dragon-ml-toolbox 20.11.0__tar.gz → 20.13.0__tar.gz

{dragon_ml_toolbox-20.11.0/dragon_ml_toolbox.egg-info → dragon_ml_toolbox-20.13.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dragon-ml-toolbox
-Version: 20.11.0
+Version: 20.13.0
 Summary: Complete pipelines and helper tools for data science and machine learning projects.
 Author-email: Karl Luigi Loza Vidaurre <luigiloza@gmail.com>
 License-Expression: MIT

{dragon_ml_toolbox-20.11.0 → dragon_ml_toolbox-20.13.0/dragon_ml_toolbox.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dragon-ml-toolbox
-Version: 20.11.0
+Version: 20.13.0
 Summary: Complete pipelines and helper tools for data science and machine learning projects.
 Author-email: Karl Luigi Loza Vidaurre <luigiloza@gmail.com>
 License-Expression: MIT

{dragon_ml_toolbox-20.11.0 → dragon_ml_toolbox-20.13.0}/ml_tools/ML_evaluation_captum/_ML_evaluation_captum.py

@@ -30,7 +30,8 @@ def captum_feature_importance(model: nn.Module,
                               save_dir: Union[str, Path],
                               target_names: Optional[list[str]] = None,
                               n_steps: int = 50,
-                              device: Union[str, torch.device] = 'cpu'):
+                              device: Union[str, torch.device] = 'cpu',
+                              verbose: int = 0):
     """
     Calculates feature importance using Captum's Integrated Gradients.
 
@@ -49,7 +50,7 @@ def captum_feature_importance(model: nn.Module,
             - If `None`, generic names (e.g., "Output_0") will be generated based on model output shape.
         n_steps (int): Number of steps for the integral approximation. Higher means more accurate but slower.
         device (str | torch.device): Torch device.
-        
+        verbose (int): Verbosity level.
     <br>
         
     ### NOTE: 
@@ -127,7 +128,8 @@ def captum_feature_importance(model: nn.Module,
             save_dir=save_dir_path,
             n_steps=n_steps,
             file_suffix=f"_{clean_name}",
-            target_name=name  # Pass original name for plotting
+            target_name=name,  # Pass original name for plotting
+            verbose=verbose
         )
 
 
@@ -139,7 +141,8 @@ def _process_single_target(ig: 'IntegratedGradients', # type: ignore
                            save_dir: Path,
                            n_steps: int,
                            file_suffix: str,
-                           target_name: str = ""):
+                           target_name: str = "",
+                           verbose: int = 0):
     """
     Private helper to run the attribution, aggregation, and saving for a single context.
     """
@@ -153,8 +156,8 @@ def _process_single_target(ig: 'IntegratedGradients', # type: ignore
                                            return_convergence_delta=True)
         # Check convergence quality
         mean_delta = torch.mean(torch.abs(delta)).item()
-        if mean_delta > 0.1:
-            _LOGGER.warning(f"Captum Convergence Delta is high ({mean_delta:.4f}). The attribution approximation may be inaccurate. Consider increasing 'n_steps'.")
+        if mean_delta > 0.1 and verbose > 0:
+            _LOGGER.warning(f"Captum Convergence Delta is high ({mean_delta:.4f}). Consider increasing 'n_steps'.")
         
     except Exception as e:
         _LOGGER.error(f"Captum attribution failed for target '{target_index}': {e}")
@@ -198,11 +201,36 @@ def _process_single_target(ig: 'IntegratedGradients', # type: ignore
         min_len = min(len(mean_abs_attr), len(feature_names))
         mean_abs_attr = mean_abs_attr[:min_len]
         feature_names = feature_names[:min_len]
+    
+    # Calculate percentages (Before Min-Max scaling to preserve relative importance)
+    total_attr_sum = np.sum(mean_abs_attr)
+    if total_attr_sum > 0:
+        attr_percentages = (mean_abs_attr / total_attr_sum) * 100.0
+    else:
+        attr_percentages = np.zeros_like(mean_abs_attr)
+    
+    # Min-Max Scaling
+    target_min = 0.01
+    target_max = 1.0
+    
+    _min = np.min(mean_abs_attr)
+    _max = np.max(mean_abs_attr)
+    
+    if _max > _min:
+        # 1. Normalize to [0, 1]
+        mean_abs_attr = (mean_abs_attr - _min) / (_max - _min)
+        # 2. Scale to [target_min, target_max]
+        mean_abs_attr = mean_abs_attr * (target_max - target_min) + target_min
+    else:
+        # Fallback: if all values are identical (e.g. all 0.0), set to target_min
+        fill_val = target_min if _max == 0 else target_max
+        mean_abs_attr = np.full_like(mean_abs_attr, fill_val)
 
