未加星标

A Crude Batching Class in Python for PyTorch and CNTK

字体大小 | |
[开发(python) 所属分类 开发(python) | 发布者 店小二05 | 时间 2019 | 作者 红领巾 ] 0人收藏点击收藏

Writing code is a skill, and like all skills, requires practice. I write code almost every day mostly because I enjoy the mental exercise but also to stay sharp.

Today I thought I’d code up a crude batching class for possible use with the CNTK and PyTorch neural network libraries. Suppose you have 10 training items. It’s usual to process a subset (usually called a batch or mini-batch) of the training items at a time. Suppose the batch size is 4.

Training is an iterative process where processing all training items is called an epoch. So, if there are 10 training items and the batch size is 4, you can get two batches in an epoch, with two training items unprocessed. Suppose the number of epochs is set to 3.

My crude batcher class returns 4 indexes on each call to function next_batch(). These represent the items in a batch.


A Crude Batching Class in Python for PyTorch and CNTK

My batcher class is crude because it doesn’t check to make sure that a training item isn’t returned more than once in an epoch.

In the demo code, for the first of three epochs, the two batches are

[8 5 0 2]
[1 3 2 5] Training items [2] and [5] will be used twice in the training in one epoch. If the number of epochs is large (several hundred or thousand) this overlap effect won’t matter.

In the end, my crude batcher code isn’t that much simpler than a more sophisticated batcher class that prevents duplicates in an epoch. But it wasn’t a wasted effort because I learned some nuances of python (related to the None object). But mostly it was a fun exercise.


A Crude Batching Class in Python for PyTorch and CNTK

Physical exercise is a bit riskier than coding/mental exercise. Left: Sometimes the picture tells the whole story. Right: I’m no expert, but my intuition tells me that exercising on a treadmill while wearing high heels is difficult.

本文开发(python)相关术语:python基础教程 python多线程 web开发工程师 软件开发工程师 软件开发流程

代码区博客精选文章
分页:12
转载请注明
本文标题:A Crude Batching Class in Python for PyTorch and CNTK
本站链接:https://www.codesec.net/view/627754.html


1.凡CodeSecTeam转载的文章,均出自其它媒体或其他官网介绍,目的在于传递更多的信息,并不代表本站赞同其观点和其真实性负责;
2.转载的文章仅代表原创作者观点,与本站无关。其原创性以及文中陈述文字和内容未经本站证实,本站对该文以及其中全部或者部分内容、文字的真实性、完整性、及时性,不作出任何保证或承若;
3.如本站转载稿涉及版权等问题,请作者及时联系本站,我们会及时处理。
登录后可拥有收藏文章、关注作者等权限...
技术大类 技术大类 | 开发(python) | 评论(0) | 阅读(208)