File size: 6,862 Bytes
7453da0
c657020
9738ed3
540056e
 
e8fd75c
5301c9c
 
540056e
00856ac
5301c9c
e8fd75c
540056e
d6695b2
e8fd75c
540056e
5301c9c
 
 
 
a2c09c5
5301c9c
 
a15b3ce
99bd104
 
c657020
 
5e6dba7
 
 
 
 
 
9738ed3
 
 
55e712f
9738ed3
 
 
 
 
 
 
c657020
a15b3ce
bf18300
a15b3ce
 
bf18300
a15b3ce
 
bf18300
a15b3ce
9738ed3
540056e
cf39162
0a67e21
 
 
 
 
 
 
 
 
396214f
 
 
 
 
c93d85e
396214f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80aa4e5
508045d
 
 
9e731de
 
 
 
508045d
455006b
aef38d7
 
 
 
 
 
455006b
 
efc72c4
508045d
fc0768e
508045d
 
 
540056e
00856ac
540056e
e8fd75c
f7b9ef5
e8fd75c
 
20fea69
508045d
e8fd75c
508045d
80aa4e5
 
 
aef38d7
396214f
 
e8fd75c
c93d85e
508045d
aef38d7
 
b9bed89
 
 
 
9375fef
b9bed89
 
 
 
 
 
 
aef38d7
b9bed89
 
776a974
 
 
 
 
 
 
 
 
 
5bddbaf
396214f
5bddbaf
396214f
 
776a974
306eb01
 
 
51691ec
a72a83f
306eb01
e0b6034
a2c4d07
 
 
306eb01
 
 
5bddbaf
6a3a19b
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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
import gradio as gr
import json
import re
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 pattern matching the expected format ("Describe this image in detail:" followed by optional space and then the rest)...
    pattern = r'^Describe this image in detail:\s*(.*)$'
    # Apply the regex pattern to extract the description text.
    match = re.search(pattern, full_sentence)
    if match:
        description = match.group(1)
        print(description)
    else:
        print("Unable to locate valid description.")

    # 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 description

def get_caption_from_MD(image_in):
    client = Client("https://vikhyatk-moondream1.hf.space/")
    result = client.predict(
		image_in,	# filepath  in 'image' Image component
		"Describe precisely the image.",	# str  in 'Question' Textbox component
		api_name="/answer_question"
    )
    print(result)
    return result

def get_magnet(prompt):
    amended_prompt = f"{prompt}"
    print(amended_prompt)
    client = Client("https://fffiloni-magnet.hf.space/")
    result = client.predict(
        "facebook/magnet-medium-10secs",	# Literal['facebook/magnet-small-10secs', 'facebook/magnet-medium-10secs', 'facebook/magnet-small-30secs', 'facebook/magnet-medium-30secs', 'facebook/audio-magnet-small', 'facebook/audio-magnet-medium']  in 'Model' Radio component
        "",	# str  in 'Model Path (custom models)' Textbox component
        amended_prompt,	# str  in 'Input Text' Textbox component
        3,	# float  in 'Temperature' Number component
        0.9,	# float  in 'Top-p' Number component
        10,	# float  in 'Max CFG coefficient' Number component
        1,	# float  in 'Min CFG coefficient' Number component
        20,	# float  in 'Decoding Steps (stage 1)' Number component
        10,	# float  in 'Decoding Steps (stage 2)' Number component
        10,	# float  in 'Decoding Steps (stage 3)' Number component
        10,	# float  in 'Decoding Steps (stage 4)' Number component
        "prod-stride1 (new!)",	# Literal['max-nonoverlap', 'prod-stride1 (new!)']  in 'Span Scoring' Radio component
        api_name="/predict_full"
    )
    print(result)
    return result[1]
    
import re
import torch
from transformers import pipeline

zephyr_model = "HuggingFaceH4/zephyr-7b-beta"
mixtral_model = "mistralai/Mixtral-8x7B-Instruct-v0.1"

pipe = pipeline("text-generation", model=mixtral_model, torch_dtype=torch.bfloat16, device_map="auto")

agent_maker_sys = f"""
You are an AI whose job is to help users create their own music which its genre 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 musical prompt for an music generation model.

For example, if a user says, "a picture of a man in a black suit and tie riding a black dragon", provide immediately a musical prompt corresponding to the image description. 
Immediately STOP after that. It should be EXACTLY in this format:
"A grand orchestral arrangement with thunderous percussion, epic brass fanfares, and soaring strings, creating a cinematic atmosphere fit for a heroic battle"
"""

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 Musical prompt: {cleaned_text}")

    music_o = get_magnet(cleaned_text)
    
    return cleaned_text.lstrip("\n"), music_o

title = "Image to Music V2",
description = "Get music from a picture"

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;">{title}</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 = "Musical prompt"
                )
                result = gr.Audio(
                    label = "Music"
                )
        with gr.Row():
            gr.Examples(
                examples = [
                    ["examples/monalisa.png"],
                    ["examples/santa.png"],
                    ["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 = False
            )

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

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