import os
import torch
from PIL import Image
import numpy as np
from PIL import Image
from lavis.models import load_model_and_preprocess
import gradio as gr

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


model, vis_processors, _ = load_model_and_preprocess(
    name="blip2_opt", model_type="pretrain_opt2.7b", is_eval=True, device=device
)


def answer_question(image, prompt):
    image = vis_processors["eval"](image).unsqueeze(0).to(device)
    response = model.generate({"image": image, "prompt": f"Question: {prompt} Answer:"})
    response = '\n'.join(response)
    return response

def generate_caption(image, caption_type):
    image = vis_processors["eval"](image).unsqueeze(0).to(device)
    
    if caption_type == "Beam Search":
        caption = model.generate({"image": image})
    else:
        caption = model.generate({"image": image}, use_nucleus_sampling=True, num_captions=3)
        
    caption = '\n'.join(caption)
    
    return caption


with gr.Blocks() as demo:
    
    gr.Markdown("## BLIP-2 Demo")
    gr.Markdown("Using `OPT2.7B` - [Github](https://github.com/salesforce/LAVIS/tree/main/projects/blip2) - [Paper](https://arxiv.org/abs/2301.12597)")
    
    
    with gr.Row():
        with gr.Column():
            input_image = gr.Image(label="Image", type="pil")
            caption_type = gr.Radio(["Beam Search", "Nucleus Sampling"], label="Caption Type", value="Beam Search")
            btn_caption = gr.Button("Generate Caption")
            
            question_txt = gr.Textbox(label="Question", lines=1)
            btn_answer = gr.Button("Generate Answer")
        
        with gr.Column():
            output_text = gr.Textbox(label="Answer", lines=5)
    
    btn_caption.click(generate_caption, inputs=[input_image, caption_type], outputs=[output_text])
    btn_answer.click(answer_question, inputs=[input_image, question_txt], outputs=[output_text])
    
    gr.Examples([['./merlion.png', 'Beam Search', 'which city is this?']], inputs=[input_image, caption_type, question_txt])
    
demo.launch()