File size: 7,323 Bytes
2252f3d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch.nn as nn
import torch
import math
import torch.nn.functional as F


class single_conv(nn.Module):
    def __init__(self, in_ch, out_ch):
        super(single_conv, self).__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(in_ch, out_ch, 3, stride=1, padding=1),
            nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True),
        )

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


class double_conv(nn.Module):
    def __init__(self, in_ch, out_ch):
        super(double_conv, self).__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(in_ch, out_ch, 3, stride=1, padding=1), nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True), nn.Conv2d(out_ch, out_ch, 3, stride=1, padding=1),
            nn.BatchNorm2d(out_ch), nn.ReLU(inplace=True)
        )

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


class double_conv_down(nn.Module):
    def __init__(self, in_ch, out_ch):
        super(double_conv_down, self).__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(in_ch, out_ch, 3, stride=2, padding=1), nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True), nn.Conv2d(out_ch, out_ch, 3, stride=1, padding=1),
            nn.BatchNorm2d(out_ch), nn.ReLU(inplace=True)
        )

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


class double_conv_up(nn.Module):
    def __init__(self, in_ch, out_ch):
        super(double_conv_up, self).__init__()
        self.conv = nn.Sequential(
            nn.UpsamplingNearest2d(scale_factor=2),
            nn.Conv2d(in_ch, out_ch, 3, stride=1, padding=1), nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True), nn.Conv2d(out_ch, out_ch, 3, stride=1, padding=1),
            nn.BatchNorm2d(out_ch), nn.ReLU(inplace=True)
        )

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


class PosEnSine(nn.Module):
    """
    Code borrowed from DETR: models/positional_encoding.py
    output size: b*(2.num_pos_feats)*h*w
    """
    def __init__(self, num_pos_feats):
        super(PosEnSine, self).__init__()
        self.num_pos_feats = num_pos_feats
        self.normalize = True
        self.scale = 2 * math.pi
        self.temperature = 10000

    def forward(self, x, pt_coord=None):
        b, c, h, w = x.shape
        if pt_coord is not None:
            z_embed = pt_coord[:, :, 2].unsqueeze(-1) + 1.
            y_embed = pt_coord[:, :, 1].unsqueeze(-1) + 1.
            x_embed = pt_coord[:, :, 0].unsqueeze(-1) + 1.
        else:
            not_mask = torch.ones(1, h, w, device=x.device)
            y_embed = not_mask.cumsum(1, dtype=torch.float32)
            x_embed = not_mask.cumsum(2, dtype=torch.float32)
            z_embed = torch.ones_like(x_embed)
        if self.normalize:
            eps = 1e-6
            z_embed = z_embed / (torch.max(z_embed) + eps) * self.scale
            y_embed = y_embed / (torch.max(y_embed) + eps) * self.scale
            x_embed = x_embed / (torch.max(x_embed) + eps) * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
        dim_t = self.temperature**(2 * (dim_t // 2) / self.num_pos_feats)

        pos_x = x_embed[:, :, :, None] / dim_t
        pos_y = y_embed[:, :, :, None] / dim_t
        pos_z = z_embed[:, :, :, None] / dim_t
        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()),
                            dim=4).flatten(3)
        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()),
                            dim=4).flatten(3)
        pos_z = torch.stack((pos_z[:, :, :, 0::2].sin(), pos_z[:, :, :, 1::2].cos()),
                            dim=4).flatten(3)
        pos = torch.cat((pos_x, pos_y, pos_z), dim=3).permute(0, 3, 1, 2)
        # if pt_coord is None:
        pos = pos.repeat(b, 1, 1, 1)
        return pos


def softmax_attention(q, k, v):
    # b x n x d x h x w
    h, w = q.shape[-2], q.shape[-1]

    q = q.flatten(-2).transpose(-2, -1)    # b x n x hw x d
    k = k.flatten(-2)    # b x n x d x hw
    v = v.flatten(-2).transpose(-2, -1)

    print('softmax', q.shape, k.shape, v.shape)

