File size: 3,210 Bytes
9651aac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (C) 2021-2022 Naver Corporation. All rights reserved.
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).

import torch
from torch import nn

class LocalfeatureIntegrationTransformer(nn.Module):
    """Map a set of local features to a fixed number of SuperFeatures """

    def __init__(self, T, N, input_dim, dim):
        """

        T: number of iterations

        N: number of SuperFeatures

        input_dim: dimension of input local features

        dim: dimension of SuperFeatures

        """
        super().__init__()
        self.T = T
        self.N = N
        self.input_dim = input_dim
        self.dim = dim
        # learnable initialization
        self.templates_init = nn.Parameter(torch.randn(1,self.N,dim))
        # qkv
        self.project_q = nn.Linear(dim, dim, bias=False)
        self.project_k = nn.Linear(input_dim, dim, bias=False)
        self.project_v = nn.Linear(input_dim, dim, bias=False)
        # layer norms
        self.norm_inputs = nn.LayerNorm(input_dim)
        self.norm_templates = nn.LayerNorm(dim)
        # for the normalization
        self.softmax = nn.Softmax(dim=-1)
        self.scale = dim ** -0.5
        # mlp
        self.norm_mlp = nn.LayerNorm(dim)
        mlp_dim = dim//2
        self.mlp = nn.Sequential(nn.Linear(dim, mlp_dim), nn.ReLU(), nn.Linear(mlp_dim, dim) )


    def forward(self, x):
        """

        input:

            x has shape BxCxHxW

        output:

            template (output SuperFeatures): tensor of shape BxCxNx1

            attn (attention over local features at the last iteration): tensor of shape BxNxHxW

        """
        # reshape inputs from BxCxHxW to Bx(H*W)xC
        B,C,H,W = x.size()
        x = x.reshape(B,C,H*W).permute(0,2,1)

        # k and v projection
        x = self.norm_inputs(x)
        k = self.project_k(x)
        v = self.project_v(x)

        # template initialization
        templates = torch.repeat_interleave(self.templates_init, B, dim=0)
        attn = None

        # main iteration loop
        for _ in range(self.T):
            templates_prev = templates

            # q projection
            templates = self.norm_templates(templates)
            q = self.project_q(templates)

            # attention
            q = q * self.scale  # Normalization.
            attn_logits =  torch.einsum('bnd,bld->bln', q, k)
            attn = self.softmax(attn_logits)
            attn = attn + 1e-8 # to avoid zero when with the L1 norm below
            attn = attn / attn.sum(dim=-2, keepdim=True)

            # update template
            templates = templates_prev + torch.einsum('bld,bln->bnd', v, attn)

            # mlp
            templates = templates + self.mlp(self.norm_mlp(templates))

        # reshape templates to BxDxNx1
        templates = templates.permute(0,2,1)[:,:,:,None]
        attn = attn.permute(0,2,1).view(B,self.N,H,W)

        return templates, attn

    def __repr__(self):
        s = str(self.__class__.__name__)
        for k in ["T","N","input_dim","dim"]:
            s += "\n  {:s}: {:d}".format(k, getattr(self,k))
        return s