app.py

import gradio as gr
import spaces
import torch
from transformers import T5Tokenizer, T5ForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
import languagecodes

favourite_langs = {"German": "de", "Romanian": "ro", "English": "en", "-----": "-----"}
all_langs = languagecodes.iso_languages

# Language options as list, add favourite languages first
options = list(favourite_langs.keys())
options.extend(list(all_langs.keys()))
models = ["Helsinki-NLP", "t5-base", "t5-small", "t5-large",
          "facebook/nllb-200-distilled-600M",
          "facebook/nllb-200-distilled-1.3B",
          "facebook/mbart-large-50-many-to-many-mmt"]

def model_to_cuda(model):
    # Move the model to GPU if available
    if torch.cuda.is_available():
        model = model.to('cuda')
        print("CUDA is available! Using GPU.")
    else:
        print("CUDA not available! Using CPU.")
    return model

@spaces.GPU
def translate_text(input_text, sselected_language, tselected_language, model_name):
    sl = all_langs[sselected_language]
    tl = all_langs[tselected_language]
    message_text = f'Translated from {sselected_language} to {tselected_language} with {model_name}'
    if model_name == "Helsinki-NLP":
        try:
            model_name = f"Helsinki-NLP/opus-mt-{sl}-{tl}"
            tokenizer = AutoTokenizer.from_pretrained(model_name)
            model = model_to_cuda(AutoModelForSeq2SeqLM.from_pretrained(model_name))
        except EnvironmentError:
            try:   
                model_name = f"Helsinki-NLP/opus-tatoeba-{sl}-{tl}"
                tokenizer = AutoTokenizer.from_pretrained(model_name)
                model = model_to_cuda(AutoModelForSeq2SeqLM.from_pretrained(model_name))
            except EnvironmentError as error:
                return f"Error finding model: {model_name}! Try other available language combination.", error
    
    if 'nllb' in model_name:
        tokenizer = AutoTokenizer.from_pretrained(model_name, src_lang=languagecodes.nllb_language_codes[sselected_language])
        model = AutoModelForSeq2SeqLM.from_pretrained(model_name, device_map="auto")
        translator = pipeline('translation', model=model, tokenizer=tokenizer, src_lang=languagecodes.nllb_language_codes[sselected_language], tgt_lang=languagecodes.nllb_language_codes[tselected_language])
        translated_text = translator(input_text, max_length=512)
        return translated_text[0]['translation_text'], message_text
    
    if model_name.startswith('facebook/mbart-large'):
        from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
        model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
        tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
        # translate source to target
        tokenizer.src_lang = languagecodes.mbart_large_languages[sselected_language]
        encoded = tokenizer(input_text, return_tensors="pt")
        generated_tokens = model.generate(
            **encoded,
            forced_bos_token_id=tokenizer.lang_code_to_id[languagecodes.mbart_large_languages[tselected_language]]
        )
        return tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0], message_text

    if model_name.startswith('t5'):
        tokenizer = T5Tokenizer.from_pretrained(model_name)
        model = T5ForConditionalGeneration.from_pretrained(model_name, device_map="auto")

    if model_name.startswith("Helsinki-NLP"):
        prompt = input_text
    else:
        prompt = f"translate {sselected_language} to {tselected_language}: {input_text}"

    input_ids = tokenizer.encode(prompt, return_tensors="pt")
    output_ids = model.generate(input_ids, max_length=512)
    translated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
    
    message_text = f'Translated from {sselected_language} to {tselected_language} with {model_name}'
    print(f'Translating from {sselected_language} to {tselected_language} with {model_name}:', f'{input_text} =  {translated_text}', sep='\n')
    return translated_text, message_text

# Define a function to swap dropdown values
def swap_languages(src_lang, tgt_lang):
    return tgt_lang, src_lang 

def create_interface():
    with gr.Blocks() as interface:
        gr.Markdown("## Machine Text Translation")

        with gr.Row():
            input_text = gr.Textbox(label="Enter text to translate:", placeholder="Type your text here...")
        
        with gr.Row():
            sselected_language = gr.Dropdown(choices=options, value = options[0], label="Source language", interactive=True)
            tselected_language = gr.Dropdown(choices=options, value = options[1], label="Target language", interactive=True)
            swap_button = gr.Button("Swap Languages")
            swap_button.click(fn=swap_languages, inputs=[sselected_language, tselected_language], outputs=[sselected_language, tselected_language])

        model_name = gr.Dropdown(choices=models, label="Select a model", value = models[4], interactive=True)
        translate_button = gr.Button("Translate")

        translated_text = gr.Textbox(label="Translated text:", placeholder="Display field for translation", interactive=False, show_copy_button=True)
        message_text = gr.Textbox(label="Messages:", placeholder="Display field for status and error messages", interactive=False)

        translate_button.click(
            translate_text, 
            inputs=[input_text, sselected_language, tselected_language, model_name], 
            outputs=[translated_text, message_text]
        )

    return interface

interface = create_interface()
interface.launch()