    N = k.shape[-1]    # ?????? maybe change to k.shape[-2]????
    attn = torch.matmul(q / N**0.5, k)
    attn = F.softmax(attn, dim=-1)
    output = torch.matmul(attn, v)

    output = output.transpose(-2, -1)
    output = output.view(*output.shape[:-1], h, w)

    return output, attn


def dotproduct_attention(q, k, v):
    # b x n x d x h x w
    h, w = q.shape[-2], q.shape[-1]

    q = q.flatten(-2).transpose(-2, -1)    # b x n x hw x d
    k = k.flatten(-2)    # b x n x d x hw
    v = v.flatten(-2).transpose(-2, -1)

    N = k.shape[-1]
    attn = None
    tmp = torch.matmul(k, v) / N
    output = torch.matmul(q, tmp)

    output = output.transpose(-2, -1)
    output = output.view(*output.shape[:-1], h, w)

    return output, attn


def long_range_attention(q, k, v, P_h, P_w):    # fixed patch size
    B, N, C, qH, qW = q.size()
    _, _, _, kH, kW = k.size()

    qQ_h, qQ_w = qH // P_h, qW // P_w
    kQ_h, kQ_w = kH // P_h, kW // P_w

    q = q.reshape(B, N, C, qQ_h, P_h, qQ_w, P_w)
    k = k.reshape(B, N, C, kQ_h, P_h, kQ_w, P_w)
    v = v.reshape(B, N, -1, kQ_h, P_h, kQ_w, P_w)

    q = q.permute(0, 1, 4, 6, 2, 3, 5)    # [b, n, Ph, Pw, d, Qh, Qw]
    k = k.permute(0, 1, 4, 6, 2, 3, 5)
    v = v.permute(0, 1, 4, 6, 2, 3, 5)

    output, attn = softmax_attention(q, k, v)    # attn: [b, n, Ph, Pw, qQh*qQw, kQ_h*kQ_w]
    output = output.permute(0, 1, 4, 5, 2, 6, 3)
    output = output.reshape(B, N, -1, qH, qW)
    return output, attn


def short_range_attention(q, k, v, Q_h, Q_w):    # fixed patch number
    B, N, C, qH, qW = q.size()
    _, _, _, kH, kW = k.size()

    qP_h, qP_w = qH // Q_h, qW // Q_w
    kP_h, kP_w = kH // Q_h, kW // Q_w

    q = q.reshape(B, N, C, Q_h, qP_h, Q_w, qP_w)
    k = k.reshape(B, N, C, Q_h, kP_h, Q_w, kP_w)
    v = v.reshape(B, N, -1, Q_h, kP_h, Q_w, kP_w)

    q = q.permute(0, 1, 3, 5, 2, 4, 6)    # [b, n, Qh, Qw, d, Ph, Pw]
    k = k.permute(0, 1, 3, 5, 2, 4, 6)
    v = v.permute(0, 1, 3, 5, 2, 4, 6)

    output, attn = softmax_attention(q, k, v)    # attn: [b, n, Qh, Qw, qPh*qPw, kPh*kPw]
    output = output.permute(0, 1, 4, 2, 5, 3, 6)
    output = output.reshape(B, N, -1, qH, qW)
    return output, attn


def space_to_depth(x, block_size):
    x_shape = x.shape
    c, h, w = x_shape[-3:]
    if len(x.shape) >= 5:
        x = x.view(-1, c, h, w)
    unfolded_x = torch.nn.functional.unfold(x, block_size, stride=block_size)
    return unfolded_x.view(*x_shape[0:-3], c * block_size**2, h // block_size, w // block_size)


def depth_to_space(x, block_size):
    x_shape = x.shape
    c, h, w = x_shape[-3:]
    x = x.view(-1, c, h, w)
    y = torch.nn.functional.pixel_shuffle(x, block_size)
    return y.view(*x_shape[0:-3], -1, h * block_size, w * block_size)


def patch_attention(q, k, v, P):
    # q: [b, nhead, c, h, w]
    q_patch = space_to_depth(q, P)    # [b, nhead, cP^2, h/P, w/P]
    k_patch = space_to_depth(k, P)
    v_patch = space_to_depth(v, P)

    # output: [b, nhead, cP^2, h/P, w/P]
    # attn: [b, nhead, h/P*w/P, h/P*w/P]
    output, attn = softmax_attention(q_patch, k_patch, v_patch)
    output = depth_to_space(output, P)    # output: [b, nhead, c, h, w]
    return output, attn