repronets-k8s
is a hub for named entity transliteration (NET) models.
It hosts trained NET models and provides simple interface to make them accesssible. All models run as serverless services on Knative. Our platform is highly scalable and economical thanks to Knative's autoscaling (even to zero)
feature
Available models:
The languages we currently support:
- English to Arabic
- English to Chinese
- English to Hebrew
- English to Japanese
- English to Korean
- English to Russian
GUI provides a tool to easily and visually compare outputs across the Transformer
and the Phonetisaurus
with different beam sizes. It shows predicted sequence(s) along with the probability of each sequence.
Link: https://boxkey.me
ReproNETs provide a public API that provides transliteration as a service via simple HTTP request.
The query parameters are:
- input: Any English word to transliterate. It must be less than or 37 characters.
- language: The target language. It must be either
ara
,chi
,heb
,jpn
,kor
, orrus
. - beam: The size of beam search. It must be an integer from 1 to 5.
- model: The name of a transliteration model. It must be either
phonetisaurus
,transformer
orall
.
This is mainly used for a real-time prediction.
curl --request GET \
--url 'https://www.boxkey.me/repronets/api/v1/predict?input=rocky&language=jpn&beam=2&model=all'
{
"data": {
"phonetisaurus": {
"data": {
"No.1": {
"prob": 0.0000018659,
"tokens": "ロッキー"
},
"No.2": {
"prob": 1.32e-8,
"tokens": "ロッキ"
}
},
"status": 200,
"message": "Successfully made predictions"
},
"transformer": {
"data": {
"No.1": {
"prob": 0.7512609939,
"tokens": "ロッキー"
},
"No.2": {
"prob": 0.0077490897,
"tokens": "ロキー"
}
},
"status": 200,
"message": "Successfully made predictions"
}
},
"status": 200
}
Thank you for your interest in contributing to the repronets-k8s
codebase.
Models can be uploaded under the ./moedls
directory ideally as a docker container.
If you need help with containerizing your model, let us know at Issues
.
Copyright 2020- Kei Nemoto, All Rights Reserved
Licensed under the Apache License 2.0: http://www.apache.org/licenses/