translation.py

import streamlit as st
from transformers import MarianTokenizer, MarianMTModel
import torch

LANGUAGES = {
    "en": ("English", "English"), "fr": ("Français", "French"), "es": ("Español", "Spanish"),
    "de": ("Deutsch", "German"), "hi": ("हिन्दी", "Hindi"), "zh": ("中文", "Chinese"),
    "ar": ("العربية", "Arabic"), "ru": ("Русский", "Russian"), "ja": ("日本語", "Japanese")
}

# Cache resource to load a specific translation model pair
@st.cache_resource
def _load_model_pair(source_lang, target_lang):
    try:
        model_name = f"Helsinki-NLP/opus-mt-{source_lang}-{target_lang}"
        tokenizer = MarianTokenizer.from_pretrained(model_name)
        model = MarianMTModel.from_pretrained(model_name)
        return tokenizer, model
    except Exception:
        return None, None

# Cache resource to load all possible model combinations
@st.cache_resource
def _load_all_models():
    models = {}
    for src in LANGUAGES.keys():
        for tgt in LANGUAGES.keys():
            if src != tgt:
                models[(src, tgt)] = _load_model_pair(src, tgt)
    return models

all_models = _load_all_models()

# Define combined_translate outside load_model with explicit parameters
def combined_translate(text, source_lang, target_lang, default_tokenizer, default_model):
    with torch.no_grad():
        if source_lang != "en":
            src_to_en_tokenizer, src_to_en_model = all_models.get((source_lang, "en"), (default_tokenizer, default_model))
            en_text = src_to_en_tokenizer.decode(src_to_en_model.generate(**src_to_en_tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=500))[0], skip_special_tokens=True)
        else:
            en_text = text
        if target_lang != "en":
            en_to_tgt_tokenizer, en_to_tgt_model = all_models.get(("en", target_lang), (default_tokenizer, default_model))
            return en_to_tgt_tokenizer.decode(en_to_tgt_model.generate(**en_to_tgt_tokenizer(en_text, return_tensors="pt", padding=True, truncation=True, max_length=1000))[0], skip_special_tokens=True)
        return en_text

# Class to handle combined translation through English pivot
class CombinedModel:
    def __init__(self, source_lang, target_lang, default_tokenizer, default_model):
        self.source_lang = source_lang
        self.target_lang = target_lang
        self.default_tokenizer = default_tokenizer
        self.default_model = default_model

    def generate(self, **kwargs):
        input_ids = kwargs.get('input_ids')
        if not input_ids:
            return torch.tensor([])
        return torch.tensor([combined_translate(tokenizer.decode(x, skip_special_tokens=True), self.source_lang, self.target_lang, self.default_tokenizer, self.default_model) for x in input_ids])

# Function to load appropriate translation model with optimized caching
@st.cache_resource
def load_model(source_lang, target_lang):
    if source_lang == target_lang:
        return _load_default_model()
    model_key = (source_lang, target_lang)
    tokenizer_model_pair = all_models.get(model_key)
    if tokenizer_model_pair and tokenizer_model_pair[0] and tokenizer_model_pair[1]:
        return tokenizer_model_pair
    # Prefer direct model if available, then pivot
    for src in [source_lang, "en"]:
        for tgt in [target_lang, "en"]:
            if src != tgt:
                pair = all_models.get((src, tgt))
                if pair and pair[0] and pair[1]:
                    return pair
    default_tokenizer, default_model = _load_default_model()
    return default_tokenizer, CombinedModel(source_lang, target_lang, default_tokenizer, default_model)

# Cache resource to load default translation model
@st.cache_resource
def _load_default_model():
    model_name = "Helsinki-NLP/opus-mt-en-hi"
    tokenizer = MarianTokenizer.from_pretrained(model_name)
    model = MarianMTModel.from_pretrained(model_name)
    return tokenizer, model

# Cache translation results to improve speed
@st.cache_data
def translate(text, source_lang, target_lang):
    if not text:
        return ""
    try:
        tokenizer, model = load_model(source_lang, target_lang)
        inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=500)
        with torch.no_grad():
            translated = model.generate(**inputs, max_length=1000 if target_lang == "hi" else 500, num_beams=4, early_stopping=True)  # Reduced to 4 beams for speed
        result = tokenizer.decode(translated[0], skip_special_tokens=True)
        return result if result.strip() else text
    except Exception as e:
        st.error(f"Translation error: {e}")
        return text