pangeamt-tea 0.2.33

TEA - Translation Engine Architect

TEA - Translation Engine Architect

A command line tool to create translation engine.

Install

First install pipx then (x being your python version):

pipx install pangeamt-tea

Usage

Step 1: Create a new project

tea new --customer customer --src_lang es --tgt_lang en --flavor automotion --version 2

This command will create the project directory structure:

├── customer_es_en_automotion_2
│   ├── config.yml
│   └── data

Then enter in the directory

cd customer_es_en_automotion_2

Step 2: Configuration

Tokenizer

A tokenizer can be applied to source and target

tea config tokenizer --src mecab  --tgt moses

To list all available tokenizer:

tea config tokenizer --help

if you would not like to use tokenizers you can run:

tea config tokenizer -s none -t none

Truecaser

tea config truecaser --src --tgt

if you would not like to use truecaser you can run:

tea config tokenizer

BPE / SentencePiece

For joint BPE:

tea config bpe -j

For not joint BPE:

 tea bpe -s -t

For using sentencepiece:

tea config bpe --sentencepiece 

and options --model_type TEXT (unigram) --vocab_size INTEGER (8000) if you would like to modify them from default

Processors

tea config processors -s "{processors}"

being processors a list of preprocesses and postprocesses.

To list all available processors:

tea config processors --list

In order to test the processors that will be applied you can run this script in the main TEA project directory:

debug_normalizers.py <config_file> <src_test> <tgt_test>

being config_file the yaml config and src_test and tgt_test the segments to test for source and target text.

Prepare

tea config prepare --shard_size 100000 --src_seq_length 400 --tgt_seq_length 400

Translation model

tea config translation-model -n onmt

Step 3:

Copy some multilingual ressources (.tmx, bilingual files, .af ) into the 'data' directory

Step 4: Run

Create workflow

tea worflow new

Clean the data passing the normalizers and validators:

tea workflow clean -n {clean_th} -d

being clean_th the number of threads.

Preprocess the data (split data in train, dev or test, tokenization, BPE):

tea workflow prepare -n {prepare_th} -s 3

being prepare_th the number of threads.

Training model

tea workflow train --gpu 0

if you do not want to use gpu do not use this parameter.

Evaluate model

tea workflow eval --step {step} --src file.src --ref file.tgt --log file.log --out file.out --gpu 0