def __init__(self, backbone, num_classes, sync_bn=False):
super(Decoder, self).__init__()
if backbone == 'resnet' or backbone == 'drn':
low_feature_size = 256
elif backbone == 'xception':
low_feature_size = 128
elif backbone == 'mobilenet':
low_feature_size = 24
elif backbone == 'inception':
low_feature_size = 192
else:
raise NotImplementedError
self.conv1 = nn.Conv2d(low_feature_size, 48, 1, bias=False)
self.bn1 = BatchNorm(48, sync_bn)
self.relu = nn.ReLU()
# here 304 = 256 + 48, is the sum size of low level feature and output feature
self.last_conv = nn.Sequential(
nn.Conv2d(304, 256, kernel_size=3, stride=1, padding=1,
bias=False), BatchNorm(256, sync_bn), nn.ReLU(),
nn.Dropout(0.5),
nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1,
bias=False), BatchNorm(256, sync_bn), nn.ReLU(),
nn.Dropout(0.1),
nn.Conv2d(256, num_classes, kernel_size=1, stride=1))
initial_weight(self.modules())
def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None, sync_bn=False):
super(BasicBlock, self).__init__()
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = BatchNorm(planes, sync_bn)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes, dilation=dilation, padding=dilation)
self.bn2 = BatchNorm(planes, sync_bn)
self.downsample = downsample
self.stride = stride
def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None, sync_bn=False):
super(Bottleneck, self).__init__()
self.conv1 = conv1x1(inplanes, planes)
self.bn1 = BatchNorm(planes, sync_bn)
self.conv2 = conv3x3(planes, planes, stride, dilation=dilation, padding=dilation)
self.bn2 = BatchNorm(planes, sync_bn)
self.conv3 = conv1x1(planes, planes * self.expansion)
self.bn3 = BatchNorm(planes * self.expansion)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
def __init__(self, backbone, output_scale, sync_bn=False):
super(ASPP, self).__init__()
if backbone == 'drn':
inplanes = 512
elif backbone == 'mobilenet':
inplanes = 320
else:
inplanes = 2048
if output_scale == 16:
dilations = [1, 6, 12, 18]
elif output_scale == 8:
dilations = [1, 12, 24, 36]
else:
raise NotImplementedError
self.layer1 = ASPPBlock(inplanes,
256,
1,
padding=0,
dilation=dilations[0],
sync_bn=sync_bn)
self.layer2 = ASPPBlock(inplanes,
256,
3,
padding=dilations[1],
dilation=dilations[1],
sync_bn=sync_bn)
self.layer3 = ASPPBlock(inplanes,
256,
3,
padding=dilations[2],
dilation=dilations[2],
sync_bn=sync_bn)
self.layer4 = ASPPBlock(inplanes,
256,
3,
padding=dilations[3],
dilation=dilations[3],
sync_bn=sync_bn)
self.global_avg_pool = nn.Sequential(
nn.AdaptiveAvgPool2d((1, 1)),
nn.Conv2d(inplanes, 256, 1, 1, bias=False),
BatchNorm(256, sync_bn), nn.ReLU())
self.conv1 = nn.Conv2d(256 * 5, 256, 1, bias=False)
self.bn1 = BatchNorm(256, sync_bn)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(0.5)
initial_weight(self.modules())
def __init__(self, block, layers, output_scale=16, sync_bn=False, zero_init_residual=False):
super(ResNet, self).__init__()
self.inplanes = 64
if output_scale == 16:
strides = [1, 2, 2, 1]
dilations = [1, 1, 1, 2]
elif output_scale == 8:
strides = [1, 2, 1, 1]
dilations = [1, 1, 2, 4]
else:
raise NotImplementedError
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1 = BatchNorm(64, sync_bn)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0], stride=strides[0], dilation=dilations[0], sync_bn=sync_bn)
self.layer2 = self._make_layer(block, 128, layers[1], stride=strides[1], dilation=dilations[1], sync_bn=sync_bn)
self.layer3 = self._make_layer(block, 256, layers[2], stride=strides[2], dilation=dilations[2], sync_bn=sync_bn)
self.layer4 = self._make_layer(block, 512, layers[3], stride=strides[3], dilation=dilations[3], sync_bn=sync_bn)
initial_weight(self.modules())
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
if zero_init_residual:
for m in self.modules():
if isinstance(m, Bottleneck):
nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
def _make_layer(self, block, planes, blocks, stride=1, dilation=1, sync_bn=False):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
conv1x1(self.inplanes, planes * block.expansion, stride),
BatchNorm(planes * block.expansion, sync_bn),
)
layers = []
layers.append(block(self.inplanes, planes, stride, dilation, downsample, sync_bn))
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes, dilation=dilation, sync_bn=sync_bn))
return nn.Sequential(*layers)
def __init__(self,
inplanes,
planes,
kernel,
padding,
dilation,
sync_bn=False):
super(ASPPBlock, self).__init__()
self.atrous_conv = nn.Conv2d(inplanes,
planes,
kernel_size=kernel,
stride=1,
padding=padding,
dilation=dilation,
bias=False)
self.bn = BatchNorm(planes, sync_bn)
# what will happen if use LeakyReLU
self.relu = nn.ReLU()
def __init__(self, in_channels, out_channels, sync_bn, **kwargs):
super(BasicConv2d, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs)
self.bn = BatchNorm(out_channels, sync_bn)