Metadata-Version: 2.1
Name: bntranslit
Version: 2.1.0
Summary: BNTRANSLIT is a deep learning based transliteration app for Bangla word
Home-page: https://github.com/sagorbrur/bntranslit
Author: Sagor Sarker
Author-email: sagorhem3532@gmail.com
License: MIT
Platform: UNKNOWN
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.6
Description-Content-Type: text/markdown
Requires-Dist: numpy
Requires-Dist: wasabi
Requires-Dist: wget

# BNTRANSLIT
__BNTRANSLIT__ is a deep learning based transliteration app for Bangla word.

## Installation
`pip install bntranslit`

## Dependency
- pytorch 1.7.0 or 1.7.0+

NB: No `GPU` Needed. Totally `CPU` based

## Pre-trained Model
- [Download bntranslit_model](https://github.com/sagorbrur/bntranslit/raw/master/model/bntranslit_model.pth)

## Usage

```py
from bntranslit import BNTransliteration


model_path = "bntranslit_model.pth"
bntrans = BNTransliteration(model_path)

word = "aami"
output = bntrans.predict(word, topk=10)
# output: ['আমি', 'আমী', 'অ্যামি', 'আমিই', 'এমি', 'আমির', 'আমিদ', 'আমই', 'আমে', 'আমিতে']

```

## Datasets and Training Details
- We used [Google Dakshina Dataset](https://github.com/google-research-datasets/dakshina)
- Thanks to [AI4Bharat](https://github.com/AI4Bharat/IndianNLP-Transliteration) for providing training notebook with details explanation
- We trained Google Bangla Dakshina lexicons train datasets for 10 epochs with batch size 128, 1e-3, embedding dim = 300, hidden dim = 512, lstm, used attention
- We evaluated our trained model with Google Bangla Dakshina lexicon test data using [AI4Bharat evaluation script](https://raw.githubusercontent.com/AI4Bharat/IndianNLP-Transliteration/jgeob-dev/tools/accuracy_reporter/accuracy_news.py) and our evaluation results insides `docs/evaluation_summary.txt`



