def __init__(self,
inplanes,
planes,
groups,
reduction,
stride=1,
downsample=None,
base_width=4):
super(SEResNeXtBottleneck, self).__init__()
width = math.floor(planes * (base_width / 64)) * groups
self.conv1 = nn.Conv2d(inplanes,
width,
kernel_size=1,
bias=False,
stride=1)
self.bn1 = SyncBN2d(width)
self.conv2 = nn.Conv2d(width,
width,
kernel_size=3,
stride=stride,
padding=1,
groups=groups,
bias=False)
self.bn2 = SyncBN2d(width)
self.conv3 = nn.Conv2d(width, planes * 4, kernel_size=1, bias=False)
self.bn3 = SyncBN2d(planes * 4)
self.relu = nn.ReLU(inplace=True)
self.se_module = SEModule(planes * 4, reduction=reduction)
self.downsample = downsample
self.stride = stride
def __init__(self,
inplanes,
planes,
groups,
reduction,
stride=1,
downsample=None):
super(SEResNetBottleneck, self).__init__()
self.conv1 = nn.Conv2d(inplanes,
planes,
kernel_size=1,
bias=False,
stride=stride)
self.bn1 = SyncBN2d(planes)
self.conv2 = nn.Conv2d(planes,
planes,
kernel_size=3,
padding=1,
groups=groups,
bias=False)
self.bn2 = SyncBN2d(planes)
self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
self.bn3 = SyncBN2d(planes * 4)
self.relu = nn.ReLU(inplace=True)
self.se_module = SEModule(planes * 4, reduction=reduction)
self.downsample = downsample
self.stride = stride
def __init__(self, in_channels, out_channels, stride=1, downsample=None):
super(Bottleneck, self).__init__()
channels = out_channels // 4
self.conv1 = nn.Conv2d(in_channels,
channels,
kernel_size=1,
bias=False)
self.bn1 = SyncBN2d(channels)
self.conv2 = nn.Conv2d(channels,
channels,
kernel_size=3,
stride=stride,
padding=1,
bias=False)
self.bn2 = SyncBN2d(channels)
self.conv3 = nn.Conv2d(channels,
out_channels,
kernel_size=1,
bias=False)
self.bn3 = SyncBN2d(out_channels)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
if (self.downsample is None) and (in_channels != out_channels):
self.downsample = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=1,
bias=False), SyncBN2d(out_channels))
self.stride = stride
def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None):
super(BasicBlock, self).__init__()
self.conv1 = conv3x3(inplanes, planes, stride, dilation)
self.bn1 = SyncBN2d(planes)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = SyncBN2d(planes)
self.downsample = downsample
self.stride = stride
def __init__(self,
in_channels,
num_classes,
backend='resnet18',
pool_scales=(1, 2, 3, 6),
pretrained='imagenet'):
'''
:param in_channels:
:param num_classes:
:param backend: only support resnet, otherwise need to have low_features for aux
and high_features methods for out
'''
super(PSPNet, self).__init__()
self.in_channes = in_channels
self.num_classes = num_classes
if hasattr(backend, 'low_features') and hasattr(backend, 'high_features') \
and hasattr(backend, 'lastconv_channel'):
self.backend = backend
elif 'resnet' in backend:
## in the future, I may retrain on os=16 (now without os)
self.backend = ResnetBackend(backend,
os=None,
pretrained=pretrained)
elif 'resnext' in backend:
self.backend = ResnetBackend(backend,
os=None,
pretrained=pretrained)
else:
raise NotImplementedError
self.pp_out_channel = 256
self.pyramid_pooling = PPBlock(self.backend.lastconv_channel,
out_channel=self.pp_out_channel,
pool_scales=pool_scales)
self.cbr_last = nn.Sequential(
nn.Conv2d(self.backend.lastconv_channel +
self.pp_out_channel * len(pool_scales),
self.pp_out_channel,
kernel_size=3,
padding=1,
bias=False), SyncBN2d(self.pp_out_channel),
nn.ReLU(inplace=True), nn.Dropout2d(0.1),
