[incubator-mxnet] branch master updated: In-place updates for Nadam, Adadelta, Adamax and SGLD (#13960)

[ha...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

haibin pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git


The following commit(s) were added to refs/heads/master by this push:
     new a4e249b  In-place updates for Nadam, Adadelta, Adamax and SGLD (#13960)
a4e249b is described below

commit a4e249bee2db7b931ecdd6fed05a94e742e7c3c5
Author: Anirudh <an...@gmail.com>
AuthorDate: Thu Feb 14 17:07:43 2019 -0800

    In-place updates for Nadam, Adadelta, Adamax and SGLD (#13960)
---
 python/mxnet/optimizer/optimizer.py | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)

diff --git a/python/mxnet/optimizer/optimizer.py b/python/mxnet/optimizer/optimizer.py
index a986f27..def2c95 100644
--- a/python/mxnet/optimizer/optimizer.py
+++ b/python/mxnet/optimizer/optimizer.py
@@ -1091,8 +1091,9 @@ class SGLD(Optimizer):
         grad = grad * self.rescale_grad
         if self.clip_gradient is not None:
             grad = clip(grad, -self.clip_gradient, self.clip_gradient)
-        weight[:] += - lr/2 * (grad + wd * weight) + normal(0, math.sqrt(lr), shape=weight.shape,
-                                                            dtype=weight.dtype, ctx=weight.context)
+        weight[:] += - lr/2 * (grad + wd * weight)
+        weight[:] += normal(0, math.sqrt(lr), shape=weight.shape,
+                            dtype=weight.dtype, ctx=weight.context)
 
 
 
@@ -1372,9 +1373,11 @@ class AdaDelta(Optimizer):
         acc_g, acc_delta = state
 
         # update g, delta
-        acc_g[:] = self.rho * acc_g + (1. - self.rho) * grad * grad
+        acc_g[:] *= self.rho
+        acc_g[:] += (1. - self.rho) * grad * grad
         current_delta = sqrt(acc_delta + self.epsilon) / sqrt(acc_g + self.epsilon) * grad
-        acc_delta[:] = self.rho * acc_delta + (1. - self.rho) * current_delta * current_delta
+        acc_delta[:] *= self.rho
+        acc_delta[:] += (1. - self.rho) * current_delta * current_delta
 
         # update weight
         weight[:] -= current_delta + wd * weight
@@ -1507,7 +1510,8 @@ class Adamax(Optimizer):
 
         # update m_t and u_t
         m_t, u_t = state
-        m_t[:] = self.beta1 * m_t + (1. - self.beta1) * grad
+        m_t[:] *= self.beta1
+        m_t[:] += (1. - self.beta1) * grad
         u_t[:] = maximum(self.beta2 * u_t, NDabs(grad))
 
         # update weight
@@ -1570,8 +1574,10 @@ class Nadam(Optimizer):
 
         # update m_t and v_t
         m_t, v_t = state
-        m_t[:] = self.beta1 * m_t + (1. - self.beta1) * grad
-        v_t[:] = self.beta2 * v_t + (1. - self.beta2) * grad * grad
+        m_t[:] *= self.beta1
+        m_t[:] += (1. - self.beta1) * grad
+        v_t[:] *= self.beta2
+        v_t[:] += (1. - self.beta2) * grad * grad
 
         grad_prime = grad / (1. - self.m_schedule)
         m_t_prime = m_t / (1. - m_schedule_next)