[
https://issues.apache.org/jira/browse/OPENNLP-830?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15103995#comment-15103995
]
Julien Subercaze edited comment on OPENNLP-830 at 1/17/16 11:53 PM:
--------------------------------------------------------------------
Hi there,
first of all, my bad, performance gain of this fix is not a 5 fold as I wrote
yesterday, I measured 1.65x improvement. I thought I deactivated all other
optims, but my code was such a mess ... Anyway I still have good news regarding
performance improvement.
To answer your question, I created a test project to measure performance impact
on both training and tagging. The project is here:
https://github.com/jsubercaze/opennlp-harness
and the README contains links to the required files.
I cannot release the data I'm working on and a major problem was to find a free
large dataset to train the model. I didn't found any, so I concatened several
ebooks from Gutenberg projects and created a 'silver' training set using the
en-maxent model.
Not so sure this is the right place, but it I introduced several other
optimizations that I imported/cleaned from my mess. The project is to be found
here (numbered 1.6.2-SNAPSHOT for testing purpose)
https://github.com/jsubercaze/opennlp-tools/commits/
Each optimization is a single commit, so you can move back
https://github.com/jsubercaze/opennlp-tools/commits/trunk
So here come the numbers (i5 2500K, Xms3G Xmx5G) :
*Building the dataset (tagging) on the Gutenberg small
1.6.0 :
- Exec time : 454502ms
- Throughput : 3295.87 sentences/sec
Hashtable fix (test pass)
- Exec time : 273929ms
- Throughput : 5468.50242456639
# Using Fast exponential (from commons.math3) in the eval method (test pass)
- Exec time : 214200ms
- Throughput : 6993.36 sentences/sec
*Training a POS Tagger on the Gutenberg small
- 1.6.0 : 356867ms
- HashTable fix : 326677ms
- Multithreading log model : 209315ms
- Writer thread for update : 201648ms
Multithreading log model : Maxent offers multhreading (must be cleaned), but
parameters is neither present and default value is 1. Change to use all cores
Writer thread for update : In the TwoPassDataIndexer, blocking I/O slows down
the process, move to non-blocking (Java 5 compatible way, there is way simpler
in >= 7).
was (Author: [email protected]):
Hi there,
first of all, my bad, performance gain of this fix is not a 5 fold as I wrote
yesterday, I measured 1.65x improvement. I thought I deactivated all other
optims, but my code was such a mess ... Anyway I still have good news regarding
performance improvement.
To answer your question, I created a test project to measure performance impact
on both training and tagging. The project is here:
https://github.com/jsubercaze/opennlp-harness
and the README contains links to the required files.
I cannot release the data I'm working on and a major problem was to find a free
large dataset to train the model. I didn't found any, so I concatened several
ebooks from Gutenberg projects and created a 'silver' training set using the
en-maxent model.
Not so sure this is the right place, but it I introduced several other
optimizations that I imported/cleaned from my mess. The project is to be found
here (numbered 1.6.2-SNAPSHOT for testing purpose)
https://github.com/jsubercaze/opennlp-tools/commits/
Each optimization is a single commit, so you can move back
https://github.com/jsubercaze/opennlp-tools/commits/trunk
So here come the numbers (i5 2500K, Xms3G Xmx5G) :
####Building the dataset (tagging) on the Gutenberg small
# 1.6.0 :
- Exec time : 454502ms
- Throughput : 3295.87 sentences/sec
# Hashtable fix (test pass)
- Exec time : 273929ms
- Throughput : 5468.50242456639
# Using Fast exponential (from commons.math3) in the eval method (test pass)
- Exec time : 214200ms
- Throughput : 6993.36 sentences/sec
###Training a POS Tagger on the Gutenberg small
# 1.6.0 : 356867ms
# HashTable fix : 326677ms
# Multithreading log model : 209315ms
# Writer thread for update : 201648ms
Multithreading log model : Maxent offers multhreading (must be cleaned), but
parameters is neither present and default value is 1. Change to use all cores
Writer thread for update : In the TwoPassDataIndexer, blocking I/O slows down
the process, move to non-blocking (Java 5 compatible way, there is way simpler
in >= 7).
> Huge runtime improvement on training (POS, Chunk, ...)
> ------------------------------------------------------
>
> Key: OPENNLP-830
> URL: https://issues.apache.org/jira/browse/OPENNLP-830
> Project: OpenNLP
> Issue Type: Improvement
> Components: Machine Learning, POS Tagger
> Affects Versions: 1.6.0
> Environment: Any
> Reporter: Julien Subercaze
> Labels: performance
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> opennlp.tools.ml.model.IndexHashTable is custom-made Hashtable that is used
> to store mapping index. This Hashtable is heavily used in openlp.tools.ml.*
> (i.e. every model) and leads to disastrous performance.
> This hashtable is probably legacy some legacy and is highly inefficient. A
> simple drop-in replacement by a java.util.HashMap wrapper solves the issue,
> doesn't break compatibility and does not add any dependency.
> Training a pos-tagger on a large dataset with custom tags, I see a factor 5
> improvement. It also seems to improve all ML models training pipeline.
> See :
> https://github.com/jsubercaze/opennlp/blob/trunk/opennlp-tools/src/main/java/opennlp/tools/ml/model/IndexHashTable.java
> For a quick fix.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
