From 73154a2f45c0a081b8e719301bdccb34fdedadd0 Mon Sep 17 00:00:00 2001
From: Thomas Young <35565423+HexToString@users.noreply.github.com>
Date: Thu, 22 Apr 2021 21:23:35 +0800
Subject: [PATCH] 123

321
---
 doc/LOD.md | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)
 create mode 100644 doc/LOD.md

diff --git a/doc/LOD.md b/doc/LOD.md
new file mode 100644
index 00000000..4e20c495
--- /dev/null
+++ b/doc/LOD.md
@@ -0,0 +1,32 @@
+# Lod Introduction
+
+(English|[简体中文](LOD_CN.md))
+
+## Principle
+
+LoD(Level-of-Detail) Tensor is an advanced feature of paddle and an extension of tensor. LoD Tensor improves training efficiency by sacrificing flexibility.
+
+**Notice：** For most users, there is no need to pay attention to the usage of LoD Tensor. Currently, serving only supports the usage of one-dimensional LOD.
+
+
+## Use
+
+**Prerequisite：** Your prediction model needs to support variable length tensor input.
+
+
+Take the visual task as an example. In the visual task, we often need to process video and image. These elements are high-dimensional objects. 
+Suppose that an existing Mini batch contains three videos, each video contains three frames, one frame and two frames respectively.
+If each frame has the same size: 640x480, the mini batch can be expressed as:
+```
+3     1  2
+口口口 口 口口
+```
+The size of the bottom tenor is (3 + 1 + 2) x640x480, and each 口 represents a 640x480 image.
+
+Then, the shape of tensor is [6,640,480], lod=[0,3,4,6].
+
+Where 0 is the starting value and 3-0 = 3; 4-3=1; 6-4 = 2, these three values just represent your variable length information. 
+
+The last element 6 in LOD should be equal to the total length of the first dimension in shape.
+
+The variable length information recorded in LOD and the first dimension information of shape in tensor should be aligned in the above way.
-- 
GitLab