Removed the constraint on positive weights during the minimization of dii (file diff_imbalance.py). Introduction of an early stopping criteria (diff_imbalance.py). Implementation of the Langevin dynamic, for the estimation of the error (diff_imbalance.py). Insertion of the calculation of the error with langevin dynamic (causal_graph.py).

Author: Matteo Allione

dadapy/causal_graph.py

@@ -28,7 +28,7 @@
 import numpy as np
 from tqdm.auto import tqdm
 
-from diff_imbalance_langevin import DiffImbalance
+from dadapy import DiffImbalance
 
 
 class CausalGraph(DiffImbalance):
@@ -482,7 +482,7 @@ def optimize_present_to_future(
                     weights_temp = np.concatenate([weights_temp, weights_langevin[1:]], axis=0)
                     imbs_training[i_var, j_tau] = np.concatenate([imbs_temp[i_var, j_tau], imbs_langevin[1:]])
                     if return_covariance:
-                        covariance_matrix[i_var, j_tau] = np.cov(weights_langevin ,rowvar=False)
+                        covariance_matrix[i_var, j_tau] = np.cov(weights_langevin[::10] ,rowvar=False)/len(weights_langevin[::10])
 
                 # save weights
                 if embedding_dim_present == 1:
@@ -497,7 +497,7 @@ def optimize_present_to_future(
                     )
                     if save_weights is True:
                         weights_training[i_var, j_tau] = weights_temp.reshape(
-                            (num_epochs + 1 + langevin_steps, self.num_variables, embedding_dim_present)
+                            (num_epochs + 1 + langevin_steps + langevin_steps, self.num_variables, embedding_dim_present)
                         )
 
         self.weights_final = weights_final

dadapy/diff_imbalance.py

@@ -181,6 +181,7 @@ def __init__(
         learning_rate=1e-2,
         learning_rate_decay=None,
         num_points_rows=None,
+        gradient_clip_value=0.0,
         early_stopping=0,
     ):
         """Initialise the DiffImbalance class."""
@@ -1635,7 +1636,7 @@ def backward_greedy_feature_selection(
             print("------------------------------------------------")
 
         return feature_sets, diis, errors, best_weights_list
-
+    
     def langevin(self, initial_params, n_epochs=1000, batch_size=100, noise=0):
         """
         We consider Langevin dynamics as a method to estimate an error of the parameters 
@@ -1700,4 +1701,4 @@ def langevin(self, initial_params, n_epochs=1000, batch_size=100, noise=0):
             imbs_langevin = imbs_langevin.at[epoch_idx].set(imb_now)
 
         return np.array(params_langevin), np.array(imbs_langevin)
-        
+