nn.Conv2d(self.pp_out_channel, num_classes, kernel_size=1))
self.cbr_deepsup = nn.Sequential(
nn.Conv2d(self.backend.lastconv_channel // 2,
self.backend.lastconv_channel // 4,
kernel_size=3,
padding=1,
bias=False),
SyncBN2d(self.backend.lastconv_channel // 4),
nn.ReLU(inplace=True), nn.Dropout2d(0.1),
nn.Conv2d(self.backend.lastconv_channel // 4,
num_classes,
kernel_size=1))
def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None):
super(Bottleneck, self).__init__()
self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
self.bn1 = SyncBN2d(planes)
self.conv2 = conv3x3(planes, planes, stride=stride, dilation=dilation)
self.bn2 = SyncBN2d(planes)
self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False)
self.bn3 = SyncBN2d(planes * self.expansion)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
def __init__(self, in_channel, out_channel):
super(unetConv2d, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(in_channel,
out_channel,
kernel_size=3,
padding=1,
bias=False), SyncBN2d(out_channel),
nn.ReLU(inplace=True),
nn.Conv2d(out_channel,
out_channel,
kernel_size=3,
padding=1,
bias=False), SyncBN2d(out_channel),
nn.ReLU(inplace=True))
def __init__(self, block, layers, output_stride=16):
if output_stride == 16:
dilations = [1, 1, 1, 2]
strides = [1, 2, 2, 1]
elif output_stride == 8:
dilations = [1, 1, 2, 4]
strides = [1, 2, 1, 1]
elif output_stride is None:
dilations = [1, 1, 1, 1]
strides = [1, 2, 2, 2]
else:
raise Warning("output_stride must be 8 or 16 or None!")
self.inplanes = 64
super(ResNet, self).__init__()
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1 = SyncBN2d(64)
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])
self.layer2 = self._make_layer(block, 128, layers[1], stride=strides[1], dilation=dilations[1])
self.layer3 = self._make_layer(block, 256, layers[2], stride=strides[2], dilation=dilations[2])
self.layer4 = self._make_layer(block, 512, layers[3], stride=strides[3], dilation=dilations[3])
# self.LS = nn.LogSoftmax(dim=1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, SyncBN2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def _make_layer(self,
block,
planes,
blocks,
groups,
reduction,
stride=1,
downsample_kernel_size=1,
downsample_padding=0):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes,
planes * block.expansion,
kernel_size=downsample_kernel_size,
stride=stride,
padding=downsample_padding,
bias=False),
SyncBN2d(planes * block.expansion),
)
layers = []
layers.append(
block(self.inplanes, planes, groups, reduction, stride,
downsample))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes, groups, reduction))
return nn.Sequential(*layers)
def __init__(self,
in_channel=2048,
out_channel=256,
pool_scales=(1, 2, 3, 6)):
'''
:param in_channel: default 2048 for resnet50 backend
:param out_channel: default 256 for resnet50 backend
:param pool_scales: default scales [1,2,3,6]
'''
super(PPBlock, self).__init__()
self.pp_bra1 = nn.Sequential(
OrderedDict([('pool1', nn.AdaptiveAvgPool2d(pool_scales[0])),
('conv1',
nn.Conv2d(in_channel,
out_channel,
kernel_size=1,
bias=False)),
('bn1', SyncBN2d(out_channel)),
('relu1', nn.ReLU(inplace=True))]))
self.pp_bra2 = nn.Sequential(
OrderedDict([('pool2', nn.AdaptiveAvgPool2d(pool_scales[1])),
('conv2',
nn.Conv2d(in_channel,
out_channel,
kernel_size=1,
bias=False)),
('bn2', SyncBN2d(out_channel)),
('relu2', nn.ReLU(inplace=True))]))
self.pp_bra3 = nn.Sequential(
