app.py

import gradio as gr
from gradio_client import Client

#fusecap_client = Client("https://noamrot-fusecap-image-captioning.hf.space/")
fuyu_client = Client("https://adept-fuyu-8b-demo.hf.space/")

def get_caption(image_in):
    
    fuyu_result = fuyu_client.predict(
	    image_in,	# str representing input in 'raw_image' Image component
	    True,	# bool  in 'Enable detailed captioning' Checkbox component
		fn_index=2
    )
    print(f"IMAGE CAPTION: {fuyu_result}")
    return fuyu_result

import re
import torch
from transformers import pipeline

pipe = pipeline("text-generation", model="HuggingFaceH4/zephyr-7b-beta", torch_dtype=torch.bfloat16, device_map="auto")

agent_maker_sys = f"""
You are an AI whose job is to help users create their own chatbot whose personality will reflect the character or scene from an image described by users.
In particular, you need to respond succintly in a friendly tone, write a system prompt for an LLM, a catchy title for the chatbot, and a very short example user input. Make sure each part is included.
The system prompt will not mention any image provided.

For example, if a user says, "a picture of a man in a black suit and tie riding a black dragon", first do a friendly response, then add the title, system prompt, and example user input. 
Immediately STOP after the example input. It should be EXACTLY in this format:
"Sure, I'd be happy to help you build a bot! I'm generating a title, system prompt, and an example input. How do they sound?
Title: Dragon Trainer
System prompt: As a LLM, You are a Dragon trainer and your job is to provide guidance and tips on mastering dragons. Use a friendly and informative tone.
Example input: How can I train a dragon to breathe fire?"

Here's another example. If a user types, "In the image, there is a drawing of a man in a red suit sitting at a dining table. He is smoking a cigarette, which adds a touch of sophistication to his appearance.", respond: 
"Sure, I'd be happy to help you build a bot! I'm generating a title, system prompt, and an example input. How do they sound? 
Title: Gentleman's Companion
System prompt: Your a sophisticated old man. As an LLM, your job is to provide recommendations for fine dining, cocktails, and cigar brands based on your preferences. Use a sophisticated and refined tone. 
Example input: Can you suggest a good cigar brand for a man who enjoys smoking while dining in style?"
"""

instruction = f"""
<|system|>
{agent_maker_sys}</s>
<|user|>
"""

def infer(image_in):
    gr.Info("Getting image caption with Fuyu...")
    user_prompt = get_caption(image_in)
    
    prompt = f"{instruction.strip()}\n{user_prompt}</s>"    
    #print(f"PROMPT: {prompt}")
    
    gr.Info("Building a system according to the image caption ...")
    outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
    

    pattern = r'\<\|system\|\>(.*?)\<\|assistant\|\>'
    cleaned_text = re.sub(pattern, '', outputs[0]["generated_text"], flags=re.DOTALL)
    
    print(f"SUGGESTED LLM: {cleaned_text}")
    
    return cleaned_text.lstrip("\n")

title = f"LLM Agent from a Picture",
description = f"Get a LLM system prompt from a picture so you can use it in <a href='https://huggingface.co/spaces/abidlabs/GPT-Baker'>GPT-Baker</a>."

css = """
#col-container{
    margin: 0 auto;
    max-width: 780px;
    text-align: left;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.HTML(f"""
        <h2 style="text-align: center;">LLM Agent from a Picture</h2>
        <p style="text-align: center;">{description}</p>
        """)
        with gr.Row():
            with gr.Column():
                image_in = gr.Image(
                    label = "Image reference",
                    type = "filepath"
                )
                submit_btn = gr.Button("Make LLM system from my pic !")
            with gr.Column():
                result = gr.Textbox(
                    label ="Suggested System",
                    max_lines = 30
                )

    submit_btn.click(
        fn = infer,
        inputs = [
            image_in
        ],
        outputs =[
            result
        ]
    )

demo.queue().launch()