File size: 8,348 Bytes
f75b087
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
be108c4
f75b087
 
 
 
 
 
 
 
7f24008
 
 
 
 
 
f75b087
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63024cf
f75b087
 
 
 
 
 
63024cf
f75b087
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63024cf
f75b087
 
 
 
 
 
f6b05e3
 
f75b087
 
 
63024cf
 
f75b087
9755f16
f75b087
 
 
 
 
 
 
f6b05e3
 
 
 
 
 
 
f75b087
 
 
1f44214
f75b087
9755f16
f75b087
 
 
 
 
 
 
 
 
 
 
 
 
e365d78
f75b087
 
e365d78
 
 
 
 
 
 
 
 
 
f75b087
7f24008
f75b087
e365d78
f75b087
 
e365d78
 
f75b087
 
7f24008
e365d78
f75b087
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1b481c7
 
 
 
 
 
 
 
 
 
5d752fd
1b481c7
 
 
 
f75b087
f6b05e3
f75b087
 
 
 
 
 
 
1f44214
f75b087
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
# after SmilingWolf/wd-tagger

import gradio as gr
import torch
import torch.nn as nn
import timm
import timm.layers.ml_decoder
from transformers import AutoModel, AutoTokenizer

import torchvision
from torchvision import transforms
import PIL
from PIL import Image
import requests
from io import BytesIO
import json
import pickle

headers = {
    "User-Agent": "Gradio 0-shot classification demo",
}

TITLE = "Danbooru 0-shot classifiction demo"
DESCRIPTION = """
Demo for 0-shot classification on Danbooru images.

Davit-tiny backbone, ML-Decoder classification head, Alibaba-NLP/gte-large-en-v1.5 text embedding model.
Training set includes IDs with <= 5,400,000 and last 3 digits in range [0, 899], inclusive.

Get image by uploading or fetching by post ID.
Get tag description by input box or fetching by tag name.
"""

def scrape_img(postID):
    postURL = "https://danbooru.donmai.us/posts/" + str(postID) + ".json"
    postData = json.loads(requests.get(postURL, headers=headers).content)
    imageURL = postData['file_url']

    print("Getting image from " + imageURL)
    response = requests.get(imageURL, headers=headers)
    image = Image.open(BytesIO(response.content))
    image.load()
    return image

def scrape_wiki(tagName):
    wikiHistoryURL = f"https://danbooru.donmai.us/wiki_page_versions.json?search[title]={tagName}"
    wikiHistory = json.loads(requests.get(wikiHistoryURL, headers=headers).content)
    wikiBody = (": " + wikiHistory[0]['body'] if len(wikiHistory) > 0 else "")
    return tagName + wikiBody

class Predictor:
    def __init__(self):
        self.img_size = (288, 288)
        self.cls_model = None
        self.tokenizer = None
        self.text_emb_model = None
        self.class_embed = None
        self.tag_names = None

        self.load_model()

    def load_model(self):
        with open('tags1588.pkl', 'rb') as f:
            classes = pickle.load(f)
        tagNames = classes[0].to_list()
        self.tag_names = tagNames

        pretrained_weights = torch.load('model.pth', map_location=torch.device('cpu'))
        self.class_embed = pretrained_weights['0.head.head.class_embed.weight']
        cls_model = timm.create_model('davit_tiny', num_classes=len(classes))

        cls_model = timm.layers.ml_decoder.add_ml_decoder_head(
            cls_model, 
            num_groups=len(classes), 
            class_embed=self.class_embed, 
            class_embed_merge='', 
            shared_fc=True)
        cls_model = nn.Sequential(cls_model)
        cls_model.load_state_dict(pretrained_weights, strict=True)
        cls_model = cls_model.eval()

        self.cls_model = cls_model

        model_path = 'Alibaba-NLP/gte-large-en-v1.5'
        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
        self.text_emb_model = AutoModel.from_pretrained(model_path, trust_remote_code=True)
        self.text_emb_model = self.text_emb_model.eval()

