Fairseq-Based Machine Translation

This function uses the Fairseq library to perform machine translation. It takes an input sentence and source and target languages, and then returns the translated sentence.

Technology Stack : Fairseq, PyTorch

Code Type : Machine translation

Code Difficulty : Intermediate

                
                    
def translate_sentence(input_sentence, source_lang, target_lang):
    import torch
    from fairseq.models import FairseqModel
    from fairseq.data.encoders import BaseSourceEncoder
    from fairseq.data import Dictionary

    # Load the model, dictionary and source encoder
    model_path = 'path_to_pretrained_model'
    model = FairseqModel.from_pretrained(model_path)
    src_dict = Dictionary.load(model_path + '/dict.txt')
    src_encoder = BaseSourceEncoder.from_pretrained(model_path, 'src')

    # Encode the input sentence
    src_tokens = src_encoder.encode(input_sentence)
    src_tokens = torch.tensor(src_tokens).unsqueeze(0)

    # Translate the encoded sentence
    with torch.no_grad():
        translated_tokens = model.translate(src_tokens)

    # Decode the translated tokens to a sentence
    translated_sentence = src_dict.decode(translated_tokens[0])

    return translated_sentence