File size: 6,017 Bytes
7453da0
c657020
540056e
 
e8fd75c
5301c9c
 
540056e
 
5301c9c
e8fd75c
540056e
d6695b2
e8fd75c
540056e
5301c9c
 
 
 
a2c09c5
5301c9c
 
a15b3ce
99bd104
 
c657020
 
5e6dba7
 
 
 
 
 
 
c657020
a15b3ce
bf18300
a15b3ce
 
bf18300
a15b3ce
 
bf18300
a15b3ce
bf18300
540056e
80aa4e5
508045d
 
 
 
 
455006b
05d65fa
 
43a6e4c
72f0d04
 
 
9375fef
318d728
bb80675
72f0d04
 
b8b8031
9375fef
b8b8031
bb80675
72f0d04
455006b
 
efc72c4
508045d
fc0768e
508045d
 
 
540056e
09d059c
540056e
e8fd75c
f7b9ef5
e8fd75c
 
20fea69
508045d
e8fd75c
508045d
80aa4e5
 
 
e8fd75c
 
5bddbaf
508045d
23bb5b3
 
b9bed89
 
 
 
9375fef
b9bed89
 
 
 
 
 
 
23bb5b3
b9bed89
 
776a974
 
 
 
 
 
 
 
 
 
5bddbaf
09d059c
5bddbaf
 
776a974
5bddbaf
c6466be
776a974
 
 
306eb01
 
 
 
e0b6034
a2c4d07
 
 
306eb01
 
 
5bddbaf
306eb01
 
b9bed89
 
 
 
 
 
 
5bddbaf
b9bed89
 
 
 
17ba671
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
import gradio as gr
import json
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/")
kosmos2_client = Client("https://ydshieh-kosmos-2.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
    )
    """

    kosmos2_result = kosmos2_client.predict(
        image_in,	# str (filepath or URL to image) in 'Test Image' Image component
        "Detailed",	# str in 'Description Type' Radio component
        fn_index=4
    )

    print(f"KOSMOS2 RETURNS: {kosmos2_result}")

    with open(kosmos2_result[1], 'r') as f:
        data = json.load(f)
    
    reconstructed_sentence = []
    for sublist in data:
        reconstructed_sentence.append(sublist[0])

    full_sentence = ' '.join(reconstructed_sentence)
    print(full_sentence)

    # Find the last occurrence of "."
    #last_period_index = full_sentence.rfind('.')

    # Truncate the string up to the last period
    #truncated_caption = full_sentence[:last_period_index + 1]

    # print(truncated_caption)
    #print(f"\n—\nIMAGE CAPTION: {truncated_caption}")
    
    return full_sentence

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: Let's say 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: Let's say You are sophisticated old man, also know as the Gentleman's Companion. 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 Kosmos2...")
    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 user_prompt, 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",
                    elem_id = "image-in"
                )
                submit_btn = gr.Button("Make LLM system from my pic !")
            with gr.Column():
                caption = gr.Textbox(
                    label = "Image caption (Kosmos2)",
                    elem_id = "image-caption"
                )
                result = gr.Textbox(
                    label = "Suggested System",
                    lines = 6,
                    max_lines = 30,
                    elem_id = "suggested-system-prompt"
                )
        with gr.Row():
            gr.Examples(
                examples = [
                    ["examples/ocean_poet.jpeg"],
                    ["examples/winter_hiking.png"],
                    ["examples/teatime.jpeg"],
                    ["examples/news_experts.jpeg"],
                    ["examples/chicken_adobo.jpeg"]
                ],
                fn = infer,
                inputs = [image_in],
                outputs = [caption, result],
                cache_examples = True
            )

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

demo.queue().launch(show_api=False)