Mailing-List: contact user-help@uima.apache.org; run by ezmlm
Precedence: bulk
Reply-To: user@uima.apache.org
From: =?utf-8?Q?Erik_F=C3=A4=C3=9Fler?= <erik.faessler@uni-jena.de>
Content-Type: multipart/alternative;
 boundary="Apple-Mail=_0D196569-A61A-419C-857D-AA8A27FE14DE"
Message-Id: <3AF412F9-7CCE-40E3-A2D3-6731EC9CFA05@uni-jena.de>
Mime-Version: 1.0 (Mac OS X Mail 9.0 \(3094\))
Subject: Re: Performance of UIMAfit JCasUtil.selectCovered() and variants
Date: Thu, 22 Oct 2015 11:05:01 +0200
References: <07371163-1101-4350-93DB-849B272935C6@uni-jena.de>
 <40AACD2C-F212-4FE8-9B69-BC89084518F9@apache.org>
To: user@uima.apache.org
In-Reply-To: <40AACD2C-F212-4FE8-9B69-BC89084518F9@apache.org>

--Apple-Mail=_0D196569-A61A-419C-857D-AA8A27FE14DE
Content-Transfer-Encoding: quoted-printable
Content-Type: text/plain;
	charset=utf-8

Small follow-up, just stumbled across by chance, did not even search for =
it:

=
http://searchivarius.org/blog/selectcovered-substantially-better-version-u=
ima-subiterator =
<http://searchivarius.org/blog/selectcovered-substantially-better-version-=
uima-subiterator>

Someone has done a performance comparison. It does not necessarily =
strike me to be the most sophisticated approach, but the code is =
available and one could use this as a hint.

Best,

Erik

> On 21 Oct 2015, at 18:26, Richard Eckart de Castilho <rec@apache.org> =
wrote:
>=20
> Hi,
>=20
> 1 uses the UIMA indexes which I believe use a binary search, so it =
should
> be something like O(log n).
>=20
> 2 is in principle O(n) but since it does a linear scan from the =
beginning
> and stops when no further annotations may be found, it practice O(n)=20=

> should be the upper bound when called for annotations towards the end =
of
> the document.
>=20
> 3 is fastest for repeated use. It should be O(n) for creating
> the index and then uses hashmap lookups.
>=20
> So 1 and 3 are better than two.
>=20
> If you need speed and need coverage information a lot, 3 should be the =
best.
>=20
> 1 and 2 are more convenient for coding.
>=20
> If you use plain UIMA and have type priorities set up, then using an
> iterator over sentences and a subiterator over tokens is likely to
> be better than 3 because it doesn't need the initial scan that 3 does.
>=20
> I'm not aware that anybody did extensive performance comparisons here.
> Some are being done in =
org.apache.uima.fit.util.JCasUtilTest.testSelectCoverRandom()
> which compares 1 and 2. Here a few lines from the test output (mind to =
increase the
> ITERATIONS variable if you try):
>=20
> ...
> Speed up factor 5.50 [naive:11 optimized:2 diff:9]
> Speed up factor 6.67 [naive:20 optimized:3 diff:17]
> Speed up factor 4.00 [naive:16 optimized:4 diff:12]
> Speed up factor 2.50 [naive:30 optimized:12 diff:18]
> Speed up factor 7.00 [naive:35 optimized:5 diff:30]
> Speed up factor 5.63 [naive:45 optimized:8 diff:37]
> Speed up factor 7.78 [naive:70 optimized:9 diff:61]
> Speed up factor 8.09 [naive:89 optimized:11 diff:78]
> ...
>=20
> Cheers,
>=20
> -- Richard
>=20
>> On 21.10.2015, at 17:07, Erik F=C3=A4=C3=9Fler =
<erik.faessler@uni-jena.de> wrote:
>>=20
>> Hi all,
>>=20
>> I=E2=80=99m wondering about the performance differences between
>>=20
>> 1) JCasUtil.selectCovered(JCas, Class<T>, AnnotationFS),
>> 2) JCasUtil.selectCovered(JCas, Class<T>, int, int) and
>> 3) JCasUtil.indexCovered(JCas, Class<T>, Class<S>)
>>=20
>> It is clear that 3) iterates once through the CAS and just returns a =
map. Once this is done, map access is swift.
>>=20
>> The Javadoc of 2) states that it is slower than 1).
>> 3) states that it is preferable to 2).
>>=20
>> Questions:
>> Is 3) also preferable over 2) when there is only one covering =
annotation or is the performance of 2) and 3) roughly equal then?
>> Main question: Is 3) also quicker than 1) if there are many covering =
annotations?
>>=20
>> Use case: I want to iterate through all sentences in paragraphs. =
Normally, I would use subiterators(), but the known type priority issue =
could be a problem for me. Should I just use 1)? Or would I still =
benefit from 3) if I have more than one paragraph?
>>=20
>> Thank you very much!
>>=20
>> Best,
>>=20
>> Erik
>=20


--Apple-Mail=_0D196569-A61A-419C-857D-AA8A27FE14DE--