def __init__(self, in_channels, num_class, scale, is_aux=False):
super(BiSeNetHead, self).__init__()
if is_aux:
self.conv_3x3 = ConvBnReLU(in_channels,
256,
kernel_size=3,
stride=1,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False)
else:
self.conv_3x3 = ConvBnReLU(in_channels,
64,
kernel_size=3,
stride=1,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False)
if is_aux:
self.classified_conv = nn.Conv2d(256,
num_class,
kernel_size=1,
stride=1,
padding=0)
else:
self.classified_conv = nn.Conv2d(64,
num_class,
kernel_size=1,
stride=1,
padding=0)
self.scale = scale
def make_decoder(in_channels, out_channels):
layers = nn.Sequential(
ConvBnReLU(in_channels, (in_channels // 4),
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False),
DeconvBnReLU((in_channels // 4), (in_channels // 4),
kernel_size=2,
stride=2,
padding=0,
has_bn=True,
has_relu=True,
inplace=True,
has_bias=False),
ConvBnReLU((in_channels // 4),
out_channels,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False))
return layers
def __init__(self, in_channels, out_channels, reduction=1):
super(FeatureFusion, self).__init__()
self.conv_1x1 = ConvBnReLU(in_channels,
out_channels,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
self.channel_attention = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
ConvBnReLU(out_channels,
out_channels // reduction,
kernel_size=1,
stride=1,
padding=0,
has_bn=False,
has_relu=True,
has_bias=False),
ConvBnReLU(out_channels // reduction,
out_channels,
kernel_size=1,
stride=1,
padding=0,
has_bn=False,
has_relu=False,
has_bias=False), nn.Sigmoid())
def __init__(self, in_channels, out_channels):
super(SpatialPath, self).__init__()
mid_channel = 64
self.conv_7x7 = ConvBnReLU(in_channels,
mid_channel,
kernel_size=7,
stride=2,
padding=3,
has_bn=True,
has_relu=True,
has_bias=False)
self.conv_3x3_1 = ConvBnReLU(mid_channel,
mid_channel,
kernel_size=3,
stride=2,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False)
self.conv_3x3_2 = ConvBnReLU(mid_channel,
mid_channel,
kernel_size=3,
stride=2,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False)
self.conv_1x1 = ConvBnReLU(mid_channel,
out_channels,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
def __init__(self, in_channels, out_channels, num_layers):
super(SegNetDecoder, self).__init__()
layers = list()
layers.append(ConvBnReLU(in_channels, in_channels // 2))
for i in range(num_layers - 1):
layers.append(ConvBnReLU(in_channels//2, in_channels // 2))
layers.append(ConvBnReLU(in_channels // 2, out_channels))
self.decoder = nn.Sequential(*layers)
def __init__(self, in_channels_low, in_channels_high, out_channels, num_class, is_train=True):
super(CascadeFeatureFusion, self).__init__()
self.is_train = is_train
self.conv_low = ConvBnReLU(in_channels_low, out_channels, kernel_size=3, padding=2, dilation=2,
has_bn=True, has_relu=False, has_bias=False)
self.conv_high = ConvBnReLU(in_channels_high, out_channels, kernel_size=3, padding=2, dilation=2,
has_bn=True, has_relu=False, has_bias=False)
self.aux_classify = nn.Conv2d(out_channels, num_class, 1, bias=False)
def __init__(self, num_class, is_train=False):
super(ICNet, self).__init__()
self.conv_resolution1 = nn.Sequential(
ConvBnReLU(3, 32, kernel_size=3, stride=2),
ConvBnReLU(32, 32, kernel_size=3, stride=2),
ConvBnReLU(32, 64, kernel_size=3, stride=2)
)
self.backbone = Backbone(50)
self.ppm = PyramidPoolingModule(512)
self.head = ICNetHead(num_class, is_train=is_train)
self.__setattr__('exclusive', ['conv_sub1', 'head'])
def __init__(self, num_class, is_train=False):
super(BiSeNet, self).__init__()
# self.first_pooling = nn.MaxPool2d(2, 2)
self.context_path = ContextPath()
