added tweedie

tensorflow_probability/python/distributions/tweedieloss.py

@@ -0,0 +1,41 @@
+import tensorflow.compat.v2 as tf
+from tensorflow_probability.python.distributions import distribution
+
+
+
+_tweedie_sample_note = """definition is [here](https://en.wikipedia.org/wiki/Tweedie_distribution), 
+                          Reference Document [here](https://arxiv.org/pdf/1912.12356.pdf)"""
+
+
+class Tweedie(distribution.AutoCompositeTensorDistribution):
+
+  def __init__(self, p=1.5,
+               name='Tweedie'):
+    """Input Paramters for the Tweedie Loss are the p value, lies bettern (1,2) for Poisson-gamma EDM   
+     |---------------|---------------------|
+     |  Value of p   |     Distribution    |
+     |---------------|---------------------|
+     |       0       |     Normal          |
+     |       1       |     Poisson         |
+     |     (1,2)     |     Poisson-Gamma   |
+     |       2       |     Gamma           |
+     |---------------|---------------------|
+     
+    """
+    self.p =p
+    self.name = name
+
+  def log_likelihood(self,y,y_pred):
+    """The Log likelihood"""
+    self.loglikelihood = - y * (tf.pow(y_pred, 1-self.p)/(1-self.p)) + (tf.pow(y_pred,2-self.p)/(2-self.p))
+    return tf.reduce_mean(self.loglikelihood)
+
+  def deviance(self,y,y_pred):
+    if self.p ==0:
+        return tf.pow(y-y_pred,2)
+    elif self.p ==1:
+        return 2*(y*log(y_pred/y) + y_pred - y)
+    elif self.p ==2:
+        return 2*(log(y_pred/y)+ y/y_pred -1)
+    else:
+        return 2*(((max(y,0)**(2-self.p))/((1-self.p)*(2-self.p))) - (y*(tf.pow(y_pred,(1-self.p)))/(1-self.p)) + (tf.pow(y_pred,2-self.p)/(2-self.p)))