 0.5 * l2_penalty * self.quadrance
 
     def objective_function_gradient(new_weights):
       self.weights = new_weights
-      return -self.LogLikelihoodGradient() + l2_penalty * self.weights
+      return -self.LogLikelihoodGradient() + l2_penalty * new_weights

[wrengr, 2017/01/11 20:38:31]

I don't think that's any cleaner than the old code

[Sharu Jiang, 2017/01/12 01:41:38]

This is because we need np_array here instead of a dict.

[wrengr, 2017/01/12 18:16:16]

Ah, right <chagrin>

 
     result = spo.minimize(
         objective_function,
         initial_weights,
         method='BFGS',
         jac=objective_function_gradient)
 
     if not result.success: # pragma: no cover
       # This should happen infrequently enough that there's no point in
       # logging it and attempting to carry on.
       raise Exception(
           'TrainableLogLinearModel.TrainWeights failed:'
           '\n\tReason: %s'
           '\n\tCurrent objective value: %s'
           '\n\tCurrent objective gradient: %s'
           '\n\tIterations: %d'
           '\n\tFunction evaluations: %d'
           '\n\tGradient evaluations: %d'
           % (result.message, result.fun, result.jac, result.nit, result.nfev,
              result.njev))
 
     # This shouldn't really be necessary, since we're resetting it
     # directly during training; but just to be safe/sure.
     self.weights = result.x