OrderedDict([('pool3', nn.AdaptiveAvgPool2d(pool_scales[2])),
('conv3',
nn.Conv2d(in_channel,
out_channel,
kernel_size=1,
bias=False)),
('bn3', SyncBN2d(out_channel)),
('relu3', nn.ReLU(inplace=True))]))
self.pp_bra4 = nn.Sequential(
OrderedDict([('pool4', nn.AdaptiveAvgPool2d(pool_scales[3])),
('conv4',
nn.Conv2d(in_channel,
out_channel,
kernel_size=1,
bias=False)),
('bn4', SyncBN2d(out_channel)),
('relu4', nn.ReLU(inplace=True))]))
def __init__(self, inp, oup, stride, dilation, expand_ratio):
super(InvertedResidual, self).__init__()
self.stride = stride
assert stride in [1, 2]
hidden_dim = round(inp * expand_ratio)
self.use_res_connect = self.stride == 1 and inp == oup
if expand_ratio == 1:
self.conv = nn.Sequential(
# dw
nn.Conv2d(hidden_dim,
hidden_dim,
3,
stride,
dilation,
dilation,
groups=hidden_dim,
bias=False),
SyncBN2d(hidden_dim),
nn.ReLU6(inplace=True),
# pw-linear
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
SyncBN2d(oup),
)
else:
self.conv = nn.Sequential(
# pw
nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
SyncBN2d(hidden_dim),
nn.ReLU6(inplace=True),
# dw
nn.Conv2d(hidden_dim,
hidden_dim,
3,
stride,
dilation,
dilation,
groups=hidden_dim,
bias=False),
SyncBN2d(hidden_dim),
nn.ReLU6(inplace=True),
# pw-linear
nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
SyncBN2d(oup),
)
def __init__(self,
in_channels,
num_classes,
backend='resnet18',
pretrained='imagenet'):
super(UnetResnetAE, self).__init__()
self.in_channes = in_channels
self.num_classes = num_classes
if 'resne' in backend:
self.encoder = ResnetBackend(backend, pretrained)
else:
raise NotImplementedError
if backend in ['resnet18', 'resnet34']:
block = BasicBlock
else:
block = Bottleneck
inter_channel = self.encoder.lastconv_channel
self.center = nn.Sequential(
nn.Conv2d(inter_channel, 512, kernel_size=3, padding=1,
bias=False), SyncBN2d(512), nn.ReLU(inplace=True),
nn.Conv2d(512, 256, kernel_size=3, padding=1, bias=False),
SyncBN2d(256), nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2, stride=2))
self.decoder5 = Decoder(block, 256 + inter_channel, 512, 64)
self.decoder4 = Decoder(block, 64 + inter_channel // 2, 256, 64)
self.decoder3 = Decoder(block, 64 + inter_channel // 4, 128, 64)
self.decoder2 = Decoder(block, 64 + inter_channel // 8, 64, 64)
self.decoder1 = Decoder(block, 64, 32, 64)
self.cbr_last = nn.Sequential(
nn.Conv2d(64 * 6, 64, kernel_size=3, padding=1, bias=False),
SyncBN2d(64), nn.ReLU(inplace=True),
nn.Conv2d(64, self.num_classes, kernel_size=1))
def __init__(self,
block,
layers,
groups,
reduction,
dropout_p=0.2,
inplanes=128,
input_3x3=True,
downsample_kernel_size=3,
downsample_padding=1,
num_classes=1000):
"""
Parameters
----------
block (nn.Module): Bottleneck class.
- For SENet154: SEBottleneck
- For SE-ResNet compare_models: SEResNetBottleneck
- For SE-ResNeXt compare_models: SEResNeXtBottleneck
layers (list of ints): Number of residual blocks for 4 layers of the
network (layer1...layer4).
groups (int): Number of groups for the 3x3 convolution in each
bottleneck block.
- For SENet154: 64
- For SE-ResNet compare_models: 1
- For SE-ResNeXt compare_models: 32
reduction (int): Reduction ratio for Squeeze-and-Excitation modules.
- For all compare_models: 16
dropout_p (float or None): Drop probability for the Dropout layer.
If `None` the Dropout layer is not used.
- For SENet154: 0.2
- For SE-ResNet compare_models: None
- For SE-ResNeXt compare_models: None
inplanes (int): Number of input channels for layer1.
