A paper I was reading, http://www.cs.toronto.edu/~ilya/pubs/2011/LANG-RNN.pdf, uses bits per character as a test metric for estimating the quality of generative computer models of text but doesn't reference how it was calculated. Googling around, I can't really find anything about it.

Does anyone know how to calculate it? Python preferably, but pseudo-code or anything works. Thanks!

asked Jul 22 '13 at 21:40
Newmu
48619
 
   
Are you talking about the stuff defined in C CHAR_BITtigcc.ticalc.org/doc/limits.html#CHAR_BIT ? –  woozyking Jul 23 '13 at 0:31
   
Nope, this is related to information theory and entropy, not actual bit size. – Newmu Jul 23 '13 at 0:46

add comment

2 Answers

Bits per character is a measure of the performance of compression methods. It's applied by compressing a string and then measuring how many bits the compressed representation takes in total, divided by how many symbols (i.e. characters) there were in the original string. The fewer bits per character the compressed version takes, the more effective the compression method is.

In other words, the authors use their generative language model, among other things, for compression and make an assumption that a higheffectiveness of the resulting compression method indicates highaccuracy of the underlying generative model.

In section 1 they state:

The goal of the paper is to demonstrate the power of large RNNs trained with the new Hessian-Free optimizer by applying them to the task of predicting the next character in a stream of text. This is an important problem because a better character-level language model could improve compression of text files (Rissanen & Langdon, 1979) [...]

The Rissanen & Langdon (1979) article is the original description ofarithmetic coding, a well-known method for text compression.

Arithmetic coding operates on the basis of a generative language model, such as the one the authors have built. Given a (possibly empty) sequence of characters, the model predicts what character may come next. Humans can do that, too, for example given the input sequence hello w, we can guess probabilities for the next character: o has high probability (because hello world is a plausible continuation), but characters like h as in hello where can I find.. or i as in hello winston also have non-zero probability. So we can establish aprobability distribution of characters for this particular input, and that's exactly what the authors' generative model does as well.

This fits naturally with arithmetic coding: Given an input sequence that has already been encoded, the bit sequence for the next character is determined by the probability distribution of possible characters: Characters with high probability get a short bit sequence, characters with low probability get a longer sequence. Then the next character is read from the input and encoded using the bit sequence that was determined from the probability distribution. If the language model is good, the character will have been predicted with high probability, so the bit sequence will be short. Then the compression continues with the next character, again using the input so far to establish a probability distribution of characters, determining bit sequences, and then reading the actual next character and encoding it accordingly.

Note that the generative model is used in every step to establish a new probability distribution. So this is an instance of adaptive arithmetic coding.

After all input has been read and encoded, the total length (in bits) of the result is measured and divided by the number of characters in the original, uncompressed input. If the model is good, it will have predicted the characters with high accuracy, so the bit sequence used for each character will have been short on average, hence the total bits per character will be low.

Regarding ready-to-use implementations

I am not aware of an implementation of arithmetic coding that allows for easy integration of your own generative language model. Most implementations build their own adaptive model on-the-fly, i.e. they adjust character frequency tables as they read input.

One option for you may be to start with arcode. I looked at the code, and it seems as though it may be possible to integrate your own model, although it's not very easy. The self._ranges member represents the language model; basically as an array of cumulative character frequencies, so self._ranges[ord('d')] is the total relative frequency of all characters that are less than d (i.e. a, b, c if we assume lower-case alphabetic characters only). You would have to modify that array after every input character and map the character probabilities you get from the generative model to character frequency ranges.

answered Jul 23 '13 at 1:40
jogojapan
24k42859
 
2  
Excellent introduction and explanation, thank you! –  Newmu Jul 23 '13 at 4:00
   
The model I'm working on generates a probability distribution for the next character given a fixed length of previous characters, so it looks like it'll be some work but I should be able to figure it out. –  Newmu Jul 23 '13 at 4:14
   
@Newmu It should definitely be possible. But I agree it will be easy to make mistakes. You'll need to test this carefully after implementing it. (Unfortunately I have no insider knowledge about arcode. The code seemed to be relatively easy to understand, which is why I suggested this module. There are others, too, though. If you do a search for "python arithmetic coding" or similar, you may be able to find implementations more suitable.) –  jogojapan Jul 23 '13 at 4:31

add comment

The sys library has a getsizeof() function, this may be helpful?http://docs.python.org/dev/library/sys

answered Jul 23 '13 at 0:57
ChrisProsser
2,5691620
  add comment

Not the answer you're looking for? Browse other questions tagged python algorithm machine-learning nlp entropy or ask your own question.

asked

7 months ago

viewed

199 times

active

7 months ago

Hot Network Questions

more hot questions

site design / logo © 2014 stack exchange inc; user contributions licensed under cc by-sa 3.0 withattribution required
rev 2014.2.25.1396
 

How to calculate bits per character of a string? (bpc) to read的更多相关文章

  1. 209. First Unique Character in a String

    Description Find the first unique character in a given string. You can assume that there is at least ...

