File size: 15,572 Bytes
226675b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 |
import torch
import torch.nn as nn
import torch.nn.functional as F
#from torchvision import models
#from base import BaseModel
#from utils.helpers import initialize_weights
from itertools import chain
#from swin_transformer import SwinTransformer
from einops import rearrange
from torch.hub import load_state_dict_from_url
GlobalAvgPool2D = lambda: nn.AdaptiveAvgPool2d(1)
model_urls = {
'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
}
class Cross_transformer_backbone(nn.Module):
def __init__(self, in_channels = 48):
super(Cross_transformer_backbone, self).__init__()
self.to_key = nn.Linear(in_channels * 2, in_channels, bias=False)
self.to_value = nn.Linear(in_channels * 2, in_channels, bias=False)
self.softmax = nn.Softmax(dim=-1)
self.gamma_cam_lay3 = nn.Parameter(torch.zeros(1))
self.cam_layer0 = nn.Sequential(
nn.Conv2d(in_channels, in_channels, kernel_size=3, padding=1),
nn.BatchNorm2d(in_channels),
nn.ReLU()
)
self.cam_layer1 = nn.Sequential(
nn.Conv2d(in_channels, in_channels, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(in_channels),
nn.ReLU(inplace=True)
)
self.cam_layer2 = nn.Sequential(
nn.Conv2d(in_channels, in_channels, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(in_channels),
nn.ReLU(inplace=True)
)
self.mlp = nn.Sequential(
nn.Conv2d(in_channels*2, in_channels, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(in_channels),
nn.ReLU(inplace=True)
)
def forward(self, input_feature, features):
Query_features = input_feature
Query_features = self.cam_layer0(Query_features)
key_features = self.cam_layer1(features)
value_features = self.cam_layer2(features)
QK = torch.einsum("nlhd,nshd->nlsh", Query_features, key_features)
softmax_temp = 1. / Query_features.size(3)**.5
A = torch.softmax(softmax_temp * QK, dim=2)
queried_values = torch.einsum("nlsh,nshd->nlhd", A, value_features).contiguous()
message = self.mlp(torch.cat([input_feature, queried_values], dim=1))
return input_feature + message
class Cross_transformer(nn.Module):
def __init__(self, in_channels = 48):
super(Cross_transformer, self).__init__()
self.fa = nn.Linear(in_channels , in_channels, bias=False)
self.fb = nn.Linear(in_channels, in_channels, bias=False)
self.fc = nn.Linear(in_channels , in_channels, bias=False)
self.fd = nn.Linear(in_channels, in_channels, bias=False)
self.softmax = nn.Softmax(dim=-1)
self.to_out = nn.Sequential(
nn.Conv2d(in_channels, in_channels, kernel_size=3, padding=1),
nn.BatchNorm2d(in_channels),
nn.ReLU(inplace=True)
)
self.gamma_cam_lay3 = nn.Parameter(torch.zeros(1))
self.fuse = nn.Sequential(
nn.Conv2d(in_channels * 4, in_channels, kernel_size=1, padding=0),
nn.BatchNorm2d(in_channels),
nn.ReLU(inplace=True)
)
def attention_layer(self, q, k, v, m_batchsize, C, height, width):
k = k.permute(0, 2, 1)
energy = torch.bmm(q, k)
energy_new = torch.max(energy, -1, keepdim=True)[0].expand_as(energy)-energy
attention = self.softmax(energy_new)
out = torch.bmm(attention, v)
out = out.view(m_batchsize, C, height, width)
return out
def forward(self, input_feature, features):
fa = input_feature
fb = features[0]
fc = features[1]
fd = features[2]
m_batchsize, C, height, width = fa.size()
fa = self.fa(fa.view(m_batchsize, C, -1).permute(0, 2, 1)).permute(0, 2, 1)
fb = self.fb(fb.view(m_batchsize, C, -1).permute(0, 2, 1)).permute(0, 2, 1)
fc = self.fc(fc.view(m_batchsize, C, -1).permute(0, 2, 1)).permute(0, 2, 1)
fd = self.fd(fd.view(m_batchsize, C, -1).permute(0, 2, 1)).permute(0, 2, 1)
qkv_1 = self.attention_layer(fa, fa, fa, m_batchsize, C, height, width)
qkv_2 = self.attention_layer(fa, fb, fb, m_batchsize, C, height, width)
qkv_3 = self.attention_layer(fa, fc, fc, m_batchsize, C, height, width)
qkv_4 = self.attention_layer(fa, fd, fd, m_batchsize, C, height, width)
atten = self.fuse(torch.cat((qkv_1, qkv_2, qkv_3, qkv_4), dim = 1))
out = self.gamma_cam_lay3 * atten + input_feature
out = self.to_out(out)
return out
class SceneRelation(nn.Module):
def __init__(self,
in_channels,
channel_list,
out_channels,
scale_aware_proj=True):
super(SceneRelation, self).__init__()
self.scale_aware_proj = scale_aware_proj
if scale_aware_proj:
