File size: 4,122 Bytes
319d3b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Code based on timm https://github.com/huggingface/pytorch-image-models

Modifications and additions for mivolo by / Copyright 2023, Irina Tolstykh, Maxim Kuprashevich
"""

import torch
import torch.nn as nn
from timm.layers.bottleneck_attn import PosEmbedRel
from timm.layers.helpers import make_divisible
from timm.layers.mlp import Mlp
from timm.layers.trace_utils import _assert
from timm.layers.weight_init import trunc_normal_


class CrossBottleneckAttn(nn.Module):
    def __init__(
        self,
        dim,
        dim_out=None,
        feat_size=None,
        stride=1,
        num_heads=4,
        dim_head=None,
        qk_ratio=1.0,
        qkv_bias=False,
        scale_pos_embed=False,
    ):
        super().__init__()
        assert feat_size is not None, "A concrete feature size matching expected input (H, W) is required"
        dim_out = dim_out or dim
        assert dim_out % num_heads == 0

        self.num_heads = num_heads
        self.dim_head_qk = dim_head or make_divisible(dim_out * qk_ratio, divisor=8) // num_heads
        self.dim_head_v = dim_out // self.num_heads
        self.dim_out_qk = num_heads * self.dim_head_qk
        self.dim_out_v = num_heads * self.dim_head_v
        self.scale = self.dim_head_qk**-0.5
        self.scale_pos_embed = scale_pos_embed

        self.qkv_f = nn.Conv2d(dim, self.dim_out_qk * 2 + self.dim_out_v, 1, bias=qkv_bias)
        self.qkv_p = nn.Conv2d(dim, self.dim_out_qk * 2 + self.dim_out_v, 1, bias=qkv_bias)

        # NOTE I'm only supporting relative pos embedding for now
        self.pos_embed = PosEmbedRel(feat_size, dim_head=self.dim_head_qk, scale=self.scale)

        self.norm = nn.LayerNorm([self.dim_out_v * 2, *feat_size])
        mlp_ratio = 4
        self.mlp = Mlp(
            in_features=self.dim_out_v * 2,
            hidden_features=int(dim * mlp_ratio),
            act_layer=nn.GELU,
            out_features=dim_out,
            drop=0,
            use_conv=True,
        )

        self.pool = nn.AvgPool2d(2, 2) if stride == 2 else nn.Identity()
        self.reset_parameters()

    def reset_parameters(self):
        trunc_normal_(self.qkv_f.weight, std=self.qkv_f.weight.shape[1] ** -0.5)  # fan-in
        trunc_normal_(self.qkv_p.weight, std=self.qkv_p.weight.shape[1] ** -0.5)  # fan-in
        trunc_normal_(self.pos_embed.height_rel, std=self.scale)
        trunc_normal_(self.pos_embed.width_rel, std=self.scale)

    def get_qkv(self, x, qvk_conv):
        B, C, H, W = x.shape

        x = qvk_conv(x)  # B, (2 * dim_head_qk + dim_head_v) * num_heads, H, W

        q, k, v = torch.split(x, [self.dim_out_qk, self.dim_out_qk, self.dim_out_v], dim=1)

        q = q.reshape(B * self.num_heads, self.dim_head_qk, -1).transpose(-1, -2)
        k = k.reshape(B * self.num_heads, self.dim_head_qk, -1)  # no transpose, for q @ k
        v = v.reshape(B * self.num_heads, self.dim_head_v, -1).transpose(-1, -2)

        return q, k, v

    def apply_attn(self, q, k, v, B, H, W, dropout=None):
        if self.scale_pos_embed:
            attn = (q @ k + self.pos_embed(q)) * self.scale  # B * num_heads, H * W, H * W
        else:
            attn = (q @ k) * self.scale + self.pos_embed(q)
        attn = attn.softmax(dim=-1)
        if dropout:
            attn = dropout(attn)

        out = (attn @ v).transpose(-1, -2).reshape(B, self.dim_out_v, H, W)  # B, dim_out, H, W
        return out

    def forward(self, x):
        B, C, H, W = x.shape

        dim = int(C / 2)
        x1 = x[:, :dim, :, :]
        x2 = x[:, dim:, :, :]

        _assert(H == self.pos_embed.height, "")
        _assert(W == self.pos_embed.width, "")

        q_f, k_f, v_f = self.get_qkv(x1, self.qkv_f)
        q_p, k_p, v_p = self.get_qkv(x2, self.qkv_p)

        # person to face
        out_f = self.apply_attn(q_f, k_p, v_p, B, H, W)
        # face to person
        out_p = self.apply_attn(q_p, k_f, v_f, B, H, W)

        x_pf = torch.cat((out_f, out_p), dim=1)  # B, dim_out * 2, H, W
        x_pf = self.norm(x_pf)
        x_pf = self.mlp(x_pf)  # B, dim_out, H, W

        out = self.pool(x_pf)
        return out