Hi Andrew, If you are using Lucene 3.6.1, you can take a look at the method which creates a single byte value out of the received float using bit manipulation at [1]. There is also a 256-element decoder table in Similarity, where each byte corresponds to a decoded float value computed by [2].
The first method encodes 0.89f to byte 123. 123 is decoded to 0.85f via the second method, so it seems that the documentation is incorrect in this regard. [1] https://github.com/apache/lucene-solr/blob/lucene_solr_3_6_1/lucene/core/src/java/org/apache/lucene/util/SmallFloat.java#L75 [2] https://github.com/apache/lucene-solr/blob/lucene_solr_3_6_1/lucene/core/src/java/org/apache/lucene/util/SmallFloat.java#L88 On Thu, Mar 5, 2015 at 3:45 AM, wangdong <hrdxwa...@gmail.com> wrote: > thank you for your disscussion. > > I am a junior user of lucene, so i am not**familiar with some deep concept > you mentioned. > my question is simple. I just want to know how to get 0.75 from > decode(encode(0.89)) in offical document. > > why not 0.875? (0.875=0.5+0.25+0.125) > > thanks > andrew > > 在 2015/3/4 22:54, Adrien Grand 写道: >> >> Norms and doc values are indeed using the same API. However >> implementations differ a bit (eg. norms are stored in memory and use >> different compression schemes). >> >> The precision loss is up to the similarity. You could write a >> similarity impl which keeps full float precision, but scoring being >> fuzzy anyway this would multiply your memory needs for norms by 4 >> while not really improving the quality of the scores of your >> documents. This precision loss is the right trade-off for most >> use-cases. >> >> On Wed, Mar 4, 2015 at 3:04 PM, Ahmet Arslan <iori...@yahoo.com.invalid> >> wrote: >>> >>> Hi Adrien, >>> >>> I read somewhere that norms are stored using docValues. >>> In my understanding, docvalues can store lossless float values. >>> So the question is, why are still several decode/encode methods exist in >>> similarity implementations? >>> Intuitively switching to docvalues for norms should prevent precision >>> loss thing. >>> >>> Ahmet >>> >>> >>> On Wednesday, March 4, 2015 3:22 PM, Adrien Grand <jpou...@gmail.com> >>> wrote: >>> Hi, >>> >>> Floats require 32 bits but norms are encoded on a single byte. So >>> there is a precision loss when encoding float values into a single >>> byte. In your example, 0.75 and 0.89 are sufficiently close to each >>> other so that they are encoded to the same byte. >>> >>> >>> On Wed, Mar 4, 2015 at 4:48 AM, wangdong <hrdxwa...@gmail.com> wrote: >>>> >>>> I read the article about the scoring section in lucene as follows: >>>> >>>> Encoding and decoding of the resulted float norm in a single byte are >>>> done >>>> by the static methods of the class Similarity:encodeNorm() >>>> >>>> <http://lucene.apache.org/core/3_6_1/api/core/org/apache/lucene/search/Similarity.html#encodeNorm%28float%29>anddecodeNorm() >>>> >>>> <http://lucene.apache.org/core/3_6_1/api/core/org/apache/lucene/search/Similarity.html#decodeNorm%28byte%29>. >>>> Due to loss of precision, it is not guaranteed that decode(encode(x)) = >>>> x, >>>> e.g. decode(encode(0.89)) = 0.75. At scoring (search) time, this norm is >>>> brought into the score of document as*norm(t, d)*, as shown by the >>>> formula >>>> inSimilarity >>>> >>>> <http://lucene.apache.org/core/3_6_1/api/core/org/apache/lucene/search/Similarity.html>. >>>> >>>> I can not understand the formula decode(encode(0.89)) = 0.75 >>>> how can i get the 0.75 from the left. >>>> >>>> Is anyone can help me ? >>>> thanks ahead! >>>> >>>> andrew >>> >>> >>> >>> -- >>> Adrien >>> >>> --------------------------------------------------------------------- >>> To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org >>> For additional commands, e-mail: java-user-h...@lucene.apache.org >>> >>> --------------------------------------------------------------------- >>> To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org >>> For additional commands, e-mail: java-user-h...@lucene.apache.org >>> >> >> > -- András --------------------------------------------------------------------- To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org For additional commands, e-mail: java-user-h...@lucene.apache.org
