[UPD] Examples ready for v2

Author: glederrey

README.md

@@ -9,7 +9,7 @@ Directed Acyclic Tabular GAN (**DATGAN**) for integrating expert knowledge in sy
 - Development Status: [Alpha](https://pypi.org/search/?q=&o=&c=Development+Status+%3A%3A+3+-+Alpha)
 - Homepage: https://github.com/glederrey/DATGAN
 
-> The preprint of the article for this model will be available on arXiv by the end of February/early March.
+> **The preprint of the article for this model should be available on arXiv by mid-March.**
 
 ## Overview
 
@@ -153,7 +153,7 @@ Once you have a **DATGAN** instance, you can call the method `fit` and passing t
 - `data`: the original DataFrame
 - `graph`: the `networkx` DAG
 - `continuous_columns`: the list of continuous columns
-- `preprocessed_data_path`: the path to the preprocessed data if done in Step 4.
+- `preprocessed_data_path`: the path to the preprocessed data if done in Step 4 or the path where to save them.
 ```python
 datgan.fit(df, graph, continuous_columns, preprocessed_data_path='./encoded_data')
 ```

datgan/datgan.py

@@ -397,7 +397,7 @@ def default_parameter_values(self, data, continuous_columns):
             if self.loss_function == 'SGAN':
                 self.g_period = 1
             elif self.loss_function == 'WGAN':
-                self.g_period = 3
+                self.g_period = 2
             elif self.loss_function == 'WGGP':
                 self.g_period = 5
 

datgan/synthesizer/losses/WGGPLoss.py

@@ -25,7 +25,7 @@ def __init__(self, metadata, var_order, name="WGGPLoss"):
         """
         super().__init__(metadata, var_order, name=name)
 
-        self.lambda_ = tf.Variable(10.0, dtype=tf.float32)
+        self.lambda_ = tf.constant(10.0, dtype=tf.float32)
 
     def gen_loss(self, synth_output, transformed_orig, transformed_synth, l2_reg):
         """
@@ -87,7 +87,7 @@ def discr_loss(self, orig_output, synth_output, interp_grad, l2_reg):
 
         # the gradient penalty loss
         interp_grad_norm = tf.sqrt(tf.reduce_sum(tf.square(interp_grad), axis=[1]))
-        grad_pen = self.lambda_ * tf.reduce_mean((interp_grad_norm - 1.0) ** 2)
+        grad_pen = tf.multiply(self.lambda_, tf.reduce_mean((interp_grad_norm - 1.0) ** 2))
 
         # Full loss
         loss = fake_loss - real_loss + grad_pen + l2_reg

datgan/synthesizer/losses/__init__.py

@@ -279,8 +279,8 @@ def initialize(self):
             self.loss = WGANLoss(self.metadata, self.var_order)
         elif self.loss_function == 'WGGP':
             # Optimizers
-            self.optimizerG = tf.keras.optimizers.Adam(learning_rate=self.learning_rate, beta_1=0, beta_2=0.9)
-            self.optimizerD = tf.keras.optimizers.Adam(learning_rate=self.learning_rate, beta_1=0, beta_2=0.9)
+            self.optimizerG = tf.keras.optimizers.Adam(learning_rate=self.learning_rate, beta_1=0.0, beta_2=0.9)
+            self.optimizerD = tf.keras.optimizers.Adam(learning_rate=self.learning_rate, beta_1=0.0, beta_2=0.9)
 
             # Loss function
             self.loss = WGGPLoss(self.metadata, self.var_order)