You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@lucene.apache.org by "bugmakerrrrrr (via GitHub)" <gi...@apache.org> on 2024/03/14 17:14:12 UTC
[I] Avoid recalculating the norm of the target vector when using cosine metric [lucene]
bugmakerrrrrr opened a new issue, #13185:
URL: https://github.com/apache/lucene/issues/13185
### Description
Currently, in the KNN retrieval process, we use `VectorSimilarityFunction#compare` to calculate the score between the target vector and the current vector. This method requires recalculating the norm of the target vector each time. To avoid this repetition, we can pass the square norm of the target vector to score method. I suggest modifying the relevant interface as follows:
```
@functionalInterface
public interface ByteVectorScorer {
float score(byte[] vector);
}
@functionalInterface
public interface FloatVectorScorer {
float score(float[] vector);
}
public enum VectorSimilarityFunction {
EUCLIDEAN {
@Override
public ByteVectorScorer getVectorScorer(byte[] target) {
return vector -> 1 / (1f + squareDistance(target, vector));
}
@Override
public FloatVectorScorer getVectorScorer(float[] target) {
return vector -> 1 / (1 + squareDistance(target, vector));
}
},
COSINE {
@Override
public ByteVectorScorer getVectorScorer(byte[] target) {
int squareNorm = dotProduct(target, target);
return vector -> (1 + cosine(target, vector, squareNorm)) / 2;
}
@Override
public FloatVectorScorer getVectorScorer(float[] target) {
double squareNorm = dotProduct(target, target);
return vector -> Math.max((1 + cosine(target, vector, squareNorm)) / 2, 0);
}
};
public abstract ByteVectorScorer getVectorScorer(byte[] target);
public abstract FloatVectorScorer getVectorScorer(float[] target);
}
```
Any thoughts？
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [I] Avoid recalculating the norm of the target vector when using cosine metric [lucene]
Posted by "benwtrent (via GitHub)" <gi...@apache.org>.
benwtrent commented on issue #13185:
URL: https://github.com/apache/lucene/issues/13185#issuecomment-1997999092
I would rather not change anything related to this enumeration until we figure out: https://github.com/apache/lucene/issues/13182
As an aside, I think cosine as a metric is fairly useless. Folks using Lucene should just normalize everything and use dot product, or use max-inner-product.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [I] Avoid recalculating the norm of the target vector when using cosine metric [lucene]
Posted by "bugmakerrrrrr (via GitHub)" <gi...@apache.org>.
bugmakerrrrrr commented on issue #13185:
URL: https://github.com/apache/lucene/issues/13185#issuecomment-2074846147
close in favour of #13281
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [I] Avoid recalculating the norm of the target vector when using cosine metric [lucene]
Posted by "bugmakerrrrrr (via GitHub)" <gi...@apache.org>.
bugmakerrrrrr commented on issue #13185:
URL: https://github.com/apache/lucene/issues/13185#issuecomment-2002493642
> I would rather not change anything related to this enumeration until we figure out: #13182
got it.
> As an aside, I think cosine as a metric is fairly useless. Folks using Lucene should just normalize everything and use dot product, or use max-inner-product.
IMO, if we think that cosine is useless, we should remove it and guide users to use the metric we believe is correct. Otherwise, we should try to optimize it.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [I] Avoid recalculating the norm of the target vector when using cosine metric [lucene]
Posted by "bugmakerrrrrr (via GitHub)" <gi...@apache.org>.
bugmakerrrrrr closed issue #13185: Avoid recalculating the norm of the target vector when using cosine metric
URL: https://github.com/apache/lucene/issues/13185
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [I] Avoid recalculating the norm of the target vector when using cosine metric [lucene]
Posted by "bugmakerrrrrr (via GitHub)" <gi...@apache.org>.
bugmakerrrrrr commented on issue #13185:
URL: https://github.com/apache/lucene/issues/13185#issuecomment-2074848114
close in favour of #13281
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org