app.py

import spaces
import torch
import re
import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from PIL import Image

if torch.cuda.is_available():
    device, dtype = "cuda", torch.float16
else:
    device, dtype = "cpu", torch.float32

model_id = "vikhyatk/moondream2"
revision = "2024-04-02"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
moondream = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, revision=revision, torch_dtype=dtype
).to(device=device)
moondream.eval()

@spaces.GPU
def answer_questions(image_tuples, prompt_text):
    result = ""
    Q_and_A = ""
    prompts = [p.strip() for p in prompt_text.split(',')]  # Splitting and cleaning prompts
    image_embeds = [img[0] for img in image_tuples if img[0] is not None]  # Extracting images from tuples, ignoring None

    print(f"\nprompts: {prompts}\n\n")
    answers = []
    for prompt in prompts:
        image_answers = moondream.batch_answer(
            images=[img.convert("RGB") for img in image_embeds],
            prompts=[prompt] * len(image_embeds),
            tokenizer=tokenizer,
        )
        answers.append(image_answers)
    
    for i, prompt in enumerate(prompts):
        Q_and_A += f"### Q: {prompt}\n"
        for j, image_tuple in enumerate(image_tuples):
            image_name = f"image{j+1}"
            answer_text = answers[i][j]  # Retrieve the answer for the i-th prompt for the j-th image
            Q_and_A += f"**{image_name} A:** \n *{answer_text}* \n\n"

    result = {'headers': prompts, 'data': answers}  # Updated result handling
    print(f"result\n{result}\n\nQ_and_A\n{Q_and_A}\n\n")
    return Q_and_A, result

with gr.Blocks() as demo:
    gr.Markdown("# moondream2 unofficial batch processing demo")
    gr.Markdown("1. Select images\n2. Enter one or more prompts separated by commas. Ex: Describe this image, What is in this image?\n\n")
    gr.Markdown("**Currently each image will be sent as a batch with the prompts thus asking each promp on each image**")
    gr.Markdown("*Running on free CPU space tier currently so results may take a bit to process compared to duplicating space and using GPU space hardware*")
    gr.Markdown("## 🌔 moondream2\nA tiny vision language model. [GitHub](https://github.com/vikhyatk/moondream)")
    with gr.Row():
        img = gr.Gallery(label="Upload Images", type="pil", preview=True, object_fit="scale-down")
    with gr.Row():
        prompt = gr.Textbox(label="Input Prompts", placeholder="Enter prompts (one prompt for each image provided) separated by commas. Ex: Describe this image, What is in this image?", lines=8)
    with gr.Row():
        submit = gr.Button("Submit")
    output = gr.Markdown(label="Questions and Answers", line_breaks=True)
    output2 = gr.Dataframe(label="Structured Dataframe", type="array", wrap=True)
    submit.click(answer_questions, [img, prompt], [output, output2])

demo.queue().launch()