From d8ff778bafce7394b45f555bbac35bf1bb45a077 Mon Sep 17 00:00:00 2001 From: Aston Zhang Date: Fri, 6 Apr 2018 21:54:20 +0000 Subject: [PATCH] wrap code --- chapter_optimization/adadelta-scratch.md | 3 ++- chapter_optimization/adam-gluon.md | 15 ++++++++------- 2 files changed, 10 insertions(+), 8 deletions(-) diff --git a/chapter_optimization/adadelta-scratch.md b/chapter_optimization/adadelta-scratch.md index c5368c8..1496590 100644 --- a/chapter_optimization/adadelta-scratch.md +++ b/chapter_optimization/adadelta-scratch.md @@ -36,7 +36,8 @@ def adadelta(params, sqrs, deltas, rho, batch_size): for param, sqr, delta in zip(params, sqrs, deltas): g = param.grad / batch_size sqr[:] = rho * sqr + (1. - rho) * nd.square(g) - cur_delta = nd.sqrt(delta + eps_stable) / nd.sqrt(sqr + eps_stable) * g + cur_delta = (nd.sqrt(delta + eps_stable) + / nd.sqrt(sqr + eps_stable) * g) delta[:] = rho * delta + (1. - rho) * cur_delta * cur_delta param[:] -= cur_delta ``` diff --git a/chapter_optimization/adam-gluon.md b/chapter_optimization/adam-gluon.md index 630f4f9..7380224 100644 --- a/chapter_optimization/adam-gluon.md +++ b/chapter_optimization/adam-gluon.md @@ -54,7 +54,14 @@ utils.optimize(batch_size=10, trainer=trainer, num_epochs=3, decay_epoch=None, * 回顾前面几章中你感兴趣的模型，将训练部分的优化算法替换成其他算法，观察并分析实验现象。 -## 优化章节回顾 +## 讨论 + +欢迎扫码直达[本节内容讨论区](https://discuss.gluon.ai/t/topic/2280)： + +![](../img/qr_adam-gluon.svg) + + +## 本章回顾 > 梯度下降可沉甸， 随机降低方差难。 @@ -73,9 +80,3 @@ utils.optimize(batch_size=10, trainer=trainer, num_epochs=3, decay_epoch=None, * 学率：学习率。 * 换：这个参数被替换掉。 * 权：指数加权移动平均。 - -## 讨论 - -欢迎扫码直达[本节内容讨论区](https://discuss.gluon.ai/t/topic/2280)： - -![](../img/qr_adam-gluon.svg) -- GitLab