File size: 4,965 Bytes
d4d8571
 
bc904d0
 
 
 
d4d8571
bc904d0
 
 
 
 
 
 
 
 
ee9e2f0
 
bc904d0
 
 
 
 
 
 
 
 
ee9e2f0
bc904d0
 
 
 
 
 
 
 
 
 
 
ee9e2f0
 
 
 
bc904d0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ee9e2f0
bc904d0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d4d8571
bc904d0
 
 
 
ee9e2f0
bc904d0
 
 
 
 
 
 
 
 
 
 
 
 
 
d4d8571
 
bc904d0
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
#!/usr/bin/env python

from __future__ import annotations

import os

import gradio as gr
import PIL.Image
import spaces
import torch
from transformers import InstructBlipForConditionalGeneration, InstructBlipProcessor

DESCRIPTION = "# InstructBLIP"

MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "1024"))

SECRET_TOKEN = os.getenv('SECRET_TOKEN', 'default_secret')

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

model_id = "Salesforce/instructblip-vicuna-7b"
processor = InstructBlipProcessor.from_pretrained(model_id)
model = InstructBlipForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")


@spaces.GPU
def run(
    secret_token: str,
    image: PIL.Image.Image,
    prompt: str,
    text_decoding_method: str = "Nucleus sampling",
    num_beams: int = 5,
    max_length: int = 256,
    min_length: int = 1,
    top_p: float = 0.9,
    repetition_penalty: float = 1.5,
    length_penalty: float = 1.0,
    temperature: float = 1.0,
) -> str:
    if secret_token != SECRET_TOKEN:
        raise gr.Error(
            f'Invalid secret token. Please fork the original space if you want to use it for yourself.')

    h, w = image.size
    scale = MAX_IMAGE_SIZE / max(h, w)
    if scale < 1:
        new_w = int(w * scale)
        new_h = int(h * scale)
        image = image.resize((new_w, new_h), resample=PIL.Image.Resampling.LANCZOS)

    inputs = processor(images=image, text=prompt, return_tensors="pt").to(device, torch.float16)
    generated_ids = model.generate(
        **inputs,
        do_sample=text_decoding_method == "Nucleus sampling",
        num_beams=num_beams,
        max_length=max_length,
        min_length=min_length,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        length_penalty=length_penalty,
        temperature=temperature,
    )
    generated_caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
    return generated_caption


with gr.Blocks(css="style.css") as demo:
    gr.Markdown(DESCRIPTION)

    with gr.Row():
        with gr.Column():
            secret_token = gr.Textbox(label="Secret token")
            input_image = gr.Image(type="pil")
            prompt = gr.Textbox(label="Prompt")
            run_button = gr.Button()
            with gr.Accordion(label="Advanced options", open=False):
                text_decoding_method = gr.Radio(
                    label="Text Decoding Method",
                    choices=["Beam search", "Nucleus sampling"],
                    value="Nucleus sampling",
                )
                num_beams = gr.Slider(
                    label="Number of Beams",
                    minimum=1,
                    maximum=10,
                    step=1,
                    value=5,
                )
                max_length = gr.Slider(
                    label="Max Length",
                    minimum=1,
                    maximum=512,
                    step=1,
                    value=256,
                )
                min_length = gr.Slider(
                    label="Minimum Length",
                    minimum=1,
                    maximum=64,
                    step=1,
                    value=1,
                )
                top_p = gr.Slider(
                    label="Top P",
                    minimum=0.1,
                    maximum=1.0,
                    step=0.1,
                    value=0.9,
                )
                repetition_penalty = gr.Slider(
                    label="Repetition Penalty",
                    info="Larger value prevents repetition.",
                    minimum=1.0,
                    maximum=5.0,
                    step=0.5,
                    value=1.5,
                )
                length_penalty = gr.Slider(
                    label="Length Penalty",
                    info="Set to larger for longer sequence, used with beam search.",
                    minimum=-1.0,
                    maximum=2.0,
                    step=0.2,
                    value=1.0,
                )
                temperature = gr.Slider(
                    label="Temperature",
                    info="Used with nucleus sampling.",
                    minimum=0.5,
                    maximum=1.0,
                    step=0.1,
                    value=1.0,
                )

        with gr.Column():
            output = gr.Textbox(label="Result")

    gr.on(
        triggers=[prompt.submit, run_button.click],
        fn=run,
        inputs=[
            secret_token,
            input_image,
            prompt,
            text_decoding_method,
            num_beams,
            max_length,
            min_length,
            top_p,
            repetition_penalty,
            length_penalty,
            temperature,
        ],
        outputs=output,
        api_name="run",
    )

if __name__ == "__main__":
    demo.queue(max_size=20).launch()