GitHub user KellenSunderland opened a pull request:
https://github.com/apache/incubator-joshua/pull/1
Performance Improvements to Joshua
Hello Joshua folks. I've got a series of patches to contribute to Joshua
including for the most part a lot of performance improvements.
**Performance Improvements**
The two main changes in terms of performance are that we've removed some
string parsing that was not strictly needed, and slowed decoding. We've also
added a post-deserialization LRU cache for objects that are commonly requested
from the packed grammar during decoding. Because the same objects are
requested over and over again from the packed grammar this reduces the cost of
unpacking them, allocating new objects, calculating their feature weights, etc.
It also dramatically reduces the work the garbage collector needs to do during
decoding.
**New Features**
In terms of new features we do add a small new feature that we will further
develop in future commits. Namely the ability to extract information via a
StructuredTranslation object that provides similar information to what is
displayed to stdout. This means if you are using Joshua as a library, the
consuming application will have programatic access to investing aspects to the
translate call, for example what the alignments in this translation were.
**New Dependancies**
These initial few patches add a new dependency, Google Guava 19.
http://mvnrepository.com/artifact/com.google.guava/guava/19.0
https://github.com/google/guava
LICENSE: https://github.com/google/guava/blob/master/COPYING (Apache 2.0)
**Testing**
We've also got a number of unit tests that we'll be contributing. In this
pull request there's only two, but there should be more to come in further
pulls. Currently these tests don't have a build target, but we can add that in
the future if you (the Joshua maintainers) would find them beneficial.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/KellenSunderland/joshua master
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/incubator-joshua/pull/1.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #1
----
commit 9501535dcd67b89e821fd686089f621c5721497f
Author: Felix Hieber <[email protected]>
Date: 2015-02-27T13:05:50Z
Decoder's Translation class now contains more members including the
possibility to store word alignment from the derivation. Allows use of Joshua
decoder class in a larger code project to extract information, rather than
relying on stdout. Also added a getter for JoshuaConfiguration in the Decoder.
commit 244e6936d8e3e7b30ebbe49ff7a9a2bd0c0c9994
Author: Felix Hieber <[email protected]>
Date: 2015-08-24T06:29:17Z
Viterbi information is now extracted from the hypergraph using a more
principled traversel functionality (WalkerFunction).
Also updated the unit tests.
commit e70677d2eab23daa7082173e6fe337d68aa12230
Author: Kellen Sunderland <[email protected]>
Date: 2015-09-22T11:37:54Z
Add an LRU cache from Google Guava to decrease allocations in the
PackedGrammer getRules() call
Results in a 1.5 times speedup in decoding and a large decrease in required
garbage collection
commit cabb52cabd5a81088b21b9e01a4668ebb2a85ffa
Author: Kellen Sunderland <[email protected]>
Date: 2015-10-14T15:13:18Z
kellens: Use Guava's memoize for expensive calls, removed unneeded members
fhieber: Important bugfix for obtaining word alignments from packedRules in
multi-threading environment
commit 5665f02ff0385db4f77bf4493db2d96bc63355d8
Author: Felix Hieber <[email protected]>
Date: 2015-12-01T12:34:47Z
Removed slow and redundant feature string parsing when constructing rules
from packed grammar (at sort time and at actual construction of feature vector).
Gets rid of String parsing features over and over again which turned out to
be slow in profiling. The solution is not perfect, but we get a nice speedup of
roughly a factor 5: If JoshuaConfiguration.amortize is set to false grammars
are forced to be sorted at decoder startup. Here are the stats: New code: Took
561.64 seconds to load pipeline. Old code: Took 2688.60 seconds to load
pipeline. Basically we are significantly reducing the time for sorting the
rules by getting rid of an intermediate string representation of the features
in a rule. Since String parsing of floats is removed now there was some float
precision change in the regression-test for which I changed the gold output.
This is fine.
commit cadd987c16ff012298b42074fb96bab8697fa84f
Author: Kellen Sunderland <[email protected]>
Date: 2016-03-29T13:46:27Z
Forced synchronization on method that still occasionally fails
multi-threaded test.
This is a fix for a very rare multithreading issue we've observed in
Joshua. We have a test that is able to reproduce the error fairly often when
run on a host with multiple physical cores. This patch fixes all errors seen
in both the patch and during runtime.
commit 2cc9996b4ed9e71ae4998a0db3eaef9586b0c69d
Author: Felix Hieber <[email protected]>
Date: 2016-03-29T14:55:07Z
Remove sorting which may rely on LOCALE of machine
Here we fixed an integration test that will fail on any machine with a
console locale set to a euro-based numbering system. So for example de-DE and
fr-FR locales would fail this test.
commit 9448ba552cd03bacad81eb4b9b5e900db360c00e
Author: Kellen Sunderland <[email protected]>
Date: 2016-03-29T15:23:23Z
Clean up Slice constructor, Fully loading source tries, lazy loading other
structures
----
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