self.business_layer = []
self.is_train = is_train
self.spatial_path = SpatialPath(3, 128)
conv_channel = 128
self.global_context = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
ConvBnReLU(512,
conv_channel,
kernel_size=1,
stride=1,
padding=0,
has_bn=False,
has_relu=True,
has_bias=False))
# stage = [512, 256, 128, 64]
self.arms = nn.Sequential(AttentionRefinement(512, conv_channel),
AttentionRefinement(256, conv_channel))
self.refines = nn.Sequential(
ConvBnReLU(conv_channel,
conv_channel,
kernel_size=3,
stride=1,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False),
ConvBnReLU(conv_channel,
conv_channel,
kernel_size=3,
stride=1,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False))
self.heads = nn.Sequential(
BiSeNetHead(conv_channel, num_class, 16, True),
BiSeNetHead(conv_channel, num_class, 8, True),
BiSeNetHead(conv_channel * 2, num_class, 8, False))
self.ffm = FeatureFusion(conv_channel * 2, conv_channel * 2, 1)
def __init__(self, in_channels, out_channels):
super(AttentionRefinement, self).__init__()
self.conv_3x3 = ConvBnReLU(in_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False)
self.channel_attention = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
ConvBnReLU(out_channels,
out_channels,
kernel_size=1,
stride=1,
padding=0,
has_bn=False,
has_relu=False,
has_bias=False), nn.Sigmoid())
def __init__(self, in_channels, out_channels, stride=1, downsample=None):
super(ResidualBlock, self).__init__()
self.conv1 = ConvBnReLU(in_channels,
out_channels,
kernel_size=3,
stride=stride,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False)
self.conv2 = ConvBnReLU(out_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1,
has_bn=True,
has_relu=False,
has_bias=False)
self.downsample = downsample
self.stride = stride
self.relu = nn.ReLU(inplace=True)
def __init__(self, num_class):
super(PSPNet, self).__init__()
resnet = resnet_model_load(50)
self.conv1 = resnet.conv1
self.layer1 = resnet.layer1
self.layer2 = resnet.layer2
self.layer3 = resnet.layer3
self.layer4 = resnet.layer4
in_channels = 2048
self.ppm = PyramidPoolingModule(in_channels)
self.classified_conv = nn.Sequential(
ConvBnReLU(in_channels * 2, in_channels // 4),
nn.Conv2d(in_channels // 4, num_class, 1))
def make_encoder(block, in_channels, out_channels):
layers = []
for i in range(0, 2):
if i is 0:
downsample = ConvBnReLU(in_channels,
out_channels,
kernel_size=1,
stride=2,
padding=0,
has_bn=True,
has_relu=False,
has_bias=False)
layers.append(block(in_channels, out_channels, 2, downsample))
else:
layers.append(block(out_channels, out_channels))
return nn.Sequential(*layers)
def __init__(self, in_channels, out_channels, atrous_rates=[6, 12, 18]):
super(AsppPlus, self).__init__()
self.conv0 = ConvBnReLU(in_channels, out_channels, kernel_size=1, stride=1, padding=0,
has_bn=True, has_relu=True, has_bias=False)
self.conv1 = ConvBnReLU(in_channels, out_channels, kernel_size=3, stride=1, padding=atrous_rates[0],
dilation=atrous_rates[0], has_bn=True, has_relu=True, has_bias=False)
self.conv2 = ConvBnReLU(in_channels, out_channels, kernel_size=3, stride=1, padding=atrous_rates[1],
dilation=atrous_rates[1], has_bn=True, has_relu=True, has_bias=False)
self.conv3 = ConvBnReLU(in_channels, out_channels, kernel_size=3, stride=1, padding=atrous_rates[2],
dilation=atrous_rates[2], has_bn=True, has_relu=True, has_bias=False)
self.conv4 = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
ConvBnReLU(in_channels, out_channels, kernel_size=1, stride=1, padding=0,
has_bn=False, has_relu=True, has_bias=False)
)
self.combine_conv = ConvBnReLU(out_channels*5, out_channels, kernel_size=1, stride=1, padding=0,
