stouputils 1.3.6__tar.gz → 1.3.7__tar.gz

{stouputils-1.3.6 → stouputils-1.3.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: stouputils
-Version: 1.3.6
+Version: 1.3.7
 Summary: Stouputils is a collection of utility modules designed to simplify and enhance the development process. It includes a range of tools for tasks such as execution of doctests, display utilities, decorators, as well as context managers, and many more.
 Project-URL: Homepage, https://github.com/Stoupy51/stouputils
 Project-URL: Issues, https://github.com/Stoupy51/stouputils/issues

{stouputils-1.3.6 → stouputils-1.3.7}/pyproject.toml

@@ -5,7 +5,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "stouputils"
-version = "1.3.6"
+version = "1.3.7"
 description = "Stouputils is a collection of utility modules designed to simplify and enhance the development process. It includes a range of tools for tasks such as execution of doctests, display utilities, decorators, as well as context managers, and many more."
 readme = "README.md"
 requires-python = ">=3.10"

{stouputils-1.3.6 → stouputils-1.3.7}/stouputils/data_science/metric_dictionnary.py

@@ -8,12 +8,16 @@ This class contains the following metrics:
 1. Main metrics:
 
    - Area Under the Curve (AUC)
+   - Area Under the Precision-Recall Curve (AUPRC)
+   - Area Under the NPV-Specificity Curve (NEGATIVE_AUPRC)
    - Specificity (True Negative Rate)
    - Recall/Sensitivity (True Positive Rate)
    - Precision (Positive Predictive Value)
    - Negative Predictive Value (NPV)
    - Accuracy
    - F1 Score
+   - Precision-Recall Average
+   - Precision-Recall Average for Negative Class
 
 2. Confusion matrix metrics:
 
@@ -39,6 +43,8 @@ This class contains the following metrics:
 
    - Youden's J statistic
    - Cost-based threshold
+   - F1 Score threshold
+   - F1 Score threshold for the negative class
 
 6. Average metrics across folds:
 
@@ -52,7 +58,9 @@ This class contains the following metrics:
 class MetricDictionnary:
 
 	# Main metrics (starting with '1:')
-	AUC: str								= "1: Area Under the Curve: AUC"
+	AUC: str								= "1: Area Under the ROC Curve: AUC / AUROC"
+	AUPRC: str								= "1: Area Under the Precision-Recall Curve: AUPRC / PR AUC"
+	NEGATIVE_AUPRC: str						= "1: Area Under the NPV-Specificity Curve: AUNPRC / NPR AUC"
 	SPECIFICITY: str						= "1: Specificity: True Negative Rate"
 	RECALL: str								= "1: Recall/Sensitivity: True Positive Rate"
 	PRECISION: str							= "1: Precision: Positive Predictive Value"
@@ -61,6 +69,8 @@ class MetricDictionnary:
 	BALANCED_ACCURACY: str					= "1: Balanced Accuracy"
 	F1_SCORE: str							= "1: F1 Score"
 	F1_SCORE_NEGATIVE: str					= "1: F1 Score for Negative Class"
+	PR_AVERAGE: str							= "1: Precision-Recall Average"
+	PR_AVERAGE_NEGATIVE: str				= "1: Precision-Recall Average for Negative Class"
 
 	# Confusion matrix metrics (starting with '2:')
 	CONFUSION_MATRIX_TN: str				= "2: Confusion Matrix: TN"
@@ -82,6 +92,8 @@ class MetricDictionnary:
 	# Optimal thresholds (starting with '5:')
 	OPTIMAL_THRESHOLD_YOUDEN: str			= "5: Optimal Threshold: Youden"
 	OPTIMAL_THRESHOLD_COST: str				= "5: Optimal Threshold: Cost"
+	OPTIMAL_THRESHOLD_F1: str				= "5: Optimal Threshold: F1"
+	OPTIMAL_THRESHOLD_F1_NEGATIVE: str		= "5: Optimal Threshold: F1 for Negative Class"
 