- For SENet154: 128
- For SE-ResNet compare_models: 64
- For SE-ResNeXt compare_models: 64
input_3x3 (bool): If `True`, use three 3x3 convolutions instead of
a single 7x7 convolution in layer0.
- For SENet154: True
- For SE-ResNet compare_models: False
- For SE-ResNeXt compare_models: False
downsample_kernel_size (int): Kernel size for downsampling convolutions
in layer2, layer3 and layer4.
- For SENet154: 3
- For SE-ResNet compare_models: 1
- For SE-ResNeXt compare_models: 1
downsample_padding (int): Padding for downsampling convolutions in
layer2, layer3 and layer4.
- For SENet154: 1
- For SE-ResNet compare_models: 0
- For SE-ResNeXt compare_models: 0
num_classes (int): Number of outputs in `last_linear` layer.
- For all compare_models: 1000
"""
super(SENet, self).__init__()
self.inplanes = inplanes
if input_3x3:
layer0_modules = [
('conv1', nn.Conv2d(3, 64, 3, stride=2, padding=1,
bias=False)),
('bn1', SyncBN2d(64)),
('relu1', nn.ReLU(inplace=True)),
('conv2', nn.Conv2d(64, 64, 3, stride=1, padding=1,
bias=False)),
('bn2', SyncBN2d(64)),
('relu2', nn.ReLU(inplace=True)),
('conv3',
nn.Conv2d(64, inplanes, 3, stride=1, padding=1, bias=False)),
('bn3', SyncBN2d(inplanes)),
('relu3', nn.ReLU(inplace=True)),
]
else:
layer0_modules = [
('conv1',
nn.Conv2d(3,
inplanes,
kernel_size=7,
stride=2,
padding=3,
bias=False)),
('bn1', SyncBN2d(inplanes)),
('relu1', nn.ReLU(inplace=True)),
]
# To preserve compatibility with Caffe weights `ceil_mode=True`
# is used instead of `padding=1`.
layer0_modules.append(('pool', nn.MaxPool2d(3,
stride=2,
ceil_mode=True)))
self.layer0 = nn.Sequential(OrderedDict(layer0_modules))
self.layer1 = self._make_layer(block,
planes=64,
blocks=layers[0],
groups=groups,
reduction=reduction,
downsample_kernel_size=1,
downsample_padding=0)
self.layer2 = self._make_layer(
block,
planes=128,
blocks=layers[1],
stride=2,
groups=groups,
reduction=reduction,
downsample_kernel_size=downsample_kernel_size,
downsample_padding=downsample_padding)
self.layer3 = self._make_layer(
block,
planes=256,
blocks=layers[2],
stride=2,
groups=groups,
reduction=reduction,
downsample_kernel_size=downsample_kernel_size,
downsample_padding=downsample_padding)
self.layer4 = self._make_layer(
block,
planes=512,
blocks=layers[3],
stride=2,
groups=groups,
reduction=reduction,
downsample_kernel_size=downsample_kernel_size,
downsample_padding=downsample_padding)
self.avg_pool = nn.AvgPool2d(7, stride=1)
self.dropout = nn.Dropout(dropout_p) if dropout_p is not None else None
self.last_linear = nn.Linear(512 * block.expansion, num_classes)
def conv_1x1_bn(inp, oup):
return nn.Sequential(nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
SyncBN2d(oup), nn.ReLU6(inplace=True))
def conv_bn(inp, oup, stride):
return nn.Sequential(nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
SyncBN2d(oup), nn.ReLU6(inplace=True))
def __init__(self, in_channel=2048, out_channel=256, os=16):
'''
:param in_channel: default 2048 for resnet101
:param out_channel: default 256 for resnet101
:param os: 16 or 8
'''
super(ASPPBlock, self).__init__()
if os == 16:
rates = [1, 6, 12, 18]
elif os == 8:
rates = [1, 12, 24, 36]
else:
