def __init__(self,
num_input_features,
growth_rate,
bn_size,
drop_rate,
norm_type='Unknown'):
super(_DenseLayer, self).__init__()
self.add_module('norm1', get_norm(norm_type, num_input_features)),
self.add_module('relu1', nn.ReLU(inplace=True)),
self.add_module(
'conv1',
nn.Conv2d(num_input_features,
bn_size * growth_rate,
kernel_size=1,
stride=1,
bias=False)),
self.add_module('norm2', get_norm(norm_type, bn_size * growth_rate)),
self.add_module('relu2', nn.ReLU(inplace=True)),
self.add_module(
'conv2',
nn.Conv2d(bn_size * growth_rate,
growth_rate,
kernel_size=3,
stride=1,
padding=1,
bias=False)),
self.drop_rate = drop_rate
def __init__(self, growth_rate=32, block_config=(6, 12, 24, 16), num_init_features=64,
norm_type='Unknown', bn_size=4, drop_rate=0, num_classes=1000, use_se=True):
super(DenseNet, self).__init__()
# First convolution
self.features = nn.Sequential(OrderedDict([('conv0',
nn.Conv2d(3, num_init_features, kernel_size=7,
stride=2, padding=3, bias=False)), (
'norm0', get_norm(norm_type, num_init_features)),
('relu0', nn.ReLU(inplace=True)), ('pool0',
nn.MaxPool2d(
kernel_size=3,
stride=2,
padding=1)), ]))
if use_se:
# Add SELayer at first convolution
self.features.add_module("SELayer_0a", SEModule(channels=num_init_features))
# Each denseblock
num_features = num_init_features
for i, num_layers in enumerate(block_config):
if use_se:
# Add a SELayer
self.features.add_module("SELayer_%da" % (i + 1), SEModule(channels=num_features))
block = _DenseBlock(num_layers=num_layers, num_input_features=num_features,
norm_type=norm_type, bn_size=bn_size, growth_rate=growth_rate,
drop_rate=drop_rate)
self.features.add_module('denseblock%d' % (i + 1), block)
num_features = num_features + num_layers * growth_rate
if i != len(block_config) - 1:
if use_se:
# Add a SELayer behind each transition block
self.features.add_module("SELayer_%db" % (i + 1),
SEModule(channels=num_features))
trans = _Transition(num_input_features=num_features,
num_output_features=num_features // 2, norm_type=norm_type)
self.features.add_module('transition%d' % (i + 1), trans)
num_features = num_features // 2
# Final batch norm
self.features.add_module('norm5', get_norm(norm_type, num_features))
# Linear layer
self.classifier = nn.Linear(num_features, num_classes)
self.num_features = num_features
# Official init from torch repo.
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.constant_(m.bias, 0)
def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
base_width=64, dilation=1, norm_type='Unknown'):
super(Bottleneck, self).__init__()
width = int(planes * (base_width / 64.)) * groups
# Both self.conv2 and self.downsample layers downsample the input when stride != 1
self.conv1 = conv1x1(inplanes, width)
self.bn1 = get_norm(norm_type, width)
self.conv2 = conv3x3(width, width, stride, groups, dilation)
self.bn2 = get_norm(norm_type, width)
self.conv3 = conv1x1(width, planes * self.expansion)
self.bn3 = get_norm(norm_type, planes * self.expansion)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
base_width=64, dilation=1, norm_type='Unknown'):
super(BasicBlock, self).__init__()
if groups != 1 or base_width != 64:
raise ValueError('BasicBlock only supports groups=1 and base_width=64')
if dilation > 1:
raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
# Both self.conv1 and self.downsample layers downsample the input when stride != 1
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = get_norm(norm_type, planes)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = get_norm(norm_type, planes)
self.downsample = downsample
self.stride = stride
def __init__(self, block, layers, num_classes=1000, norm_type='Unknown',
zero_init_residual=False, groups=1, width_per_group=64,
replace_stride_with_dilation=None, norm_layer=None):
super(ResNetFPN, self).__init__()
self.inplanes = 64
self.dilation = 1
if replace_stride_with_dilation is None:
# each element in the tuple indicates if we should replace
# the 2x2 stride with a dilated convolution instead
replace_stride_with_dilation = [False, False, False]
if len(replace_stride_with_dilation) != 3:
raise ValueError("replace_stride_with_dilation should be None "
"or a 3-element tuple, got {}".format(replace_stride_with_dilation))
self.groups = groups
self.base_width = width_per_group
self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1 = get_norm(norm_type, self.inplanes)
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],
norm_type=norm_type)
self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
dilate=replace_stride_with_dilation[0],
norm_type=norm_type)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
dilate=replace_stride_with_dilation[1],
norm_type=norm_type)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
dilate=replace_stride_with_dilation[2],
norm_type=norm_type)
self.layer5 = self._make_layer(block, 512, layers[4], stride=2,
norm_type=norm_type)
self.up2 = nn.ConvTranspose2d(self.inplanes, self.inplanes, 2, stride=2, padding=0,
bias=False)
self.merge1 = conv3x3(self.inplanes, self.inplanes)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(self.inplanes, num_classes)
self.num_features = self.inplanes
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm, nn.InstanceNorm2d)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# 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 __init__(self,
in_channels,
out_channels,
norm_type='Unknown',
**kwargs):
super(BasicConv2d, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs)
self.norm = get_norm(norm_type, out_channels, eps=0.001)
def __init__(self, num_input_features, num_output_features, norm_type='Unknown'):
super(_Transition, self).__init__()
self.add_module('norm', get_norm(norm_type, num_input_features))
self.add_module('relu', nn.ReLU(inplace=True))
self.add_module('conv',
nn.Conv2d(num_input_features, num_output_features, kernel_size=1, stride=1,
bias=False))
self.add_module('pool', nn.AvgPool2d(kernel_size=2, stride=2))
def __init__(self,
in_ch,
out_ch,
kernel_size=3,
stride=1,
padding=0,
bias=True,
norm_type='Unknown'):
super(Conv2dNormRelu, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(in_ch, out_ch, kernel_size, stride, padding, bias=bias),
get_norm(norm_type, out_ch), nn.ReLU(inplace=True))
def make_layers(cfg, batch_norm=False, norm_type='Unknown'):
layers = []
in_channels = 3
for v in cfg:
if v == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1)
if batch_norm:
layers += [
conv2d,
get_norm(norm_type, v),
nn.ReLU(inplace=True)
]
else:
layers += [conv2d, nn.ReLU(inplace=True)]
in_channels = v
return nn.Sequential(*layers)
def _make_layer(self, block, planes, blocks, stride=1, dilate=False, norm_type='Unknown'):
downsample = None
previous_dilation = self.dilation
if dilate:
self.dilation *= stride
stride = 1
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
conv1x1(self.inplanes, planes * block.expansion, stride),
get_norm(norm_type, planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
self.base_width, previous_dilation, norm_type))
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes, groups=self.groups,
base_width=self.base_width, dilation=self.dilation,
norm_type=norm_type))
return nn.Sequential(*layers)