UPD: Refacto notebook

scikit-learn-contrib · Jul 11, 2023 · 2c2244f · 2c2244f
1 parent 59db4ff
commit 2c2244f
Show file tree

Hide file tree

Showing 2 changed files with 339 additions and 993 deletions.
diff --git a/mapie/classification.py b/mapie/classification.py
@@ -610,13 +610,8 @@ def _get_true_label_cumsum_proba(
  y_pred_proba_sorted = np.take_along_axis(y_pred_proba, index_sorted, axis=1)
  y_true_sorted = np.take_along_axis(y_true, index_sorted, axis=1)
  if self.method == "ssaps":
- j_values = np.arange(1, len(self.classes_) + 1)
- penalized_residuals = 1 - np.power(residuals[:, np.newaxis], j_values)
- y_pred_proba_penalized_sorted = np.multiply(
- penalized_residuals, y_pred_proba_sorted
- )
- y_pred_proba_sorted_cumsum = np.cumsum(
- y_pred_proba_penalized_sorted, axis=1
+ y_pred_proba_sorted_cumsum = self._get_penalized_ssaps(
+ y_pred_proba_sorted, residuals
  )
  else:
  y_pred_proba_sorted_cumsum = np.cumsum(y_pred_proba_sorted, axis=1)
@@ -684,6 +679,24 @@ def _get_true_label_position(self, y_pred_proba: NDArray, y: NDArray) -> NDArray
 
  return position
 
+ def _get_penalized_ssaps(self, y_pred_proba_sorted, residuals, alpha_dim=False):
+ if alpha_dim:
+ n_alphas = y_pred_proba_sorted.shape[-1]
+ y_pred_proba_sorted = y_pred_proba_sorted[:, :, 0]
+ j_values = np.arange(1, len(self.classes_) + 1)
+ penalized_residuals = 1 - np.power(residuals[:, np.newaxis], j_values - 1)
+ penalized_residuals[:, 0] = 1.0
+ y_pred_proba_penalized_sorted = np.multiply(
+ penalized_residuals, y_pred_proba_sorted
+ )
+ y_pred_proba_sorted_cumsum = np.cumsum(y_pred_proba_penalized_sorted, axis=1)
+ if alpha_dim:
+ y_pred_proba_sorted_cumsum = y_pred_proba_sorted_cumsum[:, :, np.newaxis]
+ y_pred_proba_sorted_cumsum = np.repeat(
+ y_pred_proba_sorted_cumsum, repeats=n_alphas, axis=2
+ )
+ return y_pred_proba_sorted_cumsum
+
  def _get_last_included_proba(
  self,
  y_pred_proba: NDArray,
@@ -727,13 +740,10 @@ def _get_last_included_proba(
  y_pred_proba_sorted = np.take_along_axis(y_pred_proba, index_sorted, axis=1)
 
  if self.method == "ssaps":
- j_values = np.arange(1, len(self.classes_) + 1)
- penalized_residuals = 1 - np.power(residuals[:, np.newaxis], j_values)
- y_pred_proba_penalized_sorted = np.multiply(
- penalized_residuals[:, :, np.newaxis], y_pred_proba_sorted
- )
- y_pred_proba_sorted_cumsum = np.cumsum(
- y_pred_proba_penalized_sorted, axis=1
+ y_pred_proba_sorted_cumsum = self._get_penalized_ssaps(
+ y_pred_proba_sorted,
+ residuals,
+ alpha_dim=True,
  )
  else:
  # get sorted cumulated score

diff --git a/notebooks/ImageNet-crf.ipynb b/notebooks/ImageNet-crf.ipynb