raise NotImplementedError
self.gave_pool = nn.Sequential(
OrderedDict([('gavg', nn.AdaptiveAvgPool2d(rates[0])),
('conv0_1',
nn.Conv2d(in_channel,
out_channel,
kernel_size=1,
bias=False)),
('bn0_1', SyncBN2d(out_channel)),
('relu0_1', nn.ReLU(inplace=True))]))
self.conv1_1 = nn.Sequential(
OrderedDict([('conv0_2',
nn.Conv2d(in_channel,
out_channel,
kernel_size=1,
bias=False)),
('bn0_2', SyncBN2d(out_channel)),
('relu0_2', nn.ReLU(inplace=True))]))
self.aspp_bra1 = nn.Sequential(
OrderedDict([('conv1_1',
nn.Conv2d(in_channel,
out_channel,
kernel_size=3,
padding=rates[1],
dilation=rates[1],
bias=False)),
('bn1_1', SyncBN2d(out_channel)),
('relu1_1', nn.ReLU(inplace=True))]))
self.aspp_bra2 = nn.Sequential(
OrderedDict([('conv1_2',
nn.Conv2d(in_channel,
out_channel,
kernel_size=3,
padding=rates[2],
dilation=rates[2],
bias=False)),
('bn1_2', SyncBN2d(out_channel)),
('relu1_2', nn.ReLU(inplace=True))]))
self.aspp_bra3 = nn.Sequential(
OrderedDict([('conv1_3',
nn.Conv2d(in_channel,
out_channel,
kernel_size=3,
padding=rates[3],
dilation=rates[3],
bias=False)),
('bn1_3', SyncBN2d(out_channel)),
('relu1_3', nn.ReLU(inplace=True))]))
self.aspp_catdown = nn.Sequential(
OrderedDict([('conv_down',
nn.Conv2d(5 * out_channel,
out_channel,
kernel_size=1,
bias=False)),
('bn_down', SyncBN2d(out_channel)),
('relu_down', nn.ReLU(inplace=True)),
('drop_out', nn.Dropout(.1))]))
def __init__(self,
in_channels,
num_classes,
backend='resnet18',
os=16,
pretrained='imagenet'):
'''
:param in_channels:
:param num_classes:
:param backend: only support resnet, otherwise need to have low_features
and high_features methods for out
'''
super(DeepLabv3_plus, self).__init__()
self.in_channes = in_channels
self.num_classes = num_classes
if hasattr(backend, 'low_features') and hasattr(backend, 'high_features') \
and hasattr(backend, 'lastconv_channel'):
self.backend = backend
elif 'resnet' in backend:
self.backend = ResnetBackend(backend, os, pretrained)
elif 'resnext' in backend:
self.backend = ResnetBackend(backend,
os=None,
pretrained=pretrained)
elif 'mobilenet' in backend:
self.backend = MobileNetBackend(backend,
os=os,
pretrained=pretrained)
elif 'shufflenet' in backend:
self.backend = ShuffleNetBackend(backend,
os=os,
pretrained=pretrained)
else:
raise NotImplementedError
if hasattr(self.backend, 'interconv_channel'):
self.aspp_out_channel = self.backend.interconv_channel
else:
self.aspp_out_channel = self.backend.lastconv_channel // 8
self.aspp_pooling = ASPPBlock(self.backend.lastconv_channel,
self.aspp_out_channel, os)
self.cbr_low = nn.Sequential(
nn.Conv2d(self.aspp_out_channel,
self.aspp_out_channel // 5,
kernel_size=1,
bias=False), SyncBN2d(self.aspp_out_channel // 5),
nn.ReLU(inplace=True))
self.cbr_last = nn.Sequential(
nn.Conv2d(self.aspp_out_channel + self.aspp_out_channel // 5,
self.aspp_out_channel,
kernel_size=3,
padding=1,
bias=False), SyncBN2d(self.aspp_out_channel),
nn.ReLU(inplace=True),
nn.Conv2d(self.aspp_out_channel,
self.aspp_out_channel,
kernel_size=3,
padding=1,
bias=False), SyncBN2d(self.aspp_out_channel),
nn.ReLU(inplace=True),
nn.Conv2d(self.aspp_out_channel, self.num_classes, kernel_size=1))
