# -*- coding: utf-8 -*-
"""app.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1_HQHDuRl3mgto6slVIJGSlZ5DZeSs4El
"""

import torch
from transformers import pipeline
import gradio as gr

# Choose device: GPU if available, otherwise CPU. On Hugging Face Spaces, unless you explicitly pick a GPU runtime, you’re on CPU only
if torch.cuda.is_available():
    vqa = pipeline(
        task="visual-question-answering",
        model="Salesforce/blip-vqa-base",
        torch_dtype=torch.float16,#newer versions of TRANSFORMERS in Hugging face is torch_dtype not dtype. dtype is still working fine in Google Colab space
        device=0,          # GPU
        use_fast=False,
    )
else:
    vqa = pipeline(
        task="visual-question-answering",
        model="Salesforce/blip-vqa-base",
        device=-1,         # CPU
        use_fast=False,
    )

def answer_question(image, question):
    if not question:
        return "Please type a question about the image."
    # vqa returns a list of dicts like [{'score':..., 'answer':...}]
    result = vqa(question=question, image=image)
    return result[0]["answer"]

demo = gr.Interface(
    fn=answer_question,
    inputs=[
        gr.Image(type="pil", label="Upload an image"),
        gr.Textbox(label="Question", placeholder="e.g. What is the weather in this image?"),
    ],
    outputs=gr.Textbox(label="Answer"),
    title="BLIP Visual Question Answering",
    description="Ask a question about the uploaded image using Salesforce/blip-vqa-base.",
)

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