Move statistics update to statistics class

frechet_audio_distance/frechet_audio_distance.py

@@ -44,9 +44,37 @@ def __init__(self, dim: int):
                 tf.zeros((dim, dim), dtype=tf.float64)
             )
             self.mean = tf.Variable(tf.zeros((dim,), dtype=tf.float64))
+            self.num_items_processed = tf.Variable(0.0, dtype=tf.float64)
+
+        def update(self, data: tf.Tensor) -> None:
+            """
+            Updates the means and covariances held by an instance of this class
+            """
+            data = tf.cast(data, dtype=tf.float64)
+            num_items_this_update = tf.cast(
+                tf.shape(data)[0], dtype=tf.float64
+            )
+            self.num_items_processed.assign_add(
+                tf.cast(num_items_this_update, dtype=tf.float64)
+            )
 
-    _statistics_dim: int = VGGishParams.EMBEDDING_SIZE
-    _num_items_processed: tf.Variable  # dtype = tf.float64
+            x_norm_old = data - self.mean
+            self.mean.assign_add(
+                tf.reduce_sum(x_norm_old, axis=0) / self.num_items_processed
+            )
+            x_norm_new = data - self.mean
+
+            self.covariance.assign(
+                self.covariance
+                * (self.num_items_processed - num_items_this_update)
+                / self.num_items_processed
+            )
+            self.covariance.assign_add(
+                tf.matmul(tf.transpose(x_norm_old), x_norm_new)
+                / self.num_items_processed
+            )
+
+    _statistics_dim: int = VGGish.embedding_size
     _true_statistics: _Statistics
     _pred_statistics: _Statistics
 
@@ -99,22 +127,16 @@ def update_state(
             y_true_embedding, y_pred_embedding = map(
                 self._vggish_model, [y_true_features, y_pred_features]
             )
-            self._num_items_processed.assign_add(
-                tf.cast(tf.shape(y_true_embedding)[0], dtype=tf.float64)
-            )
             for data, statistics in zip(
                 [y_true_embedding, y_pred_embedding],
                 [self._true_statistics, self._pred_statistics],
             ):
-                self._update_statistics(
-                    statistics, data, self._num_items_processed
-                )
+                statistics.update(data)
 
     def result(self) -> tf.Tensor:  # pragma: no cover
         return self._compute_distance()
 
     def reset_state(self) -> None:
-        self._num_items_processed = tf.Variable(0, dtype=tf.float64)
         self._true_statistics = self._Statistics(self._statistics_dim)
         self._pred_statistics = self._Statistics(self._statistics_dim)