fix: stop T-norm fit from exploding in memory.

Removed the unnecessary copy of `data` and storage into a Sample object.
Now saves the stats in a dict that is picklable.

src/bob/bio/base/pipelines/score_post_processor.py

@@ -296,6 +296,11 @@ class TNormScores(TransformerMixin, BaseEstimator):
     Parameters
     ----------
 
+
+    Attributes
+    ----------
+
+    t_stats: a dictionary keeping the ``mu`` and ``std`` of each identity.
     """
 
     post_process_template = "enroll"
@@ -309,23 +314,12 @@ class TNormScores(TransformerMixin, BaseEstimator):
         super().__init__(**kwargs)
         self.top_norm = top_norm
         self.top_norm_score_fraction = top_norm_score_fraction
+        self.t_stats: dict[str, dict[str, float]] = dict()
 
     def fit(self, t_scores, y=None):
-        # TODO: THIS IS SUPER INNEFICIENT, BUT
-        # IT'S THE MOST READABLE SOLUTION
-        # Stacking scores by biometric reference
-        self.t_stats = dict()
-
+        """Computes the mean and std of each SampleSet/identity"""
         for sset in t_scores:
-            self.t_stats[sset.template_id] = Sample(
-                [s.data for s in sset], parent=sset
-            )
-
-        # Now computing the statistics in place
-        for key in self.t_stats:
-            data = self.t_stats[key].data
-
-            # Selecting the top scores
+            data = [s.data for s in sset]
             if self.top_norm:
                 # Sorting in ascending order
                 data = -np.sort(-data)
@@ -333,13 +327,10 @@ class TNormScores(TransformerMixin, BaseEstimator):
                     np.floor(len(data) * self.top_norm_score_fraction)
                 )
                 data = data[0:proportion]
-
-            self.t_stats[key].mu = np.mean(self.t_stats[key].data)
-            self.t_stats[key].std = np.std(self.t_stats[key].data)
-            # self._z_stats[key].std = legacy_std(
-            #    self._z_stats[key].mu, self._z_stats[key].data
-            # )
-            self.t_stats[key].data = []
+            self.t_stats[sset.template_id] = {
+                "mu": np.mean(data),
+                "std": np.std(data),
+            }
 
         return self
 
@@ -351,7 +342,7 @@ class TNormScores(TransformerMixin, BaseEstimator):
         def _transform_samples(X, stats):
             scores = []
             for no_normed_score in X:
-                score = (no_normed_score.data - stats.mu) / stats.std
+                score = (no_normed_score.data - stats["mu"]) / stats["std"]
 
                 t_score = Sample(score, parent=no_normed_score)
                 scores.append(t_score)
@@ -372,7 +363,7 @@ class TNormScores(TransformerMixin, BaseEstimator):
                 )
         else:
             # If it is Samples
-            t_normed_scores = _transform_samples(X)
+            t_normed_scores = _transform_samples(X)  # YD2023: Is this needed?
 
         return t_normed_scores