 	# Average metrics across folds (starting with '6:')
 	AVERAGE_METRIC: str						= "6: Average METRIC_NAME across folds"

{stouputils-1.3.6 → stouputils-1.3.7}/stouputils/data_science/metric_utils.py

@@ -117,10 +117,10 @@ class MetricUtils:
 			if mcc_metric:
 				metrics.update(mcc_metric)
 
-			# Calculate and plot ROC/AUC
-			roc_metrics: dict[str, float] = MetricUtils.roc_and_auc(true_classes, y_pred, fold_number=-1, run_name=run_name)
-			if roc_metrics:
-				metrics.update(roc_metrics)
+			# Calculate and plot (ROC Curve / AUC) and (PR Curve / AUC, and negative one)
+			curves_metrics: dict[str, float] = MetricUtils.all_curves(true_classes, y_pred, fold_number=-1, run_name=run_name)
+			if curves_metrics:
+				metrics.update(curves_metrics)
 
 		# Multiclass classification
 		elif mode == "multiclass":
@@ -287,7 +287,7 @@ class MetricUtils:
 
 	@staticmethod
 	@handle_error(error_log=DataScienceConfig.ERROR_LOG)
-	def roc_and_auc(
+	def roc_curve_and_auc(
 		true_classes: NDArray[np.intc] | NDArray[np.single],
 		pred_probs: NDArray[np.single],
 		fold_number: int = -1,
@@ -311,7 +311,7 @@ class MetricUtils:
 			>>> pred_probs = np.array([[0.9, 0.1], [0.1, 0.9], [0.1, 0.9]])
 			>>> from stouputils.ctx import Muffle
 			>>> with Muffle():
-			... 	metrics = MetricUtils.roc_and_auc(true_classes, pred_probs, run_name="")
+			... 	metrics = MetricUtils.roc_curve_and_auc(true_classes, pred_probs, run_name="")
 
 			>>> # Check metrics
 			>>> round(float(metrics[MetricDictionnary.AUC]), 2)
@@ -377,6 +377,158 @@ class MetricUtils:
 