    def embed_text(self, input_strings):
        with torch.no_grad():
            # Tokenize the input texts
            embeddingList = []
            for text in input_strings:
                batch_dict = self.tokenizer(text, padding=True, truncation=False, return_tensors='pt')
                outputs = self.text_emb_model(**batch_dict.to(self.text_emb_model.device))
                embeddings = outputs.last_hidden_state[:, 0]
                embeddingList.append(embeddings.cpu())
            embeddings = torch.cat(embeddingList)
            return embeddings
    
    def prepare_image(self, image):
        image.load()    # check if file valid
        image = image.convert("RGBA")
    
        color = (255,255,255)
    
        background = Image.new('RGB', image.size, color)
        background.paste(image, mask=image.split()[3])
        image = background
        image = transforms.Resize(self.img_size, interpolation = torchvision.transforms.InterpolationMode.BICUBIC)(image)
        image = transforms.ToTensor()(image)
        return image

    def predict(
        self,
        image,
        query,
        tag_names,
    ):
        image = self.prepare_image(image)

        image_features = self.cls_model[0].forward_features(image.unsqueeze(0))
        outputs = self.cls_model[0].head(image_features, q = query).sigmoid().float()

        general_tag_list = list(zip(tag_names, outputs[0].tolist()))
        general_tag_list.sort(key=lambda y: y[1], reverse=True)
        general_tag_preds_dict = {}
        for tag, prob in general_tag_list[:50]:
            general_tag_preds_dict[tag] = prob

        return general_tag_preds_dict

    def predict_seen_tags(
        self,
        image,
    ):
        return self.predict(image, self.class_embed, self.tag_names)
        
    
    def predict_new_tag(
        self,
        image,
        description,
    ):
        return self.predict(image, self.embed_text([description]), ["embedding"])["embedding"]


def main():
    predictor = Predictor()

    with gr.Blocks(title=TITLE) as demo:
        with gr.Column():
            gr.Markdown(
                value=f"<h1 style='text-align: center; margin-bottom: 1rem'>{TITLE}</h1>"
            )
            gr.Markdown(value=DESCRIPTION)
            with gr.Row():
                with gr.Column(variant="panel"):
                    image = gr.Image(type="pil", image_mode="RGBA", label="Input", height=600)
                    with gr.Row():
                        post_id = gr.Textbox(label="Post ID")
                        with gr.Column():
                            clear = gr.ClearButton(
                                value="Clear image",
                                components=[
                                    image,
                                ],
                                variant="secondary",
                                size="lg",
                            )
                            get_post = gr.Button(value="Get Post", variant="primary", size="lg")
                    with gr.Row():
                        submit = gr.Button(value="Predict known tags", variant="primary", size="lg")
                with gr.Column(variant="panel"):
                    tag_description = gr.Textbox(label="Tag description")
                    with gr.Row():
                        tag_name = gr.Textbox(label="Tag Name")
                        description_prediction = gr.Textbox(label="Probability")   
                    with gr.Row():
                        clear_tag_data = gr.ClearButton(value="Clear tag", variant="secondary", size="lg")
                        get_tag_description = gr.Button(value="Get tag description", variant="primary", size="lg")
                        predict_on_description = gr.Button(value="Predict described tag")
                        
                    general_bars = gr.Label(label="Known tags")
                    clear.add(
                        [
                            general_bars,
                            description_prediction,
                            post_id,
                        ]
                    )
                    clear_tag_data.add(
                        [
                            tag_description,
                            tag_name,
                            description_prediction,
                        ]
                    )


            examples = gr.Examples(
            [
                [
                    "8801249",
                    "short_over_long_sleeves"
                ],
            ],
            inputs=[
                post_id,
                tag_name,
            ],
            run_on_click=False,
            cache_examples=False,
        )
        submit.click(
            predictor.predict_seen_tags,
            inputs=[
                image,
            ],
            outputs=[general_bars],
        )
        predict_on_description.click(
            predictor.predict_new_tag,
            inputs=[image, tag_description],
            outputs=[description_prediction]
        )
        get_post.click(
            scrape_img,
            inputs=[post_id],
            outputs=[image]
        )
        get_tag_description.click(
            scrape_wiki,
            inputs=[tag_name],
            outputs=[tag_description]
        )
        

    demo.queue(max_size=10)
    demo.launch()


if __name__ == "__main__":
    main()