  2. PHP icov转码报错解决方法，iconv(): Detected an illegal character in input string

    iconv(): Detected an illegal character in input string 错误解决方法 //转码 function iconv_gbk_to_uft8($strin ...

  3. LeetCode_387. First Unique Character in a String

    387. First Unique Character in a String Easy Given a string, find the first non-repeating character ...

  4. PHP出现iconv(): Detected an illegal character in input string

    PHP传给JS字符串用ecsape转换加到url里,又用PHP接收,再用网上找的unscape函数转换一下,这样得到的字符串是UTF-8的,但我需要的是GB2312,于是用iconv转换 开始是这样用 ...

  5. [LeetCode] First Unique Character in a String 字符串第一个不同字符

    Given a string, find the first non-repeating character in it and return it's index. If it doesn't ex ...

  6. 387. First Unique Character in a String

    Given a string, find the first non-repeating character in it and return it's index. If it doesn't ex ...

  7. LeetCode 387. First Unique Character in a String

    Problem: Given a string, find the first non-repeating character in it and return it's index. If it d ...

  8. LeetCode First Unique Character in a String

    原题链接在这里:https://leetcode.com/problems/first-unique-character-in-a-string/ 题目: Given a string, find t ...

  9. leetcode修炼之路——387. First Unique Character in a String

    最近公司搬家了,有两天没写了,今天闲下来了,继续开始算法之路. leetcode的题目如下: Given a string, find the first non-repeating characte ...

随机推荐

  1. 面试题中关于String的常见操作

    题目1: 将用户输入的一段话,每个单词的首字母大写, 每个单词之间的空格调整为只有一个,遇到数字,将数字与后一个单词用下划线 "_" 进行连接 题目2:将 i @@ am @@@ ...

  2. 在vscode中，自定义代码片段，例vue组件的模板

    1---- 2----  输入vue,  选 vue.json 3----  在vue.json中编辑, 有说明 a.  tab符,要用空格, 也可以转义 4----   新建vue文件, 输入自定义 ...

  3. Type mismatch: cannot convert from javax.servlet.http.Cookie[] to org.apache.tomcat.util.http.parser.Cookie[] 的一种可能

    今天用到Cookie时,写了一个Cookie数组,发现报错“Type mismatch: cannot convert from javax.servlet.http.Cookie[] to org. ...

  4. LRU缓存机制

    运用你所掌握的数据结构,设计和实现一个  LRU (最近最少使用) 缓存机制.它应该支持以下操作: 获取数据 get 和 写入数据 put . 获取数据 get(key) - 如果密钥 (key) 存 ...

  5. Unity --- 如何降低UI的填充率

    1.首先简单介绍一下什么叫填充率: Fill Rate(填充率)是指显卡每帧或者说每秒能够渲染的像素数.在每帧绘制中,如果一个像素被反复绘制的次数越多,那么它占用的资源也必然更多.目前在移动设备上,F ...

  6. 有关C#中List排序的总结

    这里有一篇文章作者总结的就比较详细: https://blog.csdn.net/jimo_lonely/article/details/51711821 在这里只记录一点: 对list或者数组中的数 ...

  7. python中函数与函数式编程(二)

    首先要明白为什么要用到返回值,返回值的作用就是为了分情况来处理下面的程序(个人见解总结) 1.函数返回值 def test1(): pass def test2(): return 0 def tes ...

  8. （转）c# 属性与索引器

    属性是一种成员,它提供灵活的机制来读取.写入或计算私有字段的值. 属性可用作公共数据成员,但它们实际上是称为“访问器”的特殊方法. 这使得可以轻松访问数据,还有助于提高方法的安全性和灵活性. 一个简单 ...

  9. 文献导读 - Machine Learning Identifies Stemness Features Associated with Oncogenic Dedifferentiation

    参考: Machine Learning Identifies Stemness Features Associated with Oncogenic Dedifferentiation 前所未有!1 ...

  10. ASA与N6K对接

    ASA5545配置interface GigabitEthernet0/0 channel-group 10 mode active no nameif no security-level no ip ...