has_bn=True, has_relu=True, has_bias=False)
def __init__(self, num_class):
super(DFANet, self).__init__()
self.conv1 = ConvBnReLU(3,
8,
kernel_size=3,
stride=2,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False)
self.enc2_1 = DFANetEncoder(in_channels=8, out_channels=48, stage=4)
self.enc3_1 = DFANetEncoder(in_channels=48, out_channels=96, stage=6)
self.enc4_1 = DFANetEncoder(in_channels=96, out_channels=192, stage=4)
self.fca_1 = FcAttention(192, 192)
self.enc2_2 = DFANetEncoder(in_channels=240, out_channels=48, stage=4)
self.enc3_2 = DFANetEncoder(in_channels=144, out_channels=96, stage=6)
self.enc4_2 = DFANetEncoder(in_channels=288, out_channels=192, stage=4)
self.fca_2 = FcAttention(192, 192)
self.enc2_3 = DFANetEncoder(in_channels=240, out_channels=48, stage=4)
self.enc3_3 = DFANetEncoder(in_channels=144, out_channels=96, stage=6)
self.enc4_3 = DFANetEncoder(in_channels=288, out_channels=192, stage=4)
self.fca_3 = FcAttention(192, 192)
# fuse to decoder
self.enc2_1_to_decoder = ConvBnReLU(48,
32,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
self.enc2_2_to_decoder = ConvBnReLU(48,
32,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
self.enc2_3_to_decoder = ConvBnReLU(48,
32,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
self.fca_1_to_decoder = ConvBnReLU(192,
32,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
self.fca_2_to_decoder = ConvBnReLU(192,
32,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
self.fca_3_to_decoder = ConvBnReLU(192,
32,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
self.merge_conv = ConvBnReLU(32,
32,
kernel_size=1,
stride=1,
padding=0,
has_bn=True,
has_relu=True,
has_bias=False)
self.classified_conv = nn.Conv2d(32, num_class, 1, 1, bias=False)
def __init__(self, in_channels, num_class):
super(DeepLabHead, self).__init__()
self.conv1x1 = ConvBnReLU(in_channels, in_channels, kernel_size=1, stride=1, padding=0,
has_bn=True, has_relu=True, has_bias=False)
self.conv3x3 = ConvBnReLU(in_channels*2, num_class, kernel_size=3, stride=1, padding=1,
has_bn=True, has_relu=True, has_bias=False)
def __init__(self, num_class):
super(LinkNet, self).__init__()
filters = [64, 128, 256, 512]
self.feature_scale = 1
filters = [x // self.feature_scale for x in filters]
self.inchannels = filters[0]
# # Encoder
self.encoder_before = nn.Sequential(
ConvBnReLU(3,
filters[0],
kernel_size=7,
stride=2,
padding=3,
has_bn=True,
has_relu=True,
has_bias=False),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),
)
block = ResidualBlock
self.encoder1 = make_encoder(block=block,
in_channels=filters[0],
out_channels=filters[0])
self.encoder2 = make_encoder(block=block,
in_channels=filters[0],
out_channels=filters[1])
self.encoder3 = make_encoder(block=block,
in_channels=filters[1],
out_channels=filters[2])
self.encoder4 = make_encoder(block=block,
in_channels=filters[2],
out_channels=filters[3])
self.avgpool = nn.AvgPool2d(7)
# # Decoder
self.decoder4 = make_decoder(in_channels=filters[3],
out_channels=filters[2])
self.decoder3 = make_decoder(in_channels=filters[2],
out_channels=filters[1])
self.decoder2 = make_decoder(in_channels=filters[1],
out_channels=filters[0])
self.decoder1 = make_decoder(in_channels=filters[0],
out_channels=filters[0])
# # Final classify
self.final_deconv1 = DeconvBnReLU(filters[0],
32 // self.feature_scale,
kernel_size=2,
stride=2,
padding=0,
has_bn=True,
has_relu=True,
inplace=True,
has_bias=False)
self.final_conv2 = ConvBnReLU(32 // self.feature_scale,
32 // self.feature_scale,
kernel_size=3,
stride=1,
padding=1,
has_bn=True,
has_relu=True,
has_bias=False)
self.classified_conv = nn.ConvTranspose2d(32 // self.feature_scale,
num_class, 2, 2)