app.py

import gradio as gr
from transformers import BlipProcessor, BlipForConditionalGeneration, MarianMTModel, MarianTokenizer
from PIL import Image
import torch

# Load translation model
translator_model_ar = MarianMTModel.from_pretrained("Helsinki-NLP/opus-mt-en-ar")
translator_tokenizer_ar = MarianTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-ar")

# Load BLIP model (fine-tuned)
model_path = "saja003/MuniVis"
processor_en = BlipProcessor.from_pretrained(model_path)
model_en = BlipForConditionalGeneration.from_pretrained(model_path)
model_en.eval()

# Function to describe image
def describe_image(image, language):
    if language == "Arabic":
        inputs = processor_en(image, return_tensors="pt")
        with torch.no_grad():
            out = model_en.generate(**inputs)
        description = processor_en.decode(out[0], skip_special_tokens=True)
        inputs_ar = translator_tokenizer_ar(description, return_tensors="pt")
        with torch.no_grad():
            translated_tokens = translator_model_ar.generate(**inputs_ar)
        arabic_description = translator_tokenizer_ar.decode(translated_tokens[0], skip_special_tokens=True)
        return arabic_description

    elif language == "English":
        inputs_en = processor_en(image, return_tensors="pt")
        with torch.no_grad():
            out_en = model_en.generate(**inputs_en)
        description_en = processor_en.decode(out_en[0], skip_special_tokens=True)
        return description_en

# Gradio UI
iface = gr.Interface(
    fn=describe_image,
    inputs=[
        gr.Image(type="pil", label="Upload an Image"),
        gr.Dropdown(choices=["Arabic", "English"], label="Select Language")
    ],
    outputs="text",
    title="Image Captioning with Arabic Translation",
    description="Select the language and upload an image to get a description."
)

if __name__ == "__main__":
    iface.launch()