self.scene_encoder = nn.ModuleList(
[nn.Sequential(
nn.Conv2d(in_channels, out_channels, 1),
nn.ReLU(True),
nn.Conv2d(out_channels, out_channels, 1),
) for _ in range(len(channel_list))]
)
else:
# 2mlp
self.scene_encoder = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 1),
nn.ReLU(True),
nn.Conv2d(out_channels, out_channels, 1),
)
self.content_encoders = nn.ModuleList()
self.feature_reencoders = nn.ModuleList()
for c in channel_list:
self.content_encoders.append(
nn.Sequential(
nn.Conv2d(c, out_channels, 1),
nn.BatchNorm2d(out_channels),
nn.ReLU(True)
)
)
self.feature_reencoders.append(
nn.Sequential(
nn.Conv2d(c, out_channels, 1),
nn.BatchNorm2d(out_channels),
nn.ReLU(True)
)
)
self.normalizer = nn.Sigmoid()
def forward(self, scene_feature, features: list):
content_feats = [c_en(p_feat) for c_en, p_feat in zip(self.content_encoders, features)]
scene_feats = [op(scene_feature) for op in self.scene_encoder]
relations = [self.normalizer(sf) * cf for sf, cf in
zip(scene_feats, content_feats)]
return relations
class PSPModule(nn.Module):
def __init__(self, in_channels, bin_sizes=[1, 2, 4, 6]):
super(PSPModule, self).__init__()
out_channels = in_channels // len(bin_sizes)
self.stages = nn.ModuleList([self._make_stages(in_channels, out_channels, b_s)
for b_s in bin_sizes])
self.bottleneck = nn.Sequential(
nn.Conv2d(in_channels+(out_channels * len(bin_sizes)), in_channels,
kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(in_channels),
nn.ReLU(inplace=True),
nn.Dropout2d(0.1)
)
def _make_stages(self, in_channels, out_channels, bin_sz):
conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
bn = nn.BatchNorm2d(out_channels)
relu = nn.ReLU(inplace=True)
return nn.Sequential(conv, bn, relu)
def forward(self, features):
h, w = features.size()[2], features.size()[3]
pyramids = [features]
pyramids.extend([F.interpolate(stage(features), size=(h, w), mode='bilinear',
align_corners=True) for stage in self.stages])
output = self.bottleneck(torch.cat(pyramids, dim=1))
return output
class Change_detection(nn.Module):
# Implementing only the object path
def __init__(self, num_classes=2, use_aux=True, fpn_out=48, freeze_bn=False, **_):
super(Change_detection, self).__init__()
f_channels = [64, 128, 256, 512]
# CNN-backbone
self.PPN = PSPModule(f_channels[-1])
# Relation-aware
self.Cross_transformer_backbone_a3 = Cross_transformer_backbone(in_channels = f_channels[3])
self.Cross_transformer_backbone_a2 = Cross_transformer_backbone(in_channels = f_channels[2])
self.Cross_transformer_backbone_a1 = Cross_transformer_backbone(in_channels = f_channels[1])
self.Cross_transformer_backbone_a0 = Cross_transformer_backbone(in_channels = f_channels[0])
self.Cross_transformer_backbone_a33 = Cross_transformer_backbone(in_channels = f_channels[3])
self.Cross_transformer_backbone_a22 = Cross_transformer_backbone(in_channels = f_channels[2])
self.Cross_transformer_backbone_a11 = Cross_transformer_backbone(in_channels = f_channels[1])
self.Cross_transformer_backbone_a00 = Cross_transformer_backbone(in_channels = f_channels[0])
self.Cross_transformer_backbone_b3 = Cross_transformer_backbone(in_channels = f_channels[3])
self.Cross_transformer_backbone_b2 = Cross_transformer_backbone(in_channels = f_channels[2])
self.Cross_transformer_backbone_b1 = Cross_transformer_backbone(in_channels = f_channels[1])
self.Cross_transformer_backbone_b0 = Cross_transformer_backbone(in_channels = f_channels[0])
self.Cross_transformer_backbone_b33 = Cross_transformer_backbone(in_channels = f_channels[3])
self.Cross_transformer_backbone_b22 = Cross_transformer_backbone(in_channels = f_channels[2])
self.Cross_transformer_backbone_b11 = Cross_transformer_backbone(in_channels = f_channels[1])
self.Cross_transformer_backbone_b00 = Cross_transformer_backbone(in_channels = f_channels[0])
# Scale-aware
self.sig = nn.Sigmoid()
self.gap = GlobalAvgPool2D()
self.sr1 = SceneRelation(in_channels = f_channels[3], channel_list = f_channels, out_channels = f_channels[3], scale_aware_proj=True)
self.sr2 = SceneRelation(in_channels = f_channels[2], channel_list = f_channels, out_channels = f_channels[2], scale_aware_proj=True)
self.sr3 = SceneRelation(in_channels = f_channels[1], channel_list = f_channels, out_channels = f_channels[1], scale_aware_proj=True)
