app.py

import streamlit as st
from transformers import T5Tokenizer, T5ForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM, pipeline, logging
import torch
import os
import httpx

logging.set_verbosity_error()

# Create the app layout
st.header("Text Machine Translation")
input_text = st.text_input("Enter text to translate:")
# Create a list of options for the select box
options = ["German", "Romanian", "English", "French", "Spanish", "Italian"]
langs = {"English":"en", "Romanian":"ro", "German":"de", "French":"fr", "Spanish":"es", "Italian":"it"}
models = ["Helsinki-NLP", "t5-base", "t5-small", "t5-large", "Unbabel/Tower-Plus-2B", "Unbabel/TowerInstruct-Mistral-7B-v0.2", 'Google', 'Argos']

# Create two columns
scol, tcol = st.columns(2)
# Place select boxes in columns
with scol:
    sselected_language = st.selectbox("Source language:", options, index=0, placeholder="Select source language")
with tcol:
    tselected_language = st.selectbox("Target language:", options, index=1, placeholder="Select target language")
model_name = st.selectbox("Select a model:", models, index=0, placeholder="Select language model")

sl = langs[sselected_language]
tl = langs[tselected_language]

st.session_state["sselected_language"] = sselected_language
st.session_state["tselected_language"] = tselected_language
st.session_state["model_name"] = model_name

if model_name == 'Helsinki-NLP':
    try:
        model_name = f"Helsinki-NLP/opus-mt-{sl}-{tl}"
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
    except EnvironmentError:
        model_name = f"Helsinki-NLP/opus-tatoeba-{sl}-{tl}"
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
if model_name.startswith('t5'):
    tokenizer = T5Tokenizer.from_pretrained(model_name)
    model = T5ForConditionalGeneration.from_pretrained(model_name)

st.write("Selected language combination:", sselected_language, " - ", tselected_language, "Selected model:", model_name)
submit_button = st.button("Translate")
translated_textarea = st.text("")

# Handle the submit button click
if submit_button:
    if model_name.startswith('Helsinki-NLP'):
        prompt = input_text
        print(prompt)
        input_ids = tokenizer.encode(prompt, return_tensors='pt')
        # Perform translation
        output_ids = model.generate(input_ids)
        # Decode the translated text
        translated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
    elif model_name.startswith('Google'): 
        url = os.environ['GCLIENT'] + f'sl={sl}&tl={tl}&q={input_text}'
        response = httpx.get(url)
        translated_text = response.json()[0][0][0]
        print(response.json()[0][0])
    elif model_name.startswith('t5'):
        prompt = f'translate {sselected_language} to {tselected_language}: {input_text}'
        print(prompt)
        input_ids = tokenizer.encode(prompt, return_tensors='pt')
        # Perform translation
        output_ids = model.generate(input_ids)
        # Decode the translated text
        translated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
    if 'Unbabel' in model_name:   
        pipe = pipeline("text-generation", model=model_name, torch_dtype=torch.bfloat16, device_map="auto")
        # We use the tokenizer’s chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating
        messages = [{"role": "user",
                     "content": f"Translate the following text from {sselected_language} into {tselected_language}.\n{sselected_language}: {input_text}.\n{tselected_language}:"}]
        prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
        outputs = pipe(prompt, max_new_tokens=256, do_sample=False)
        translated_text = outputs[0]["generated_text"]
        start_marker = "<end_of_turn>"
        if start_marker in translated_text:
            translated_text = translated_text.split(start_marker)[1].strip()
        translated_text = translated_text.replace('Answer:', '').strip() if translated_text.startswith('Answer:') else translated_text
    if 'Argos' in model_name:   
        import argostranslate.package
        import argostranslate.translate
        
        from_code = sl
        to_code = tl
        
        # Download and install Argos Translate package
        argostranslate.package.update_package_index()
        available_packages = argostranslate.package.get_available_packages()
        package_to_install = next(
            filter(
                lambda x: x.from_code == from_code and x.to_code == to_code, available_packages
            )
        )
        argostranslate.package.install_from_path(package_to_install.download())
        
        # Translate
        translated_text = argostranslate.translate.translate(input_text, from_code, to_code)
        print(translated_text)
        
    # Display the translated text
    print(translated_text)
    st.write(f"Translated text from {sselected_language} to {tselected_language} using {model_name}:")
    translated_textarea = st.text(translated_text)