From e424d209a7b1338b59bc39cd82b03ea9e8c5a646 Mon Sep 17 00:00:00 2001
From: Peter Hawkins <phawkins@google.com>
Date: Wed, 12 Jul 2017 13:10:45 -0700
Subject: [PATCH] [TF:XLA] Use a more numerically accurate formulation of
 ResourceApplyRMSProp.

PiperOrigin-RevId: 161706120
---
 tensorflow/compiler/tf2xla/kernels/training_ops.cc | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tensorflow/compiler/tf2xla/kernels/training_ops.cc b/tensorflow/compiler/tf2xla/kernels/training_ops.cc
index 42104b951ac..a2ecbca124c 100644
--- a/tensorflow/compiler/tf2xla/kernels/training_ops.cc
+++ b/tensorflow/compiler/tf2xla/kernels/training_ops.cc
@@ -352,9 +352,9 @@ class ResourceApplyRMSProp : public XlaOpKernel {
                b->Sub(XlaHelpers::FloatLiteral(b, type, 1.0), rho)));
     xla::ComputationDataHandle new_mom =
         b->Add(b->Mul(mom, momentum),
-               b->Div(b->Mul(grad, lr),
+               b->Mul(b->Mul(grad, lr),
                       b->Pow(b->Add(new_ms, epsilon),
-                             XlaHelpers::FloatLiteral(b, type, 0.5))));
+                             XlaHelpers::FloatLiteral(b, type, -0.5))));
     xla::ComputationDataHandle new_var = b->Sub(var, new_mom);
 
     OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, type, new_var));