renamed dataloader , rescaled target, removed target scale, target ce…

…nter

configs/prediction_models/TFTpytorch.gin

@@ -22,7 +22,7 @@ model/hyperparameter.lr_scheduler = "exponential"
 
 # Dataset params
 PredictionDatasetTFTpytorch.max_encoder_length = 24
-PredictionDatasetTFTpytorch.max_prediction_length = 24
+PredictionDatasetTFTpytorch.max_prediction_length = 1
 PredictionDatasetTFTpytorch.target="label"
 PredictionDatasetTFTpytorch.time_varying_known_reals=["time_idx"]
 PredictionDatasetTFTpytorch.add_relative_time_idx=False

configs/prediction_models/common/DLCommon.gin

@@ -15,7 +15,7 @@ base_regression_preprocessor.generate_features = False
 # Train params
 train_common.optimizer = @Adam
 train_common.epochs = 1000
-train_common.batch_size = 1
+train_common.batch_size = 64
 train_common.patience = 20
 train_common.min_delta = 1e-4
 

configs/prediction_models/common/DLTuning.gin

@@ -1,5 +1,5 @@
 # Hyperparameter tuner settings for Deep Learning.
 tune_hyperparameters.scopes = ["model", "optimizer"]
 tune_hyperparameters.n_initial_points = 5
-tune_hyperparameters.n_calls = 30
+tune_hyperparameters.n_calls = 15
 tune_hyperparameters.folds_to_tune_on = 2

configs/tasks/Regression.gin

@@ -26,7 +26,7 @@ preprocess.use_static = True
 
 # SPECIFYING REGRESSION OUTCOME SCALING
 base_regression_preprocessor.outcome_min = 0
-base_regression_preprocessor.outcome_max = 15
+base_regression_preprocessor.outcome_max = 168
 
 # SELECTING DATASET
 PredictionDataset.vars = %vars

icu_benchmarks/data/loader.py

@@ -429,8 +429,8 @@ def __getitem__(self, idx: int) -> Tuple[Tensor, Tensor, Tensor]:
         return from_numpy(window.values).to(float32)
 
 
-@gin.configurable("PredictionDatasetTFTpytorch")
-class PredictionDatasetTFTpytorch(TimeSeriesDataSet):
+@gin.configurable("PredictionDatasetpytorch")
+class PredictionDatasetpytorch(TimeSeriesDataSet):
     """Subclass of timeseries dataset works with pyotrch forecasting library .
 
     Args:
@@ -498,8 +498,8 @@ def __init__(
             time_varying_unknown_categoricals=time_varying_unknown_categoricals,
             time_varying_unknown_reals=time_varying_unknown_reals,
             add_relative_time_idx=add_relative_time_idx,
-            add_target_scales=True,
-            add_encoder_length=True,
+            # add_target_scales=True,
+            # add_encoder_length=True,
             predict_mode=True,
             target_normalizer=None,
         )

icu_benchmarks/models/train.py

@@ -22,7 +22,7 @@
 from icu_benchmarks.data.loader import (
     PredictionDataset,
     ImputationDataset,
-    PredictionDatasetTFTpytorch,
+    PredictionDatasetpytorch,
 )
 from icu_benchmarks.models.utils import save_config_file, JSONMetricsLogger
 from icu_benchmarks.contants import RunMode
@@ -111,7 +111,7 @@ def train_common(
         ImputationDataset
         if mode == RunMode.imputation
         else (
-            PredictionDatasetTFTpytorch if (pytorch_forecasting) else PredictionDataset
+            PredictionDatasetpytorch if (pytorch_forecasting) else PredictionDataset
         )
     )
 
@@ -277,31 +277,37 @@ def train_common(
 
     if explain:
 
-        attributions_Saliency = model.explantation_captum(
-            test_loader=test_loader,
-            method=Saliency, log_dir=log_dir, plot=True
-        )
-        print("saliency", attributions_Saliency)
+        # attributions_Saliency = model.explantation_captum(
+        #    test_loader=test_loader,
+        #    method=Saliency, log_dir=log_dir, plot=True
+        # )
+
+        # print("saliency", attributions_Saliency)
+        """
         attributions_shap = model.explantation_captum(
             test_loader=test_loader,
             method=ShapleyValueSampling, log_dir=log_dir, plot=True, n_samples=10
         )
         print("shap", attributions_shap)
-        attributions_IG = model.explantation_captum(
-            test_loader=test_loader,
-            method=IntegratedGradients, log_dir=log_dir, plot=True, n_steps=20
-        )
-        print("IG", attributions_IG)
+        """
+       # attributions_IG = model.explantation_captum(
+       #     test_loader=test_loader,
+       #     method=IntegratedGradients, log_dir=log_dir, plot=True, n_steps=20
+       # )
+       # print("IG", attributions_IG)
         Interpertations = model.interpertations(test_loader, log_dir, plot=True)
         print("attention", Interpertations)
         attributions_dict = {
             "attributions_Saliency": attributions_Saliency.tolist(),
             "attributions_IG": attributions_IG.tolist(),
-            "attributions_shap": attributions_shap.tolist(),
+            # "attributions_shap": attributions_shap.tolist(),
             "attention_weights": Interpertations["attention"].tolist(),
             "static_variables": Interpertations["static_variables"].tolist(),
             "encoder_variables": Interpertations["encoder_variables"].tolist()
         }
+        variable_importance = np.concatenate((Interpertations["static_variables"].cpu().detach().numpy(),
+                                              Interpertations["encoder_variables"].cpu().detach().numpy()))
+        ind = np.argpartition(variable_importance, -5)[-5:]
 