 		return metrics
 
+	@staticmethod
+	@handle_error(error_log=DataScienceConfig.ERROR_LOG)
+	def pr_curve_and_auc(
+		true_classes: NDArray[np.intc] | NDArray[np.single],
+		pred_probs: NDArray[np.single],
+		fold_number: int = -1,
+		run_name: str = ""
+	) -> dict[str, float]:
+		""" Calculate Precision-Recall curve and AUC score. (and NPV-Specificity curve and AUC)
+
+		Args:
+			true_classes  (NDArray[np.intc | np.single]):  True class labels (one-hot encoded or class indices)
+			pred_probs    (NDArray[np.single]):            Predicted probabilities (must be probability scores, not class indices)
+			fold_number   (int):                           Fold number, used for naming the plot file, usually
+				-1 for final model with test set,
+				0 for final model with validation set,
+				>0 for other folds with their validation set
+			run_name      (str):                           Name for saving the plot
+		Returns:
+			dict[str, float]:	Dictionary containing AUC score and optimal thresholds
+
+		Examples:
+			>>> true_classes = np.array([0, 1, 0])
+			>>> pred_probs = np.array([[0.9, 0.1], [0.1, 0.9], [0.1, 0.9]])
+			>>> from stouputils.ctx import Muffle
+			>>> with Muffle():
+			... 	metrics = MetricUtils.pr_curve_and_auc(true_classes, pred_probs, run_name="")
+
+			>>> # Check metrics
+			>>> round(float(metrics[MetricDictionnary.AUPRC]), 2)
+			0.75
+			>>> round(float(metrics[MetricDictionnary.NEGATIVE_AUPRC]), 2)
+			0.92
+			>>> round(float(metrics[MetricDictionnary.PR_AVERAGE]), 2)
+			0.5
+			>>> round(float(metrics[MetricDictionnary.PR_AVERAGE_NEGATIVE]), 2)
+			0.33
+			>>> round(float(metrics[MetricDictionnary.OPTIMAL_THRESHOLD_F1]), 2)
+			0.9
+			>>> round(float(metrics[MetricDictionnary.OPTIMAL_THRESHOLD_F1_NEGATIVE]), 2)
+			0.1
+		"""
+		auc_value, average_precision, precision, recall, thresholds = Utils.get_pr_curve_and_auc(true_classes, pred_probs)
+		neg_auc_value, average_precision_neg, npv, specificity, neg_thresholds = (
+			Utils.get_pr_curve_and_auc(true_classes, pred_probs, negative=True)
+		)
+
+		# Calculate metrics
+		metrics: dict[str, float] = {
+			MetricDictionnary.AUPRC: auc_value,
+			MetricDictionnary.NEGATIVE_AUPRC: neg_auc_value,
+			MetricDictionnary.PR_AVERAGE: average_precision,
+			MetricDictionnary.PR_AVERAGE_NEGATIVE: average_precision_neg
+		}
+
+		# Calculate optimal thresholds for both PR curves
+		for is_negative in (False, True):
+
+			# Get the right values based on positive/negative case
+			if not is_negative:
+				curr_precision = precision
+				curr_recall = recall
+				curr_thresholds = thresholds
+				curr_auc = auc_value
+				curr_ap = average_precision
+			else:
+				curr_precision = npv
+				curr_recall = specificity
+				curr_thresholds = neg_thresholds
+				curr_auc = neg_auc_value
+				curr_ap = average_precision_neg
+
+			# Calculate F-score for each threshold
+			fscore: NDArray[np.single] = (2 * curr_precision * curr_recall) / (curr_precision + curr_recall)
+			fscore = fscore[~np.isnan(fscore)]
+
+			# Get optimal threshold (maximum F-score)
+			if len(fscore) > 0:
+				optimal_idx: int = int(np.argmax(fscore))
+				optimal_threshold: float = curr_thresholds[optimal_idx]
+			else:
+				optimal_idx: int = 0
+				optimal_threshold = float('inf')
+
+			# Store in metrics dictionary
+			if not is_negative:
+				metrics[MetricDictionnary.OPTIMAL_THRESHOLD_F1] = optimal_threshold
+			else:
+				metrics[MetricDictionnary.OPTIMAL_THRESHOLD_F1_NEGATIVE] = optimal_threshold
+
+			# Plot ROC curve if not nan
+			if run_name:
+				label: str = "Precision - Recall" if not is_negative else "Negative Predictive Value - Specificity"
+				plt.figure(figsize=(12, 6))
+				plt.plot(curr_recall, curr_precision, "b", label=f"{label} curve (AUC = {curr_auc:.2f}, AP = {curr_ap:.2f})")
+
+				# Prepare the path
+				fold_name: str = ""
+				if fold_number > 0:
+					fold_name = f"_fold_{fold_number}_val"
+				elif fold_number == 0:
+					fold_name = "_val"
+				elif fold_number == -1:
+					fold_name = "_test"
+				elif fold_number == -2:
+					fold_name = "_train"
+				pr: str = "pr" if not is_negative else "negative_pr"
+				curve_path: str = f"{DataScienceConfig.TEMP_FOLDER}/{run_name}_{pr}_curve{fold_name}.png"
+
+				plt.plot(
+					curr_recall[optimal_idx], curr_precision[optimal_idx], 'go', label=f"Optimal threshold (t={optimal_threshold:.2f})"
+				)
+
+				plt.xlim([-0.01, 1.01])
+				plt.ylim([-0.01, 1.01])
+				plt.xlabel("Recall" if not is_negative else "Specificity")
+				plt.ylabel("Precision" if not is_negative else "Negative Predictive Value")
+				plt.title(f"{label} Curve")
+				plt.legend(loc="lower right")
+				plt.savefig(curve_path)
+				mlflow.log_artifact(curve_path)
+				os.remove(curve_path)
+				plt.close()
+
+		return metrics
+
+	@staticmethod
+	@handle_error(error_log=DataScienceConfig.ERROR_LOG)
+	def all_curves(
+		true_classes: NDArray[np.intc] | NDArray[np.single],
+		pred_probs: NDArray[np.single],
+		fold_number: int = -1,
+		run_name: str = ""
+	) -> dict[str, float]:
+		""" Run all X_curve_and_auc functions and return a dictionary of metrics.
+
+		Args:
+			true_classes  (NDArray[np.intc | np.single]):  True class labels (one-hot encoded or class indices)
+			pred_probs    (NDArray[np.single]):            Predicted probabilities (must be probability scores, not class indices)
+			fold_number   (int):                           Fold number, used for naming the plot file, usually
+				-1 for final model with test set,
+				0 for final model with validation set,
+				>0 for other folds with their validation set
+			run_name      (str):                           Name for saving the plot
+		Returns:
+			dict[str, float]: Dictionary containing AUC score and optimal thresholds for ROC and PR curves
+		"""
+		metrics: dict[str, float] = {}
+		metrics.update(MetricUtils.roc_curve_and_auc(true_classes, pred_probs, fold_number, run_name))
+		metrics.update(MetricUtils.pr_curve_and_auc(true_classes, pred_probs, fold_number, run_name))
+		return metrics
+
 
 	@staticmethod
 	@handle_error(error_log=DataScienceConfig.ERROR_LOG)
@@ -609,7 +761,7 @@ class MetricUtils:
 
 				# Find all local minima
 				from scipy.signal import argrelextrema
-				local_minima_idx: NDArray[np.intp] = argrelextrema(y_array, np.less)[0]
+				local_minima_idx: NDArray[np.intp] = np.array(argrelextrema(y_array, np.less)[0], dtype=np.intp)
 				distinct_candidates = np.unique(np.append(local_minima_idx, best_idx))
 		else:
 			assert 0 <= best_idx < len(x_array), "Best x index is out of bounds"

{stouputils-1.3.6 → stouputils-1.3.7}/stouputils/data_science/models/model_interface.py

@@ -851,22 +851,13 @@ class ModelInterface(AbstractModel):
 		self.history.append(history)
 		self.evaluation_results.append(eval_results)
 
-		# Generate and save ROC curve for this fold
-		MetricUtils.roc_and_auc(
-			true_classes=true_classes,
-			pred_probs=predictions,
-			fold_number=fold_number,
-			run_name=self.run_name
-		)
+		# Generate and save ROC Curve and PR Curve for this fold
+		MetricUtils.all_curves(true_classes, predictions, fold_number, run_name=self.run_name)
 
-		# If final model, also log the ROC curve for the train set
+		# If final model, also log the ROC curve and PR curve for the train set
 		if fold_number == 0:
-			MetricUtils.roc_and_auc(
-				true_classes=training_true_classes,
-				pred_probs=training_predictions,
-				fold_number=-2,
-				run_name=self.run_name
-			)
+			fold_number = -2	# -2 is the train set
+			MetricUtils.all_curves(training_true_classes, training_predictions, fold_number, run_name=self.run_name)
 
 		# Log visualization artifacts if they were generated
 		if temp_dir is not None:

{stouputils-1.3.6 → stouputils-1.3.7}/stouputils/data_science/utils.py

@@ -197,22 +197,89 @@ class Utils:
 		assert y_pred.ndim > 1 and y_pred.shape[1] > 1, "Predictions must be probability scores in one-hot format"
 		pred_probs: NDArray[np.single] = y_pred[:, 1]  # Take probability of positive class only
 
-		# Convert true labels to class indices if they're one-hot encoded
-		true_classes: NDArray[np.intc] = Utils.convert_to_class_indices(y_true)
-
 		# Calculate ROC curve and AUC score using probabilities
 		with Muffle(mute_stderr=True):	# Suppress "UndefinedMetricWarning: No positive samples in y_true [...]"
 
