app.py

import gradio as gr
import tempfile
from transformers import  MT5ForConditionalGeneration, MT5Tokenizer,ViltProcessor, ViltForQuestionAnswering, AutoTokenizer
import torch
from PIL import Image

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# English to Persian model
fa_en_translation_tokenizer = MT5Tokenizer.from_pretrained("SeyedAli/Persian-to-English-Translation-mT5-V1")
fa_en_translation_model = MT5ForConditionalGeneration.from_pretrained("SeyedAli/Persian-to-English-Translation-mT5-V1").to(device)

def run_fa_en_transaltion_model(input_string, **generator_args):
    input_ids = fa_en_translation_tokenizer.encode(input_string, return_tensors="pt")
    res = fa_en_translation_model.generate(input_ids, **generator_args)
    output = fa_en_translation_tokenizer.batch_decode(res, skip_special_tokens=True)
    return output

# Persian to English model
en_fa_translation_tokenizer = MT5Tokenizer.from_pretrained("SeyedAli/English-to-Persian-Translation-mT5-V1")
en_fa_translation_model = MT5ForConditionalGeneration.from_pretrained("SeyedAli/English-to-Persian-Translation-mT5-V1").to(device)

def run_en_fa_transaltion_model(input_string, **generator_args):
    input_ids = en_fa_translation_tokenizer.encode(input_string, return_tensors="pt")
    res = en_fa_translation_model.generate(input_ids, **generator_args)
    output = en_fa_translation_tokenizer.batch_decode(res, skip_special_tokens=True)
    return output

# Visual Question Answering model
VQA_processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-finetuned-vqa")
VQA_model = ViltForQuestionAnswering.from_pretrained("dandelin/vilt-b32-finetuned-vqa").to(device)


def VQA(image,text):
    with tempfile.NamedTemporaryFile(suffix=".png") as temp_image_file:
        # Copy the contents of the uploaded image file to the temporary file
        Image.fromarray(image).save(temp_image_file.name)
        # Load the image file using Pillow
        image = Image.open(temp_image_file.name)
        # prepare inputs
        encoding = processor(image, run_fa_en_transaltion_model(text), return_tensors="pt")
        # forward pass
        outputs = model(**encoding)
        logits = outputs.logits
        idx = logits.argmax(-1).item()
        output=[]
        for item in model.config.id2label[idx]:
            output.append(run_en_fa_transaltion_model(item))
        return output

iface = gr.Interface(fn=VQA, inputs=["image","text"], outputs="text")
iface.launch(share=False)