File size: 6,884 Bytes
88a6a7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
from PIL import Image

import torch
import torch.nn as nn
from torchvision.transforms.functional import to_tensor, resize
from typing import List, Tuple, Union
from huggingface_hub import PyTorchModelHubMixin, hf_hub_download

class Bottleneck(nn.Module):

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        stride: int = 1,
        add_downsample: bool = False
    ) -> None:
        super().__init__()
        width = out_channels // 4
        self.conv1 = nn.Conv2d(in_channels, width, kernel_size=1)
        self.conv2 = nn.Conv2d(width, width, kernel_size=3, stride=stride, padding=1)
        self.conv3 = nn.Conv2d(width, out_channels, kernel_size=1)

        self.relu = nn.ReLU(inplace=True)

        self.downsample = None
        if add_downsample:
            self.downsample = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        identity = x

        out = self.conv1(x)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.relu(out)

        out = self.conv3(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

class DeepDanbooruModel(nn.Module, PyTorchModelHubMixin):

    def __init__(
        self,
        block_out_channels: Tuple[int, ...],
        blocks_per_layer: Tuple[int, ...],
        num_classes: int,
        *,
        tag_file: str,
        resolution: int = 512
    ) -> None:
        super().__init__()
        in_channels = 64

        self.block_out_channels = block_out_channels
        self.blocks_per_layer = blocks_per_layer
        self.num_classes = num_classes
        self._tags = []
        with open(tag_file, "r", encoding="utf-8") as f:
            self._tags = [line.strip() for line in f if line.strip()]
        self.resolution = resolution

        self.conv1 = nn.Conv2d(3, in_channels, kernel_size=7, stride=2, padding=3)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=0, ceil_mode=True)

        self.layers = nn.ModuleList([])
        input_channel = in_channels
        for i, (num_block, output_channel) in enumerate(zip(blocks_per_layer, block_out_channels)):
            stride = 1 if i == 0 else 2
            self.layers.append(
                self._make_layer(num_block, input_channel, output_channel, stride=stride)
            )
            input_channel = output_channel

        self.fc = nn.Conv2d(in_channels=block_out_channels[-1], out_channels=num_classes, kernel_size=1, bias=False)
        self.activation = nn.Sigmoid()

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                ks = m.kernel_size[0]
                if ks > 1:
                    m.padding_mode = "constant"
                    if m.stride[0] == 1:
                        m._reversed_padding_repeated_twice = (1, 1, 1, 1)
                    else:
                        m._reversed_padding_repeated_twice = (
                            ks // 2 - 1, ks // 2,
                            ks // 2 - 1, ks // 2
                        )

    def _make_layer(
        self,
        num_block: int,
        in_channels: int,
        out_channels: int,
        stride: int = 1,
    ) -> nn.Sequential:

        layers = nn.Sequential()
        layers.append(
            Bottleneck(in_channels, out_channels, stride=stride, add_downsample=True)
        )

        for _ in range(1, num_block):
            layers.append(
                Bottleneck(out_channels, out_channels, stride=1, add_downsample=False)
            )

        return layers

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.conv1(x)
        x = self.relu(x)
        x = self.maxpool(x)
        
        for layer in self.layers:
            x = layer(x)

        x = self.fc(x)
        x = nn.functional.avg_pool2d(x, kernel_size=x.shape[-2:])
        
        x = torch.flatten(x, 1)
        x = self.activation(x)
        
        return x
    
    @torch.no_grad()
    def tag(
        self,
        image: Union[Image.Image, List[Image.Image], torch.Tensor],
        threshold: float = 0.5
    ) -> List[List[str]]:
        
        # Convert PIL Images to tensors if needed and stack
        if isinstance(image, Image.Image):
            image = [image, ]
        if isinstance(image, List):
            images = torch.stack([
                resize(to_tensor(img), [self.resolution, self.resolution])
            for img in image])
        
        assert images.ndim == 4 and images.shape[-1] == self.resolution and images.shape[-2] == self.resolution, f"Expected 4D tensor (N, C, 512, 512), got shape {images.shape}"
        device = next(self.parameters()).device
        images = images.to(device)

        # Model forward pass
        probs = self(images)

        # Thresholding and tag lookup
        results = []
        for prob_vector in probs:
            selected = (prob_vector > threshold).nonzero(as_tuple=True)[0].cpu().tolist()
            results.append([self._tags[i] for i in selected])

        # Return single result or batch
        return results

    def save_pretrained(
        self,
        save_directory: str,
        config: dict = None,
        **kwargs
    ):
        # 1. Call super to save model and config
        super().save_pretrained(save_directory, config=config, **kwargs)

        readme = os.path.join(save_directory, "README.md")
        if os.path.exists(readme):
            os.remove(readme)

        # 2. Write tags.txt into save_directory
        tags_path = os.path.join(save_directory, "tags.txt")
        with open(tags_path, "w", encoding="utf-8") as f:
            for tag in getattr(self, '_tags', []):
                f.write(f"{tag}\n")

    @classmethod
    def from_pretrained(
        cls,
        pretrained_model_name_or_path: str,
        force_download: bool = False,
        cache_dir: str = None,
        local_files_only: bool = False,
        **model_kwargs
    ):
        model_id = str(pretrained_model_name_or_path)
        if os.path.isdir(model_id):
            tag_file = os.path.join(model_id, "tags.txt")
        else:
            tag_file = hf_hub_download(
                repo_id=model_id,
                filename="tags.txt",
                force_download=force_download,
                cache_dir=cache_dir,
                local_files_only=local_files_only,
            )
        model_kwargs["tag_file"] = tag_file
        
        instance = super().from_pretrained(
            pretrained_model_name_or_path,
            force_download=force_download,
            cache_dir=cache_dir,
            local_files_only=local_files_only,
            **model_kwargs
        )
        return instance