File size: 6,699 Bytes
ccb8f0e
07da240
 
 
4b2b363
07da240
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4b2b363
07da240
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4b2b363
 
7d66afc
4b2b363
 
 
 
07da240
 
 
 
 
 
 
4b2b363
 
 
07da240
 
 
 
 
42713e7
 
07da240
 
 
 
 
4b2b363
07da240
42713e7
07da240
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dc036e7
 
 
07da240
 
 
 
 
dc036e7
07da240
 
 
 
 
 
 
 
42713e7
63a8a75
42713e7
63a8a75
 
42713e7
 
63a8a75
42713e7
63a8a75
 
42713e7
07da240
42713e7
 
 
 
07da240
 
 
42713e7
07da240
8e25602
07da240
8adf5fe
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
import os, re, cv2 
from typing import Mapping, Tuple, Dict
import gradio as gr
import numpy as np
import io
import pandas as pd
from PIL import Image
from huggingface_hub import hf_hub_download
from onnxruntime import InferenceSession

# noinspection PyUnresolvedReferences
def make_square(img, target_size):
    old_size = img.shape[:2]
    desired_size = max(old_size)
    desired_size = max(desired_size, target_size)

    delta_w = desired_size - old_size[1]
    delta_h = desired_size - old_size[0]
    top, bottom = delta_h // 2, delta_h - (delta_h // 2)
    left, right = delta_w // 2, delta_w - (delta_w // 2)

    color = [255, 255, 255]
    return cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)


# noinspection PyUnresolvedReferences
def smart_resize(img, size):
    # Assumes the image has already gone through make_square
    if img.shape[0] > size:
        img = cv2.resize(img, (size, size), interpolation=cv2.INTER_AREA)
    elif img.shape[0] < size:
        img = cv2.resize(img, (size, size), interpolation=cv2.INTER_CUBIC)
    else:  # just do nothing
        pass

    return img


class WaifuDiffusionInterrogator:
    def __init__(
            self,
            repo='SmilingWolf/wd-v1-4-vit-tagger',
            model_path='model.onnx',
            tags_path='selected_tags.csv',
            mode: str = "auto"
    ) -> None:
        self.__repo = repo
        self.__model_path = model_path
        self.__tags_path = tags_path
        self._provider_mode = mode

        self.__initialized = False
        self._model, self._tags = None, None
        self.cache = {}
    def _init(self) -> None:
        if self.__initialized:
            return

        model_path = hf_hub_download(self.__repo, filename=self.__model_path)
        tags_path = hf_hub_download(self.__repo, filename=self.__tags_path)

        self._model = InferenceSession(str(model_path))
        self._tags = pd.read_csv(tags_path)

        self.__initialized = True

    def _calculation(self, image: Image.Image) -> pd.DataFrame:
        self._init()

        # code for converting the image and running the model is taken from the link below
        # thanks, SmilingWolf!
        # https://huggingface.co/spaces/SmilingWolf/wd-v1-4-tags/blob/main/app.py

        # convert an image to fit the model
        _, height, _, _ = self._model.get_inputs()[0].shape

        # alpha to white
        image = image.convert('RGBA')
        new_image = Image.new('RGBA', image.size, 'WHITE')
        new_image.paste(image, mask=image)
        image = new_image.convert('RGB')
        image = np.asarray(image)

        # PIL RGB to OpenCV BGR
        image = image[:, :, ::-1]

        image = make_square(image, height)
        image = smart_resize(image, height)
        image = image.astype(np.float32)
        image = np.expand_dims(image, 0)

        # evaluate model
        input_name = self._model.get_inputs()[0].name
        label_name = self._model.get_outputs()[0].name
        confidence = self._model.run([label_name], {input_name: image})[0]

        full_tags = self._tags[['name', 'category']].copy()
        full_tags['confidence'] = confidence[0]

        return full_tags
    def interrogate(self, image: Image) -> Tuple[Dict[str, float], Dict[str, float]]:
        
        imgByteArr = io.BytesIO()
        image.save(imgByteArr, format="png")
        imgByteArr = imgByteArr.getvalue()
        if imgByteArr in cache:
            return cache[imgByteArr]

        full_tags = self._calculation(image)

        # first 4 items are for rating (general, sensitive, questionable, explicit)
        ratings = dict(full_tags[full_tags['category'] == 9][['name', 'confidence']].values)

        # rest are regular tags
        tags = dict(full_tags[full_tags['category'] != 9][['name', 'confidence']].values)
        self.cache[imgByteArr] = (ratings, tags)
        if len(cache) > 25:
            cache.popitem(last=False)

        return ratings, tags


WAIFU_MODELS: Mapping[str, WaifuDiffusionInterrogator] = {
    'chen-vit': WaifuDiffusionInterrogator(),
    'chen-convnext': WaifuDiffusionInterrogator(
        repo='SmilingWolf/wd-v1-4-convnext-tagger'
    ),
}
RE_SPECIAL = re.compile(r'([\\()])')


def image_to_wd14_tags(image: Image.Image, model_name: str, threshold: float,
                       use_spaces: bool, use_escape: bool, include_ranks=False, score_descend=True) \
        -> Tuple[Mapping[str, float], str, Mapping[str, float]]:
    model = WAIFU_MODELS[model_name]
    ratings, tags = model.interrogate(image)

    filtered_tags = {
        tag: score for tag, score in tags.items()
        if score >= threshold
    }

    text_items = []
    tags_pairs = filtered_tags.items()
    if score_descend:
        tags_pairs = sorted(tags_pairs, key=lambda x: (-x[1], x[0]))
    for tag, score in tags_pairs:
        tag_outformat = tag
        if use_spaces:
            tag_outformat = tag_outformat.replace('_', ' ')
            tag_outformat.replace(' ', '-')
        else:
            tag_outformat.replace('_', '-')
        if use_escape:
            tag_outformat = re.sub(RE_SPECIAL, r'\\\1', tag_outformat)
        if include_ranks:
            tag_outformat = f"({tag_outformat}:{score:.3f})"
        text_items.append(tag_outformat)
    output_text = ' '.join(text_items)

    return ratings, output_text, filtered_tags


if __name__ == '__main__':
    with gr.Blocks() as demo:
        with gr.Row():
            with gr.Column():
                gr_input_image = gr.Image(type='pil', label='Chen Chen')
                with gr.Row():
                    gr_model = gr.Radio(list(WAIFU_MODELS.keys()), value='chen-vit', label='Chen')
                    gr_threshold = gr.Slider(0.0, 1.0, 0.5, label='Chen Chen Chen Chen Chen')
                with gr.Row():
                    gr_space = gr.Checkbox(value=True, label='Chen " " Chen Chen "_"')
                    gr_escape = gr.Checkbox(value=True, label='Chen Text Escape')

                gr_btn_submit = gr.Button(value='橙', variant='primary')

            with gr.Column():
                gr_ratings = gr.Label(label='橙 橙')
                with gr.Tabs():
                    with gr.Tab("Chens"):
                        gr_tags = gr.Label(label='Chens')
                    with gr.Tab("Chen Text"):
                        gr_output_text = gr.TextArea(label='Chen Text')

        gr_btn_submit.click(
            image_to_wd14_tags,
            inputs=[gr_input_image, gr_model, gr_threshold, gr_space, gr_escape],
            outputs=[gr_ratings, gr_output_text, gr_tags],
            api_name="classify"
        )
    demo.queue(os.cpu_count()).launch()