         # Path to the JSON file in log_dir
         json_file_path = f"{log_dir}/Attributions.json"
@@ -310,26 +316,24 @@ def train_common(
         with open(json_file_path, 'w') as json_file:
             json.dump(attributions_dict, json_file)
         if XAI_metric:
-
             random_attributions = np.random.normal(size=24)
-
-            F_baseline = model.Faithfulness_Correlation(test_loader, random_attributions, pertrub='Noise')
+            F_baseline = model.Faithfulness_Correlation(test_loader, random_attributions, pertrub='Noise', ind=ind)
             print('Random normal faithfulness correlation', F_baseline)
-
             F_attribution = model.Faithfulness_Correlation(test_loader, attributions_IG, pertrub='Noise')
             print('Attributions faithfulness correlation', F_attribution)
-            F_attention = model.Faithfulness_Correlation(test_loader, Interpertations["attention"], pertrub='Noise')
+            F_attention = model.Faithfulness_Correlation(
+                test_loader, Interpertations["attention"], pertrub='Noise', ind=ind)
             print('Attention faithfulness correlation', F_attention)
             F_saliency = model.Faithfulness_Correlation(test_loader, attributions_Saliency, pertrub='Noise')
             print('Saliency faithfulness correlation', F_saliency)
-            F_shap = model.Faithfulness_Correlation(test_loader, attributions_shap, pertrub='Noise')
-            print('shap faithfulness correlation', F_shap)
+           # F_shap = model.Faithfulness_Correlation(test_loader, attributions_shap, pertrub='Noise')
+           # print('shap faithfulness correlation', F_shap)
             XAI_dict = {
                 "Faithfulness_correlation_normal_random": F_baseline,
                 "Faithfulness_correlation_IG": F_attribution,
                 "Faithfulness_correlation_attention": F_attention,
                 "Faithfulness_correlation_saliency": F_saliency,
-                "Faithfulness_correlation_shap": F_shap
+                #    "Faithfulness_correlation_shap": F_shap
             }
 
         # Path to the JSON file in log_dir

icu_benchmarks/models/wrappers.py

@@ -529,7 +529,7 @@ def explantation_captum(self, test_loader, method, log_dir=".", plot=False, **kw
             explantation = method(self.forward_captum)
             # Reformat attributions.
             attr_all_timesteps = []
-            for time_step in range(23, 24):
+            for time_step in range(0, 24):
                 if method is not captum.attr.Saliency:
                     attr = explantation.attribute(
                         data, target=(time_step),  baselines=baselines, **kwargs
@@ -593,7 +593,7 @@ def explantation_captum(self, test_loader, method, log_dir=".", plot=False, **kw
         return means
 
     def Faithfulness_Correlation(
-        self, test_loader, attribution, similarity_func=None, nr_runs=2, pertrub=None, subset_size=6, features=False
+        self, test_loader, attribution, similarity_func=None, nr_runs=2, pertrub=None, subset_size=3, ind=[]
     ):
         """
         Implementation of faithfulness correlation by Bhatt et al., 2020.
@@ -654,16 +654,24 @@ def Faithfulness_Correlation(
 
                 if pertrub == "Noise":
                     # add normal noise to input
+
                     noise = torch.randn_like(x["encoder_cont"])
+                    if (len(ind) == 0):
+                        x["encoder_cont"][:, a_ix, :] += noise[:, a_ix, :]
+                    else:
+
+                        x["encoder_cont"][:, a_ix[:, None], ind] += noise[:, a_ix[:, None], ind]
 
-                    x["encoder_cont"][:, a_ix, :] += noise[:, a_ix, :]
                 elif pertrub == "baseline":
                     # Create a mask tensor with zeros at specified time steps and ones everywhere else
                     # pytorch bug need to change to cpu for next step and then revert
                     mask = torch.ones_like(x["encoder_cont"]).cpu()
-
-                    mask[:, a_ix, :] = 0
+                    if (len(ind) == 0):
+                        mask[:, a_ix, :] = 0
+                    else:
+                        mask[:, a_ix, ind] = 0
                     mask = mask.to(x["encoder_cont"].device)
+
                     x["encoder_cont"] = x["encoder_cont"] * mask
                 data = (
                     x["encoder_cat"],