File size: 6,097 Bytes
332190f |
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 |
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
"""
@Author : Qingping Zheng
@Contact : [email protected]
@File : ddgcn.py
@Time : 10/01/21 00:00 PM
@Desc :
@License : Licensed under the Apache License, Version 2.0 (the "License");
@Copyright : Copyright 2022 The Authors. All Rights Reserved.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn.functional as F
import torch.nn as nn
from inplace_abn import InPlaceABNSync
class SpatialGCN(nn.Module):
def __init__(self, plane, abn=InPlaceABNSync):
super(SpatialGCN, self).__init__()
inter_plane = plane // 2
self.node_k = nn.Conv2d(plane, inter_plane, kernel_size=1)
self.node_v = nn.Conv2d(plane, inter_plane, kernel_size=1)
self.node_q = nn.Conv2d(plane, inter_plane, kernel_size=1)
self.conv_wg = nn.Conv1d(inter_plane, inter_plane, kernel_size=1, bias=False)
self.bn_wg = nn.BatchNorm1d(inter_plane)
self.softmax = nn.Softmax(dim=2)
self.out = nn.Sequential(nn.Conv2d(inter_plane, plane, kernel_size=1),
abn(plane))
self.gamma = nn.Parameter(torch.zeros(1))
def forward(self, x):
# b, c, h, w = x.size()
node_k = self.node_k(x)
node_v = self.node_v(x)
node_q = self.node_q(x)
b,c,h,w = node_k.size()
node_k = node_k.view(b, c, -1).permute(0, 2, 1)
node_q = node_q.view(b, c, -1)
node_v = node_v.view(b, c, -1).permute(0, 2, 1)
# A = k * q
# AV = k * q * v
# AVW = k *(q *v) * w
AV = torch.bmm(node_q,node_v)
AV = self.softmax(AV)
AV = torch.bmm(node_k, AV)
AV = AV.transpose(1, 2).contiguous()
AVW = self.conv_wg(AV)
AVW = self.bn_wg(AVW)
AVW = AVW.view(b, c, h, -1)
# out = F.relu_(self.out(AVW) + x)
out = self.gamma * self.out(AVW) + x
return out
class DDualGCN(nn.Module):
"""
Feature GCN with coordinate GCN
"""
def __init__(self, planes, abn=InPlaceABNSync, ratio=4):
super(DDualGCN, self).__init__()
self.phi = nn.Conv2d(planes, planes // ratio * 2, kernel_size=1, bias=False)
self.bn_phi = abn(planes // ratio * 2)
self.theta = nn.Conv2d(planes, planes // ratio, kernel_size=1, bias=False)
self.bn_theta = abn(planes // ratio)
# Interaction Space
# Adjacency Matrix: (-)A_g
self.conv_adj = nn.Conv1d(planes // ratio, planes // ratio, kernel_size=1, bias=False)
self.bn_adj = nn.BatchNorm1d(planes // ratio)
# State Update Function: W_g
self.conv_wg = nn.Conv1d(planes // ratio * 2, planes // ratio * 2, kernel_size=1, bias=False)
self.bn_wg = nn.BatchNorm1d(planes // ratio * 2)
# last fc
self.conv3 = nn.Conv2d(planes // ratio * 2, planes, kernel_size=1, bias=False)
self.bn3 = abn(planes)
self.local = nn.Sequential(
nn.Conv2d(planes, planes, 3, groups=planes, stride=2, padding=1, bias=False),
abn(planes),
nn.Conv2d(planes, planes, 3, groups=planes, stride=2, padding=1, bias=False),
abn(planes),
nn.Conv2d(planes, planes, 3, groups=planes, stride=2, padding=1, bias=False),
abn(planes))
self.gcn_local_attention = SpatialGCN(planes, abn)
self.final = nn.Sequential(nn.Conv2d(planes * 2, planes, kernel_size=1, bias=False),
abn(planes))
self.gamma1 = nn.Parameter(torch.zeros(1))
def to_matrix(self, x):
n, c, h, w = x.size()
x = x.view(n, c, -1)
return x
def forward(self, feat):
# # # # Local # # # #
x = feat
local = self.local(feat)
local = self.gcn_local_attention(local)
local = F.interpolate(local, size=x.size()[2:], mode='bilinear', align_corners=True)
spatial_local_feat = x * local + x
# # # # Projection Space # # # #
x_sqz, b = x, x
x_sqz = self.phi(x_sqz)
x_sqz = self.bn_phi(x_sqz)
x_sqz = self.to_matrix(x_sqz)
b = self.theta(b)
b = self.bn_theta(b)
b = self.to_matrix(b)
# Project
z_idt = torch.matmul(x_sqz, b.transpose(1, 2)) # channel
# # # # Interaction Space # # # #
z = z_idt.transpose(1, 2).contiguous()
z = self.conv_adj(z)
z = self.bn_adj(z)
z = z.transpose(1, 2).contiguous()
# Laplacian smoothing: (I - A_g)Z => Z - A_gZ
z += z_idt
z = self.conv_wg(z)
z = self.bn_wg(z)
# # # # Re-projection Space # # # #
# Re-project
y = torch.matmul(z, b)
n, _, h, w = x.size()
y = y.view(n, -1, h, w)
y = self.conv3(y)
y = self.bn3(y)
# g_out = x + y
# g_out = F.relu_(x+y)
g_out = self.gamma1*y + x
# cat or sum, nearly the same results
out = self.final(torch.cat((spatial_local_feat, g_out), 1))
return out
class DDualGCNHead(nn.Module):
def __init__(self, inplanes, interplanes, abn=InPlaceABNSync):
super(DDualGCNHead, self).__init__()
self.conva = nn.Sequential(nn.Conv2d(inplanes, interplanes, 3, padding=1, bias=False),
abn(interplanes))
self.dualgcn = DDualGCN(interplanes, abn)
self.convb = nn.Sequential(nn.Conv2d(interplanes, interplanes, 3, padding=1, bias=False),
abn(interplanes))
self.bottleneck = nn.Sequential(
nn.Conv2d(inplanes + interplanes, interplanes, kernel_size=3, padding=1, dilation=1, bias=False),
abn(interplanes)
)
def forward(self, x):
output = self.conva(x)
output = self.dualgcn(output)
output = self.convb(output)
output = self.bottleneck(torch.cat([x, output], 1))
return output
|