fix a bug in catnet

Author: SAAS R7 User1

exnn/exnn.py

@@ -6,7 +6,7 @@ class ExNN(BaseNet):
     """
     Enhanced explainable neural network (ExNN) based on sparse, orthogonal and smooth constraints.
 
-    ExNN is based on our paper (Yang et al. 2018) with the following implementation details:
+    ExNN is based on our paper (Yang et al. 2020 TNNLS) with the following implementation details:
 
     1. Categorical variables should be first converted by one-hot encoding, and we directly link each of the dummy variables as a bias term to final output.
 
@@ -80,7 +80,7 @@ class ExNN(BaseNet):
 
     References
     ----------
-    .. Yang, Zebin, Aijun Zhang, and Agus Sudjianto. "Enhancing Explainability of Neural Networks through Architecture Constraints." arXiv preprint arXiv:1901.03838 (2019).
+    .. Yang, Zebin, Aijun Zhang, and Agus Sudjianto. "Enhancing Explainability of Neural Networks through Architecture Constraints." TNNLS (2020).
 
     """
 

exnn/gamnet.py

@@ -4,7 +4,7 @@
 
 class GAMNet(BaseNet):
     """
-    Generalized additive model vai neural network implementation. It is just a simplified version of sosxnn with identity projection layer.
+    Generalized additive model vai neural network implementation. It is just a simplified version of exnn with identity projection layer.
 
     Parameters
     ----------

exnn/layers.py

@@ -46,15 +46,15 @@ def __init__(self, category_num, bn_flag=True, cagetnet_id=0):
         self.cagetnet_id = cagetnet_id
 
         self.categ_bias = self.add_weight(name="cate_bias_" + str(self.cagetnet_id),
-                                                 shape=[self.depth, 1],
+                                                 shape=[self.category_num, 1],
                                                  initializer=tf.zeros_initializer(),
                                                  trainable=True)
         self.moving_mean = self.add_weight(name="mean" + str(self.cagetnet_id), shape=[1], initializer=tf.zeros_initializer(), trainable=False)
         self.moving_norm = self.add_weight(name="norm" + str(self.cagetnet_id), shape=[1], initializer=tf.ones_initializer(), trainable=False)
 
     def call(self, inputs, training=False):
 
-        dummy = tf.one_hot(indices=tf.cast(inputs[:, 0], tf.int32), depth=self.depth)
+        dummy = tf.one_hot(indices=tf.cast(inputs[:, 0], tf.int32), depth=self.category_num)
         self.output_original = tf.matmul(dummy, self.categ_bias)
 
         if training: