File size: 5,512 Bytes
9890667
 
 
 
 
 
 
323d67d
9890667
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
323d67d
9890667
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
323d67d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import clip
import copy
from torch.autograd import Function

from collections import OrderedDict
from torchvision import transforms


def convert_state_dict(state_dict):
    new_state_dict = OrderedDict()
    for k, v in state_dict.items():
        if k.startswith("module."):
            k = k.replace("module.", "")
        new_state_dict[k] = v
    return new_state_dict


def convert_weights_float(model: nn.Module):
    """Convert applicable model parameters to fp32"""

    def _convert_weights_to_fp32(l):
        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
            l.weight.data = l.weight.data.float()
            if l.bias is not None:
                l.bias.data = l.bias.data.float()

        if isinstance(l, nn.MultiheadAttention):
            for attr in [
                *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
                "in_proj_bias",
                "bias_k",
                "bias_v",
            ]:
                tensor = getattr(l, attr)
                if tensor is not None:
                    tensor.data = tensor.data.float()

        for name in ["text_projection", "proj"]:
            if hasattr(l, name):
                attr = getattr(l, name)
                if attr is not None:
                    attr.data = attr.data.float()

    model.apply(_convert_weights_to_fp32)


class ReverseLayerF(Function):
    @staticmethod
    def forward(ctx, x, alpha):
        ctx.alpha = alpha

        return x.view_as(x)

    @staticmethod
    def backward(ctx, grad_output):
        output = grad_output.neg() * ctx.alpha

        return output, None


## taken from https://github.com/moein-shariatnia/OpenAI-CLIP/blob/master/modules.py
class ProjectionHead(nn.Module):
    def __init__(self, embedding_dim, projection_dim, dropout=0):
        super().__init__()
        self.projection = nn.Linear(embedding_dim, projection_dim)
        self.gelu = nn.GELU()
        self.fc = nn.Linear(projection_dim, projection_dim)
        self.dropout = nn.Dropout(dropout)
        self.layer_norm = nn.LayerNorm(projection_dim)

    def forward(self, x):
        projected = self.projection(x)
        x = self.gelu(projected)
        x = self.fc(x)
        x = self.dropout(x)
        x = x + projected
        x = self.layer_norm(x)
        return x


def init_weights(m):  # TODO: do we need init for layernorm?
    if isinstance(m, nn.Linear):
        torch.nn.init.xavier_uniform_(m.weight)
        if m.bias is not None:
            nn.init.normal_(m.bias, std=1e-6)


class CSD_CLIP(nn.Module):
    """backbone + projection head"""

    def __init__(self, name="vit_large", content_proj_head="default", model_path=None):
        super(CSD_CLIP, self).__init__()
        self.content_proj_head = content_proj_head
        if name == "vit_large":
            if model_path is None:
                clipmodel, _ = clip.load("ViT-L/14")
            else:
                clipmodel, _ = clip.load(model_path)
            self.backbone = clipmodel.visual
            self.embedding_dim = 1024
        elif name == "vit_base":
            if model_path is None:
                clipmodel, _ = clip.load("ViT-B/16")
            else:
                clipmodel, _ = clip.load(model_path)
            self.backbone = clipmodel.visual
            self.embedding_dim = 768
            self.feat_dim = 512
        else:
            raise Exception("This model is not implemented")

        convert_weights_float(self.backbone)
        self.last_layer_style = copy.deepcopy(self.backbone.proj)
        if content_proj_head == "custom":
            self.last_layer_content = ProjectionHead(self.embedding_dim, self.feat_dim)
            self.last_layer_content.apply(init_weights)

        else:
            self.last_layer_content = copy.deepcopy(self.backbone.proj)

        self.backbone.proj = None

    @property
    def dtype(self):
        return self.backbone.conv1.weight.dtype

    def forward(self, input_data, alpha=None):
        feature = self.backbone(input_data)

        if alpha is not None:
            reverse_feature = ReverseLayerF.apply(feature, alpha)
        else:
            reverse_feature = feature

        style_output = feature @ self.last_layer_style
        style_output = nn.functional.normalize(style_output, dim=1, p=2)

        # if alpha is not None:
        if self.content_proj_head == "custom":
            content_output = self.last_layer_content(reverse_feature)
        else:
            content_output = reverse_feature @ self.last_layer_content
        content_output = nn.functional.normalize(content_output, dim=1, p=2)
        return feature, content_output, style_output


def create_model_and_transforms(model_path="models/csd_clip.pth"):
    # init model
    model = CSD_CLIP("vit_large", "default")

    # load model
    checkpoint = torch.load(model_path, map_location="cpu")
    state_dict = convert_state_dict(checkpoint["model_state_dict"])
    model.load_state_dict(state_dict, strict=False)

    # normalization
    normalize = transforms.Normalize(
        (0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)
    )
    preprocess = transforms.Compose(
        [
            transforms.Resize(
                size=224, interpolation=transforms.functional.InterpolationMode.BICUBIC
            ),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            normalize,
        ]
    )
    return model, preprocess, preprocess