From 1d04b19ce86ccf055f58955142447aab577d6171 Mon Sep 17 00:00:00 2001
From: Abhinav Arora <abhinavarora28@gmail.com>
Date: Wed, 6 Dec 2017 01:55:12 +0530
Subject: [PATCH] Fix the rendering of latex equation for adamax op (#6294)

* Using latex fraction syntax in sigmoid and logsigmoid op

* Fixing the rendering of the latex equations in adamax operator
---
 paddle/operators/activation_op.cc |  8 ++++----
 paddle/operators/adamax_op.cc     | 10 ++++++----
 2 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/paddle/operators/activation_op.cc b/paddle/operators/activation_op.cc
index 154c618e8e7..83262f950e1 100644
--- a/paddle/operators/activation_op.cc
+++ b/paddle/operators/activation_op.cc
@@ -44,9 +44,9 @@ class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "Input of Sigmoid operator");
     AddOutput("Y", "Output of Sigmoid operator");
     AddComment(R"DOC(
-Sigmoid Activation Operator.
+Sigmoid Activation Operator
 
-$y = 1 / (1 + e^{-x})$
+$$y = \frac{1}{1 + e^{-x}}$$
 
 )DOC");
   }
@@ -60,9 +60,9 @@ class LogSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "Input of LogSigmoid operator");
     AddOutput("Y", "Output of LogSigmoid operator");
     AddComment(R"DOC(
-Logsigmoid Activation Operator.
+Logsigmoid Activation Operator
 
-$y = \log(1 / (1 + e^{-x}))$
+$$y = \log \frac{1}{1 + e^{-x}}$$
 
 )DOC");
   }
diff --git a/paddle/operators/adamax_op.cc b/paddle/operators/adamax_op.cc
index d5bbc672e18..867ddd97907 100644
--- a/paddle/operators/adamax_op.cc
+++ b/paddle/operators/adamax_op.cc
@@ -107,10 +107,12 @@ Adam algorithm based on the infinity norm.
 
 Adamax updates:
 
-$$momentOut = \beta_1 * moment + (1 - \beta_1) * grad \break
-infNormOut = max(\beta_2 * infNorm + \epsilon, |grad|) \break
-learningRate = learningRate /(1 - \beta_1_{pow}) \break
-paramOut = param - learningRate * momentPut / infNormOut$$
+$$
+  momentOut = \beta_{1} * moment + (1 - \beta_{1}) * grad \\
+  infNormOut = max(\beta_{2} * infNorm + \epsilon, |grad|) \\
+  learningRate = \frac{learningRate}{1 - \beta_{1}^{Beta1Pow}} \\
+  paramOut = param - learningRate * \frac{momentOut}{infNormOut}
+$$
 
 The original paper does not have an epsilon attribute.
 However, it is added here for numerical stability to prevent the
-- 
GitLab