app.py

import gradio as gr
from transformers import BlipProcessor, BlipForConditionalGeneration, NllbTokenizer, AutoModelForSeq2SeqLM
from PIL import Image
import torch

# Load model 1: English image captioning
blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")

# Load model 2: Translate EN → VI
translator_tokenizer = NllbTokenizer.from_pretrained("facebook/nllb-200-distilled-600M")
translator_model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M")
lang_code = "vie_Latn"

# Manually map lang_code to token ID (based on tokenizer's config)
lang_token_id = translator_tokenizer.convert_tokens_to_ids(lang_code)

def caption_translate(image):
    # Step 1: Get English caption
    inputs = blip_processor(image, return_tensors="pt")
    out = blip_model.generate(**inputs)
    eng_caption = blip_processor.decode(out[0], skip_special_tokens=True)

    # Step 2: Translate to Vietnamese
    inputs = translator_tokenizer(eng_caption, return_tensors="pt", src_lang="eng_Latn")
    translated = translator_model.generate(
        **inputs,
        forced_bos_token_id=lang_token_id,
        max_length=100
    )
    vi_caption = translator_tokenizer.decode(translated[0], skip_special_tokens=True)

    return vi_caption

iface = gr.Interface(
    fn=caption_translate,
    inputs=gr.Image(type="pil"),
    outputs="text",
    title="🧠 AI Mô Tả Hình Ảnh Bằng Tiếng Việt",
    description="Upload ảnh, hệ thống sẽ mô tả nội dung bằng tiếng Việt bằng cách kết hợp 2 mô hình: caption → translate"
)

iface.launch()