self.sr4 = SceneRelation(in_channels = f_channels[0], channel_list =f_channels, out_channels = f_channels[0], scale_aware_proj=True)
# Cross transformer
self.Cross_transformer1 = Cross_transformer(in_channels = f_channels[3])
self.Cross_transformer2 = Cross_transformer(in_channels = f_channels[2])
self.Cross_transformer3 = Cross_transformer(in_channels = f_channels[1])
self.Cross_transformer4 = Cross_transformer(in_channels = f_channels[0])
# Generate change map
self.conv_fusion = nn.Sequential(
nn.Conv2d(960 , fpn_out, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(fpn_out),
nn.ReLU(inplace=True)
)
self.output_fill = nn.Sequential(
nn.ConvTranspose2d(fpn_out , fpn_out, kernel_size=2, stride = 2, bias=False),
nn.BatchNorm2d(fpn_out),
nn.ReLU(inplace=True),
nn.Conv2d(fpn_out, num_classes, kernel_size=3, padding=1)
)
self.active = nn.Sigmoid()
def forward(self, x):
# CNN-backbone
features1, features2 = x
features, features11, features22= [], [],[]
# Relation-aware
for i in range(len(features1)):
if i == 0:
features11.append(self.Cross_transformer_backbone_a00(features1[i] , self.Cross_transformer_backbone_a0(features1[i], features2[i])))
features22.append(self.Cross_transformer_backbone_b00(features2[i], self.Cross_transformer_backbone_b0(features2[i], features1[i])))
elif i == 1:
features11.append(self.Cross_transformer_backbone_a11(features1[i] , self.Cross_transformer_backbone_a1(features1[i], features2[i])))
features22.append(self.Cross_transformer_backbone_b11(features2[i], self.Cross_transformer_backbone_b1(features2[i], features1[i])))
elif i == 2:
features11.append(self.Cross_transformer_backbone_a22(features1[i] , self.Cross_transformer_backbone_a2(features1[i], features2[i])))
features22.append(self.Cross_transformer_backbone_b22(features2[i], self.Cross_transformer_backbone_b2(features2[i], features1[i])))
elif i == 3:
features11.append(self.Cross_transformer_backbone_a33(features1[i] , self.Cross_transformer_backbone_a3(features1[i], features2[i])))
features22.append(self.Cross_transformer_backbone_b33(features2[i], self.Cross_transformer_backbone_b3(features2[i], features1[i])))
# The distance between features from two input images.
for i in range(len(features1)):
features.append(abs(features11[i] - features22[i]))
features[-1] = self.PPN(features[-1])
# Scale-aware and cross transformer
H, W = features[0].size(2), features[0].size(3)
c6 = self.gap(features[-1])
c7 = self.gap(features[-2])
c8 = self.gap(features[-3])
c9 = self.gap(features[-4])
features1, features2, features3, features4 = [], [], [], []
features1[:] = [F.interpolate(feature, size=(64, 64), mode='nearest') for feature in features[:]]
list_3 = self.sr1(c6, features1)
fe3 = self.Cross_transformer1(list_3[-1], [list_3[-2], list_3[-3], list_3[-4]])
features2[:] = [F.interpolate(feature, size=(64, 64), mode='nearest') for feature in features[:]]
list_2 = self.sr2(c7, features2)
fe2 = self.Cross_transformer2(list_2[-2], [list_2[-1], list_2[-3], list_2[-4]])
features3[:] = [F.interpolate(feature, size=(64, 64), mode='nearest') for feature in features[:]]
list_1 = self.sr3(c8, features3)
fe1 = self.Cross_transformer3(list_1[-3], [list_1[-1], list_1[-2], list_1[-4]])
features4[:] = [F.interpolate(feature, size=(128, 128), mode='nearest') for feature in features[:]]
list_0 = self.sr4(c9, features4)
fe0 = self.Cross_transformer4(list_0[-4], [list_0[-1], list_0[-2], list_0[-3]])
refined_fpn_feat_list = [fe3, fe2, fe1, fe0]
# Upsampling
refined_fpn_feat_list[0] = F.interpolate(refined_fpn_feat_list[0], scale_factor=4, mode='nearest')
refined_fpn_feat_list[1] = F.interpolate(refined_fpn_feat_list[1], scale_factor=4, mode='nearest')
refined_fpn_feat_list[2] = F.interpolate(refined_fpn_feat_list[2], scale_factor=4, mode='nearest')
refined_fpn_feat_list[3] = F.interpolate(refined_fpn_feat_list[3], scale_factor=2, mode='nearest')
# Generate change map
x = self.conv_fusion(torch.cat((refined_fpn_feat_list), dim=1))
x = self.output_fill(x)
return x
if __name__ == '__main__':
xa = torch.randn(4, 3, 256, 256)
xb = torch.randn(4, 3, 256, 256)
net = Change_detection()
out = net(xa, xb)
print(out.shape) |