1.7x performance improvement

experiments/performance.ipynb

@@ -15,7 +15,7 @@
  },
  {
  "cell_type": "code",
- "execution_count": 4,
+ "execution_count": 2,
  "id": "0b07fa87-3c3c-40d4-bca7-605cfe5052c9",
  "metadata": {},
  "outputs": [],
@@ -37,16 +37,16 @@
  "name": "stdout",
  "output_type": "stream",
  "text": [
- "0.9599\n",
- "CPU times: user 163 ms, sys: 96.5 ms, total: 259 ms\n",
- "Wall time: 14min 11s\n"
+ "0.9604\n",
+ "CPU times: user 136 ms, sys: 119 ms, total: 255 ms\n",
+ "Wall time: 17 s\n"
  ]
  }
  ],
  "source": [
  "%%time\n",
  "clf = ProbabilisticScoringList([-1, 1, 2])\n",
- "print(f\"{cross_val_score(clf, X, y, fit_params=dict(l=2), cv=ShuffleSplit(100, test_size=.2, random_state=42), n_jobs=-1).mean():.4f}\")"
+ "print(f\"{cross_val_score(clf, X, y, fit_params=dict(l=1), cv=ShuffleSplit(100, test_size=.2, random_state=42), n_jobs=-1).mean():.4f}\")"
  ]
  },
  {
@@ -65,9 +65,9 @@
  ],
  "metadata": {
  "kernelspec": {
- "display_name": "Python 3 (ipykernel)",
+ "display_name": "skpsl",
  "language": "python",
- "name": "python3"
+ "name": "skpsl"
  },
  "language_info": {
  "codemirror_mode": {

skpsl/probabilistic_scoring_list.py

@@ -20,35 +20,29 @@ def __init__(self, features, scores):
 
  self.logger = logging.getLogger(__name__)
  self.scores_vec = np.array(scores)
- self.score_sums = {}
  self.probabilities = {}
  self.entropy = None
- self.calibrator = None
 
  def fit(self, X, y) -> "_ClassifierAtK":
- n = X.shape[0]
- relevant_scores = self._relevant_scores(X)
+ scores = self._scores_per_record(X)
+ n = scores.size
 
  # compute all possible total scores using subset-summation
- self.score_sums = {0}
+ total_scores = {0}
  for score in self.scores_vec:
- self.score_sums |= {prev_sum + score for prev_sum in self.score_sums}
+ total_scores |= {prev_sum + score for prev_sum in total_scores}
+ total_scores = np.array(sorted(total_scores))
 
- # calibrate probabilities
- self.calibrator = IsotonicRegression(y_min=0.0, y_max=1.0, increasing=True, out_of_bounds="clip")
- self.calibrator.fit(relevant_scores, y)
+ # compute probabilities
+ calibrator = IsotonicRegression(y_min=0.0, y_max=1.0, increasing=True, out_of_bounds="clip")
+ calibrator.fit(scores, y)
+ self.probabilities = {T: p for T, p in zip(total_scores, calibrator.transform(total_scores))}
 
- # compute calibrated probabilities
- sigmaK = np.array(sorted(self.score_sums))
- cal_ps = self.calibrator.predict(sigmaK)
-
- # set calibrated probabilities
- self.probabilities = {T: p for T, p in zip(sigmaK, cal_ps)}
- self.entropy = 0
- for ti, pi in self.probabilities.items():
- Ni = np.count_nonzero(relevant_scores == ti)
- Hi = stats_entropy([pi, 1 - pi], base=2)
- self.entropy += (Ni / n) * Hi
+ # TODO. this should actually be inside of a score function. the actual fitting is finished at this point
+ total_scores, score_freqs = np.unique(scores, return_counts=True)
+ score_probas = np.array([self.probabilities[ti] for ti in total_scores])
+ entropy_values = stats_entropy([score_probas, 1 - score_probas], base=2)
+ self.entropy = np.sum((score_freqs / n) * entropy_values)
 
  return self
 
@@ -58,13 +52,16 @@ def predict(self, X):
  def predict_proba(self, X):
  """Predicts the probability for
  """
- proba_true = np.vectorize(self.probabilities.get)(self._relevant_scores(X))
+ scores = self._scores_per_record(X)
+ proba_true = np.empty_like(scores, dtype=float)
+ for total_score in np.unique(scores):
+ proba_true[scores == total_score] = self.probabilities[total_score]
  proba = np.vstack([1 - proba_true, proba_true]).T
  return proba
 
  # Helper functions
- def _relevant_scores(self, X):
- return np.sum(X[:, self.features] * self.scores_vec, axis=1)
+ def _scores_per_record(self, X):
+ return X[:, self.features] @ self.scores_vec
 
 
 class ProbabilisticScoringList(BaseEstimator, ClassifierMixin):