fe1ixxu / better_mmt Goto Github PK

This multilingual machine translation model is trained on 12 layers with 4 experts and total 73 million parallel dataset. One can download mmodel from google drive: https://drive.google.com/file/d/1pDH3zueCK9t4lHEvWsqjEPzh34PbISZY/view?usp=sharing or from our local machine if you have access: /brtx/601-nvme1/haoranxu/BETTER_MMT/models/m6_moe_eom_4_xxl/

conda create -n better-moe python=3.8
git clone https://github.com/facebookresearch/fairscale.git
cd fairscale
git checkout origin/experts_lt_gpus_moe_reload_fix
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=10.2 -c pytorch
pip install -e ./

To translate from any languauge from or to English, run the command (Note that you should have at least 2^n(n>1) GPUs to use the model to translate, e.g., 2, 4, 8, 16....):

bash translate.sh ${INPUT} ${OUTPUT} ${SRC_LANG} ${TGT_LANG} ${MODEL_PATH}

Five inputs should be:

• ${INPUT}: input file that is needed to be translated
• ${OUTPUT}: output file
• ${SRC_LANG}: source language id
• ${TGT_LANG}: target language id
• ${MODEL_PATH}: Model path, now located at /brtx/601-nvme1/haoranxu/BETTER_MMT/models/m6_moe_eom_4_xxl/

Note that language id only supports languages that used in BETTER:

• fas: Farsi
• zho_Hans: Chinese (Simplified)
• rus: Russian
• kor: Korean
• ara_Arab: Arabic
• eng: English