app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

import fasttext
import urllib.request
import os

MODEL_PATH = "lid.176.ftz"

if not os.path.exists(MODEL_PATH):
    urllib.request.urlretrieve(
        "https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.ftz",
        MODEL_PATH
    )

lang_model = fasttext.load_model(MODEL_PATH)

def detect_language(text):
    text = text.strip().lower()

    # tiny Kinyarwanda keyword safety net
    rw_keywords = ["muraho", "amakuru", "neza", "murakoze", "yego", "oya"]
    if any(word in text for word in rw_keywords):
        return "rw"

    pred = lang_model.predict(text.replace("\n", " "))
    return pred[0][0].replace("__label__", "")

def auto_translate(text):
    lang = detect_language(text)

    if lang == "rw":
        return rw_to_en(text)

    if lang == "en":
        return text, lang

    return f"Detected '{lang}'. Only English and Kinyarwanda supported."

MODEL_ID = "mbazaNLP/Nllb_finetuned_general_en_kin"

print("Loading NLLB model once… please wait.")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID)

def translate(text, src_lang, tgt_lang):
    tokenizer.src_lang = src_lang
    inputs = tokenizer(text, return_tensors="pt")

    tokens = model.generate(
        **inputs,
        forced_bos_token_id=tokenizer.convert_tokens_to_ids(tgt_lang),
        max_length=200,
        repetition_penalty=1.2,
        length_penalty=1.0,
        do_sample=False, 
        num_beams=1
    )

    return tokenizer.batch_decode(tokens, skip_special_tokens=True)[0]

# Function 1: Kinyarwanda → English
def rw_to_en(text):
    return translate(text, "kin_Latn", "eng_Latn")

# Function 2: English → Kinyarwanda
def en_to_rw(text):
    return translate(text, "eng_Latn", "kin_Latn")

app = gr.TabbedInterface(
    [
        gr.Interface(fn=rw_to_en, inputs="text", outputs="text", title="Kiny → English"),
        gr.Interface(fn=en_to_rw, inputs="text", outputs="text", title="English → Kiny"),
        gr.Interface(fn=auto_translate, inputs="text", outputs="text", title="Auto Detect"),
    ],
    tab_names=["Kiny → English", "English → Kiny", "Auto Detect"]
)

app.launch()