[
https://issues.apache.org/jira/browse/NLPCRAFT-67?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17127793#comment-17127793
]
Gleb edited comment on NLPCRAFT-67 at 6/7/20, 10:32 PM:
--------------------------------------------------------
Batching is finished. Current benchmarks are (done on my PC: AMD Ryzen 9 3900X
12-Core Processor; GeForce RTX 2070 SUPER Graphic card):
N | CUDA | CPU
1 | 58 | 141
5 | 79 | 479
100 | 500 | 3300
1000 | 5000 | 33000
Where N is number of sentences in batches, number is CUDA and CPU columns are
time in milliseconds. I believe that for big batches further optimization could
be done (CUDA), where batch processing by Bert model is small percentage. It
seems that big batches could not be processed fast on CPU. For example, out of
33 seconds of 1000 words processing on CPU, 30 seconds was spent on forwarding
Bert (computing weights).
was (Author: ifropc):
Batching is finished. Current benchmarks are (done on my PC: AMD Ryzen 9 3900X
12-Core Processor; GeForce RTX 2070 SUPER):
N | CUDA | CPU
1 | 58 | 141
5 | 79 | 479
100 | 500 | 3300
1000 | 5000 | 33000
Where N is number of sentences in batches, number is CUDA and CPU columns are
time in milliseconds. I believe that for big batches further optimization could
be done (CUDA), where batch processing by Bert model is small percentage. It
seems that big batches could not be processed fast on CPU. For example, out of
33 seconds of 1000 words processing on CPU, 30 seconds was spent on forwarding
Bert (computing weights).
> Python machine learning module
> -------------------------------
>
> Key: NLPCRAFT-67
> URL: https://issues.apache.org/jira/browse/NLPCRAFT-67
> Project: NLPCraft
> Issue Type: New Feature
> Reporter: Gleb
> Assignee: Gleb
> Priority: Major
> Fix For: 0.7.0
>
>
> Python part which consists of Bert masked word prediction and FastText
> synonyms filter
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
