Introduced the kwargs activation_func to margarine for choosing activ… (

#57)

* Introduced the kwargs activation_func to margarine for choosing activation function. The default one is tanh.

* Added description of the new activation_func kwargs to the class MAF. Added activation_fuc kwarg to clustered.py in line 229.

* Added description of the new activation_func arguments to clustered.py.

README.rst

@@ -7,7 +7,7 @@ Introduction
 
 :margarine: Marginal Bayesian Statistics
 :Authors: Harry T.J. Bevins
-:Version: 1.2.4
+:Version: 1.2.5
 :Homepage:  https://github.com/htjb/margarine
 :Documentation: https://margarine.readthedocs.io/
 

margarine/clustered.py

@@ -49,6 +49,11 @@ class clusterMAF():
                 each neural network. The default is two hidden layers with
                 50 nodes each and each network in the chain has the same hidden
                 layer structure.
+        
+        activation_func: **string / default = 'tanh'**
+            | The choice of activation function. It must be an activation
+            function keyword recognisable by TensorFlow. The default is
+            'tanh', the hyperbolic tangent activation function.
 
         cluster_labels: **list / default = None**
             | If clustering has been performed externally to margarine you can
@@ -77,6 +82,7 @@ def __init__(self, theta, **kwargs):
         self.cluster_labels = kwargs.pop('cluster_labels', None)
         self.cluster_number = kwargs.pop('cluster_number', None)
         self.parameters = kwargs.pop('parameters', None)
+        self.activation_func = kwargs.pop('activation_func', 'tanh')
 
         # Avoid unintended side effects by copying theta
         theta = theta.copy()
@@ -232,7 +238,8 @@ def __init__(self, theta, **kwargs):
                                  learning_rate=self.learning_rate,
                                  hidden_layers=self.hidden_layers,
                                  theta_min=self.theta_min,
-                                 theta_max=self.theta_max))
+                                 theta_max=self.theta_max,
+                                 activation_func=self.activation_func))
 
     def train(self, epochs=100, early_stop=False, loss_type='sum'):
 

margarine/maf.py

@@ -48,6 +48,11 @@ class MAF:
                 50 nodes each and each network in the chain has the same hidden
                 layer structure.
 
+        activation_func: **string / default = 'tanh'**
+            | The choice of activation function. It must be an activation
+            function keyword recognisable by TensorFlow. The default is
+            'tanh', the hyperbolic tangent activation function.
+
         theta_max: **numpy array**
             | The true upper limits of the priors used to generate the samples
                 that we want the MAF to learn.
@@ -85,6 +90,7 @@ def __init__(self, theta, **kwargs):
         self.learning_rate = kwargs.pop('learning_rate', 1e-3)
         self.hidden_layers = kwargs.pop('hidden_layers', [50, 50])
         self.parameters = kwargs.pop('parameters', None)
+        self.activation_func = kwargs.pop('activation_func', 'tanh')
 
         # Avoids unintended side effects outside the class
         if not isinstance(theta, tf.Tensor):
@@ -176,7 +182,7 @@ def gen_mades(self):
         """Generating the masked autoregressive flow."""
 
         self.mades = [tfb.AutoregressiveNetwork(params=2,
-                      hidden_units=self.hidden_layers, activation='tanh',
+                      hidden_units=self.hidden_layers, activation=self.activation_func,
                       input_order='random')
                       for _ in range(self.number_networks)]