     # --- Save Data to CSV ---
     summary_df = pd.DataFrame({
         CaptumKeys.FEATURE_COLUMN: feature_names,
-        CaptumKeys.IMPORTANCE_COLUMN: mean_abs_attr
+        CaptumKeys.IMPORTANCE_COLUMN: mean_abs_attr,
+        CaptumKeys.PERCENT_COLUMN: attr_percentages
     }).sort_values(CaptumKeys.IMPORTANCE_COLUMN, ascending=False)
     
     csv_name = f"{CaptumKeys.SAVENAME}{file_suffix}.csv"
@@ -210,10 +238,13 @@ def _process_single_target(ig: 'IntegratedGradients', # type: ignore
     summary_df.to_csv(csv_path, index=False)
 
     # --- Generate Plot ---
-    plot_df = summary_df.head(20).sort_values(CaptumKeys.IMPORTANCE_COLUMN, ascending=True)
-    plt.figure(figsize=(10, 8), dpi=200)
-    plt.barh(plot_df[CaptumKeys.FEATURE_COLUMN], plot_df[CaptumKeys.IMPORTANCE_COLUMN], color='mediumpurple')
-    plt.xlabel("Mean Absolute Attribution")
+    plot_df = summary_df.head(20).sort_values(CaptumKeys.PERCENT_COLUMN, ascending=True)
+    plt.figure(figsize=(10, 8), dpi=300)
+    plt.barh(plot_df[CaptumKeys.FEATURE_COLUMN], plot_df[CaptumKeys.PERCENT_COLUMN], color='mediumpurple')
+    # plt.xlim(0, 1.05) # standardized scale # Removed to reflect actual percentages
+    plt.xlim(left=0) # start at 0
+    # plt.xlabel("Scaled Mean Absolute Attribution")
+    plt.xlabel("Relative Importance (%)")
     
     title = "Feature Importance"
     

{dragon_ml_toolbox-20.11.0 → dragon_ml_toolbox-20.13.0}/ml_tools/ML_trainer/_dragon_trainer.py

@@ -33,8 +33,7 @@ from ..ML_configuration import (FormatRegressionMetrics,
                             FinalizeMultiTargetRegression,
                             FinalizeRegression)
 
-from ..path_manager import make_fullpath
-from ..keys._keys import PyTorchLogKeys, PyTorchCheckpointKeys, DatasetKeys, MLTaskKeys, MagicWords, DragonTrainerKeys, ScalerKeys
+from ..keys._keys import PyTorchLogKeys, PyTorchCheckpointKeys, DatasetKeys, MLTaskKeys, DragonTrainerKeys, ScalerKeys
 from .._core import get_logger
 
 from ._base_trainer import _BaseDragonTrainer
@@ -824,7 +823,8 @@ class DragonTrainer(_BaseDragonTrainer):
                        n_samples: int = 100,
                        feature_names: Optional[list[str]] = None,
                        target_names: Optional[list[str]] = None,
-                       n_steps: int = 50):
+                       n_steps: int = 50,
+                       verbose: int = 0):
         """
         Explains model predictions using Captum's Integrated Gradients.
         
@@ -927,7 +927,8 @@ class DragonTrainer(_BaseDragonTrainer):
                 save_dir=save_dir,
                 target_names=target_names,
                 n_steps=n_steps,
-                device=self.device
+                device=self.device,
+                verbose=verbose
             )
 
     def _attention_helper(self, dataloader: DataLoader):

{dragon_ml_toolbox-20.11.0 → dragon_ml_toolbox-20.13.0}/ml_tools/keys/_keys.py

@@ -99,8 +99,9 @@ class SHAPKeys:
 
 class CaptumKeys:
     """Keys for Captum functions"""
-    FEATURE_COLUMN = "feature"
-    IMPORTANCE_COLUMN = "importance"
+    FEATURE_COLUMN = "Feature"
+    IMPORTANCE_COLUMN = "Scaled Mean Attribution"
+    PERCENT_COLUMN = "Relative Importance(%)"
     SAVENAME = "captum_summary"
     PLOT_NAME = "captum_importance_plot"
 

{dragon_ml_toolbox-20.11.0 → dragon_ml_toolbox-20.13.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dragon-ml-toolbox"
-version = "20.11.0"
+version = "20.13.0"
 description = "Complete pipelines and helper tools for data science and machine learning projects."
 authors = [
     { name = "Karl Luigi Loza Vidaurre", email = "luigiloza@gmail.com" }