ML-ImageHash/Models.py at master · apaz-cli/ML-ImageHash · GitHub

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
import torch
from torch import nn
import torchvision.transforms
import torch.nn.functional as F
from collections import OrderedDict
from torch.nn.init import dirac_, xavier_normal_


class ReZero(nn.Module):
    def __init__(self, layer):
        super(ReZero, self).__init__()
        # Suppress pylint bug
        # pylint: disable=E1101
        self.layer = layer
        self.α = nn.Parameter(torch.zeros(1), requires_grad=True)
        # pylint: enable=E1101

    def forward(self, x):
        return x + self.α*self.layer(x)


class OntoF16(nn.Module):
    def __init__(self):
        super(OntoF16, self).__init__()
        self.f16_map = 2**12 - 1

    def forward(self, x):
        return x * self.f16_map


class ResnetHasher(nn.Module):
    def __init__(self):
        super(ResnetHasher, self).__init__()

        self.resnet = torch.hub.load(
            'pytorch/vision:v0.6.0', 'resnet18', pretrained=True)
        self.enc_layers = OrderedDict([
            ('resnet', self.resnet),
            ('relu_r', nn.ReLU(1000)),
            ('norm_r', nn.BatchNorm1d(1000)),

            ('full_1', nn.Linear(1000, 500)),
            ('relu_1', nn.ReLU()),
            ('norm_1', nn.BatchNorm1d(500)),

            ('full_2', nn.Linear(500, 250)),
            ('relu_2', nn.ReLU()),
            ('norm_2', nn.BatchNorm1d(250)),

            ('output', nn.Linear(250, 8)),
            ('tanh', nn.Tanh()),
            #('ontoF16', OntoF16())
        ])
        self.encoder = nn.Sequential(self.enc_layers)

    def forward(self, x):
        return self.encoder(x)

    def visualize(self):
        from torchsummary import summary
        summary(self, (3, 64, 64))


class HasherBlock(nn.Module):
    def __init__(self, block_num, in_channels, out_channels, kernel_size, kernel_stride, conv_padding, pool_num=0):
        super(HasherBlock, self).__init__()

        self.block_num = block_num

        # Set layers
        self.layers = OrderedDict()
        self.layers[f'conv{self.block_num}'] = nn.Conv2d(
            in_channels, out_channels, kernel_size, kernel_stride, padding=conv_padding)
        if pool_num != 0:
            self.layers[f'pool{self.block_num}'] = nn.MaxPool2d(pool_num)
        self.layers[f'relu{self.block_num}'] = nn.ReLU()
        self.layers[f'norm{self.block_num}'] = nn.BatchNorm2d(out_channels)

        self.model = nn.Sequential(self.layers)

    def init(self):
        dirac_(self.layers[f'conv{self.block_num}'].weight.data)

    def forward(self, x):
        return self.model(x)

    def visualize(self):
        from torchsummary import summary
        summary(self, (3, 64, 64))


class CustomHasher(nn.Module):
    def __init__(self):
        super(CustomHasher, self).__init__()

        self.layers = OrderedDict([
            # block_num, in_channels, out_channels, kernel_size, kernel_stride, conv_padding pool_num
            ('block_1', HasherBlock(1,  3, 12, 3, 1, 1, 2)),
            ('block_2', HasherBlock(2, 12, 24, 3, 1, 1, 2)),
            ('block_3', HasherBlock(3, 24, 48, 3, 1, 1, 2)),
            ('block_4', HasherBlock(4, 48, 96, 3, 1, 1, 2)),
            ('adapt', nn.AdaptiveMaxPool2d(1)),
            ('flat ', nn.Flatten()),
            ('full_1', nn.Linear(96, 36)),
            ('full_2', nn.Linear(36, 8)),
            ('tanh', nn.Tanh()),
            #('ontoF16', OntoF16())
        ])
        self.encoder = nn.Sequential(self.layers)

    def forward(self, x):
        return self.encoder(x)

    def visualize(self):
        from torchsummary import summary
        summary(self, (3, 64, 64))


class SmallHasher(nn.Module):
    def __init__(self):
        super(SmallHasher, self).__init__()

        self.layers = OrderedDict([
            # block_num, in_channels, out_channels, kernel_size, kernel_stride, conv_padding pool_num
            ('block_1', HasherBlock(1,  3, 16, 3, 2, 1, 1)),
            ('block_2', HasherBlock(2, 16, 20, 3, 2, 1, 1)),
            ('block_3', HasherBlock(3, 20, 28, 3, 2, 1, 1)),
            ('block_4', HasherBlock(4, 28, 32, 3, 2, 1, 1)),
            ('block_5', HasherBlock(5, 32, 40, 3, 2, 1, 1)),
            ('adapt', nn.AdaptiveMaxPool2d(1)),
            ('flat ', nn.Flatten()),
            ('full_1', nn.Linear(40, 24)),
            ('relu_1', nn.ReLU()),
            ('full_2', nn.Linear(24, 16)),
            ('tanh', nn.Tanh())
        ])

        # Careful initialization
        self.layers['block_1'].init()
        self.layers['block_2'].init()
        self.layers['block_3'].init()
        self.layers['block_4'].init()
        self.layers['block_5'].init()
        xavier_normal_(self.layers['full_1'].weight.data)
        xavier_normal_(self.layers['full_2'].weight.data)

        self.encoder = nn.Sequential(self.layers)

    def forward(self, x):
        return self.encoder(x)

    def visualize(self):
        from torchsummary import summary
        summary(self, (3, 64, 64))