translation.py

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

# Define supported languages
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")
}

# Load a specific translation model pair with caching
@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 as e:
        return None, None  # Suppress error message, return None for fallback

# Load all possible model combinations with caching
@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

# Preload all models
all_models = _load_all_models()

# Perform combined translation through intermediate languages
def combined_translate(text, source_lang, target_lang, default_tokenizer, default_model):
    try:
        if source_lang == target_lang:  # No translation needed if languages are same
            return text
        if source_lang != "en":
            src_to_inter_tokenizer, src_to_inter_model = None, None
            for inter in ["en", "fr", "es", "de", "ru"]:  # Try multiple intermediates
                pair = all_models.get((source_lang, inter))
                if pair and pair[0] and pair[1]:
                    src_to_inter_tokenizer, src_to_inter_model = pair
                    break
            inter_text = src_to_inter_tokenizer.decode(src_to_inter_model.generate(**src_to_inter_tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=500))[0], skip_special_tokens=True) if src_to_inter_tokenizer else text
        else:
            inter_text = text
        if target_lang != "en":
            inter_to_tgt_tokenizer, inter_to_tgt_model = None, None
            for inter in ["en", "fr", "es", "de", "ru"]:
                pair = all_models.get((inter, target_lang))
                if pair and pair[0] and pair[1]:
                    inter_to_tgt_tokenizer, inter_to_tgt_model = pair
                    break
            translated = inter_to_tgt_tokenizer.decode(inter_to_tgt_model.generate(**inter_to_tgt_tokenizer(inter_text, return_tensors="pt", padding=True, truncation=True, max_length=1000))[0], skip_special_tokens=True) if inter_to_tgt_tokenizer else inter_text
            return translated if translated.strip() else text
        return inter_text
    except Exception:
        return text  # Suppress error, return source text

# Class to handle combined translation
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):
        try:
            input_ids = kwargs.get('input_ids')
            if not input_ids or input_ids.size(0) == 0:
                return torch.tensor([])
            inputs = self.default_tokenizer.batch_decode(input_ids, skip_special_tokens=True)
            translated_texts = [combined_translate(text, self.source_lang, self.target_lang, self.default_tokenizer, self.default_model) for text in inputs]
            encoded_outputs = [self.default_tokenizer.encode(t, return_tensors="pt", padding=True, truncation=True, max_length=500)[0] for t in translated_texts]
            if not encoded_outputs:
                return torch.tensor([])
            return torch.stack(encoded_outputs)  # Stack tensors to ensure proper shape
        except Exception:
            return torch.tensor([])  # Suppress error, return empty tensor

# Load appropriate translation model with caching
@st.cache_resource
def load_model(source_lang, target_lang):
    try:
        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
        for inter in LANGUAGES.keys():
            if inter != source_lang and inter != target_lang:
                pair1 = all_models.get((source_lang, inter))
                pair2 = all_models.get((inter, target_lang))
                if pair1 and pair1[0] and pair1[1] and pair2 and pair2[0] and pair2[1]:
                    return pair1
        default_tokenizer, default_model = _load_default_model()
        return default_tokenizer, CombinedModel(source_lang, target_lang, default_tokenizer, default_model)
    except Exception:
        raise  # Allow higher-level handling if needed

# Load default translation model with caching
@st.cache_resource
def _load_default_model():
    try:
        model_name = "Helsinki-NLP/opus-mt-en-hi"
        tokenizer = MarianTokenizer.from_pretrained(model_name)
        model = MarianMTModel.from_pretrained(model_name)
        return tokenizer, model
    except Exception:
        raise  # Allow higher-level handling if needed

# Translate text with caching
@st.cache_data
def translate(text, source_lang, target_lang):
    try:
        if not text:
            return ""
        tokenizer, model = load_model(source_lang, target_lang)
        inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=500)
        if inputs['input_ids'].size(0) > 1:
            inputs = {k: v[0].unsqueeze(0) for k, v in inputs.items()}
        with torch.no_grad():
            translated_ids = model.generate(**inputs, max_length=1000 if target_lang in ["hi", "zh", "ja"] else 500, num_beams=4, early_stopping=True)
        result = tokenizer.decode(translated_ids[0], skip_special_tokens=True) if translated_ids.size(0) > 0 else None
        if result and result.strip():
            return result
        else:
            st.warning("This translation is not possible at this moment. Please try another language.")
            return text  # Return source text without additional note
    except Exception:
        st.warning("This translation is not possible at this moment. Please try another language.")
        return text  # Return source text without additional note