[torch]parallel

最新推荐文章于 2024-08-27 22:56:32 发布

三枚目

最新推荐文章于 2024-08-27 22:56:32 发布

阅读量1k

点赞数

CC 4.0 BY-SA版权

分类专栏： torch

本文链接：https://blog.youkuaiyun.com/apsvvfb/article/details/51013859

torch 专栏收录该内容

21 篇文章

订阅专栏

本文介绍了DataParallel模块在FBCunn中的使用方法及其内部工作原理。DataParallel通过将输入拆分为不同的列，并在这些列上并行运行相同的模型来实现数据并行处理。此外，还列举了一些未被充分记录的方法。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

examples: imagenet
https://github.com/soumith/imagenet-multiGPU.torch

http://facebook.github.io/fbcunn/fbcunn/#fbcunn.fbcunn.DataParallel.dok
fbcunn.DataParallel

DataParallel splits the input along separate columns, that run the same models on distinct partitions of the input.

Pictorially

                    +--------+
    column 1        |        |         column 3
       +------------+  Input +-------------+
       |            |        |             |
       |            +----+---+             |
       |                 |                 |
       |                 |                 |
  +----+---+        +----+---+        +----+---+
  |        |        |        |        |        |
  | Linear |        | Linear |        | Linear |       row 1
  |        |        |        |        |        |
  +----+---+        +----+---+        +----+---+
       |                 |                 |
       |                 |                 |
  +----+---+        +----+---+        +----+---+
  |        |        |        |        |        |
  |  Tanh  |        |  Tanh  |        |  Tanh  |       row 2
  |        |        |        |        |        |
  +----+---+        +----+---+        +----+---+
       |                 |                 |
       |                 |                 |
       |                 |                 |
       |            +----+---+             |
       |            |        |             |
       +------------+ Output +-------------+
                    |        |
                    +--------+

Undocumented methods

fbcunn.DataParallel:name()
fbcunn.DataParallel:updateGradInput(_input, gradOutput)
fbcunn.DataParallel:accUpdateGradParameters(_input, _gradOutput, lr)

https://github.com/clementfarabet/lua—parallel