 			# Import functions
 			try:
-				from sklearn.metrics import auc, roc_curve
+				from sklearn.metrics import roc_auc_score, roc_curve
 			except ImportError as e:
 				raise ImportError("scikit-learn is required for ROC curve calculation. Install with 'pip install scikit-learn'") from e
 
-			results: tuple[Any, Any, Any] = roc_curve(true_classes, pred_probs, drop_intermediate=False)
+			# Convert y_true to class indices for both functions
+			y_true_indices: NDArray[np.intc] = Utils.convert_to_class_indices(y_true)
+
+			# Calculate AUC score directly using roc_auc_score
+			auc_value: float = float(roc_auc_score(y_true_indices, pred_probs))
+
+			# Calculate ROC curve points
+			results: tuple[Any, Any, Any] = roc_curve(y_true_indices, pred_probs, drop_intermediate=False)
 			fpr: NDArray[np.single] = results[0]
 			tpr: NDArray[np.single] = results[1]
 			thresholds: NDArray[np.single] = results[2]
-			auc_value: float = float(auc(fpr, tpr))
+
 		return auc_value, fpr, tpr, thresholds
 
+	@staticmethod
+	@handle_error(error_log=DataScienceConfig.ERROR_LOG)
+	def get_pr_curve_and_auc(
+		y_true: NDArray[np.intc | np.single],
+		y_pred: NDArray[np.single],
+		negative: bool = False
+	) -> tuple[float, float, NDArray[np.single], NDArray[np.single], NDArray[np.single]]:
+		""" Calculate Precision-Recall Curve (or Negative Precision-Recall Curve) and AUC score.
+
+		Args:
+			y_true  (NDArray[intc | single]):   True class labels (either one-hot encoded or class indices)
+			y_pred  (NDArray[single]):          Predicted probabilities (must be probability scores, not class indices)
+			negative (bool):                    Whether to calculate the negative Precision-Recall Curve
+		Returns:
+			tuple[float, NDArray[np.single], NDArray[np.single], NDArray[np.single]]:
+				Tuple containing either:
+					- AUC score, Average Precision, Precision, Recall, and Thresholds
+					- AUC score, Average Precision, Negative Predictive Value, Specificity, and Thresholds for the negative class
+
+		Examples:
+			>>> # Binary classification example
+			>>> y_true = np.array([0.0, 1.0, 0.0, 1.0, 0.0])
+			>>> y_pred = np.array([[0.2, 0.8], [0.1, 0.9], [0.8, 0.2], [0.2, 0.8], [0.7, 0.3]])
+			>>> auc_value, average_precision, precision, recall, thresholds = Utils.get_pr_curve_and_auc(y_true, y_pred)
+			>>> round(auc_value, 2)
+			0.92
+			>>> round(average_precision, 2)
+			0.83
+			>>> [round(x, 2) for x in precision.tolist()]
+			[0.4, 0.5, 0.67, 1.0, 1.0]
+			>>> [round(x, 2) for x in recall.tolist()]
+			[1.0, 1.0, 1.0, 0.5, 0.0]
+			>>> [round(x, 2) for x in thresholds.tolist()]
+			[0.2, 0.3, 0.8, 0.9]
+		"""
+		# For predictions, assert they are probabilities (one-hot encoded)
+		assert y_pred.ndim > 1 and y_pred.shape[1] > 1, "Predictions must be probability scores in one-hot format"
+		pred_probs: NDArray[np.single] = y_pred[:, 1] if not negative else y_pred[:, 0]
+
+		# Calculate Precision-Recall Curve and AUC score using probabilities
+		with Muffle(mute_stderr=True):	# Suppress "UndefinedMetricWarning: No positive samples in y_true [...]"
+
+			# Import functions
+			try:
+				from sklearn.metrics import auc, average_precision_score, precision_recall_curve
+			except ImportError as e:
+				raise ImportError("scikit-learn is required for PR Curve calculation. Install with 'pip install scikit-learn'") from e
+
+			# Convert y_true to class indices for both functions
+			y_true_indices: NDArray[np.intc] = Utils.convert_to_class_indices(y_true)
+
+			results: tuple[Any, Any, Any] = precision_recall_curve(
+				y_true=y_true_indices,
+				probas_pred=pred_probs,
+				pos_label=1 if not negative else 0
+			)
+			precision: NDArray[np.single] = results[0]
+			recall: NDArray[np.single] = results[1]
+			thresholds: NDArray[np.single] = results[2]
+			auc_value: float = float(auc(recall, precision))
+			average_precision: float = float(average_precision_score(y_true_indices, pred_probs))
+		return auc_value, average_precision, precision, recall, thresholds
+