def __init__(self, in_channels=3):
super(DenseNet, self).__init__()
self.conv1 = Conv2dCOB(in_channels=in_channels,
out_channels=3,
kernel_size=3,
padding=1)
self.conv2 = Conv2dCOB(in_channels=3,
out_channels=3,
kernel_size=3,
padding=1)
self.conv3 = Conv2dCOB(in_channels=6,
out_channels=3,
kernel_size=3,
padding=1)
self.conv4 = Conv2dCOB(in_channels=3,
out_channels=3,
kernel_size=3,
padding=1)
self.relu1 = ReLUCOB()
self.relu2 = ReLUCOB()
self.relu3 = ReLUCOB()
self.relu4 = ReLUCOB()
self.concat1 = Concat()
self.flatten = FlattenCOB()
self.fc1 = LinearCOB(in_channels * 32 * 32, 10)
def __init__(self,
num_input_features,
growth_rate,
bn_size,
drop_rate,
memory_efficient=False):
super(_DenseLayerCOB, self).__init__()
self.add_module('norm1', BatchNorm2dCOB(num_input_features)),
self.add_module('relu1', ReLUCOB(inplace=True)),
self.add_module(
'conv1',
Conv2dCOB(num_input_features,
bn_size * growth_rate,
kernel_size=1,
stride=1,
bias=False)),
self.add_module('norm2', BatchNorm2dCOB(bn_size * growth_rate)),
self.add_module('relu2', ReLUCOB(inplace=True)),
self.add_module(
'conv2',
Conv2dCOB(bn_size * growth_rate,
growth_rate,
kernel_size=3,
stride=1,
padding=1,
bias=False)),
self.drop_rate = float(drop_rate)
self.memory_efficient = memory_efficient
self.concat = Concat()
self.dropout = DropoutCOB(self.drop_rate)
def __init__(self):
super(ResidualNet5, self).__init__()
nb_channels = 3
self.conv1 = Conv2dCOB(in_channels=1,
out_channels=nb_channels,
kernel_size=3,
padding=1)
self.conv2 = Conv2dCOB(in_channels=nb_channels,
out_channels=nb_channels,
kernel_size=3,
padding=1)
self.conv3 = Conv2dCOB(in_channels=nb_channels,
out_channels=nb_channels,
kernel_size=3,
padding=1)
self.conv4 = Conv2dCOB(in_channels=nb_channels,
out_channels=nb_channels,
kernel_size=3)
self.conv11 = Conv2dCOB(in_channels=nb_channels,
out_channels=nb_channels,
kernel_size=3,
padding=1)
self.relu1 = ReLUCOB()
self.relu2 = ReLUCOB()
self.relu3 = ReLUCOB()
self.relu4 = ReLUCOB()
self.add1 = Add()
self.add2 = Add()
self.relu5 = ReLUCOB()
def __init__(self):
super(ResidualNet2, self).__init__()
self.conv1 = Conv2dCOB(in_channels=1,
out_channels=3,
kernel_size=3,
padding=1)
self.conv2 = Conv2dCOB(in_channels=3,
out_channels=3,
kernel_size=3,
padding=1)
self.conv3 = Conv2dCOB(in_channels=3,
out_channels=3,
kernel_size=3,
padding=1)
self.conv4 = Conv2dCOB(in_channels=3, out_channels=3, kernel_size=3)
self.relu1 = ReLUCOB()
self.relu2 = ReLUCOB()
self.relu3 = ReLUCOB()
self.relu4 = ReLUCOB()
self.add1 = Add()
self.add2 = Add()
self.flatten = FlattenCOB()
self.fc1 = LinearCOB(2028, 10)
self.relu5 = ReLUCOB()
def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
groups=1,
base_width=64,
dilation=1,
norm_layer=None):
super(BasicBlockCOB, self).__init__()
if norm_layer is None:
norm_layer = BatchNorm2dCOB
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 = norm_layer(planes)
self.relu1 = ReLUCOB(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = norm_layer(planes)
self.downsample = downsample
self.add = Add()
self.relu2 = ReLUCOB(inplace=True)
self.stride = stride
def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
groups=1,
base_width=64,
dilation=1,
norm_layer=None):
super(BottleneckCOB, self).__init__()
if norm_layer is None:
norm_layer = BatchNorm2dCOB
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 = norm_layer(width)
self.conv2 = conv3x3(width, width, stride, groups, dilation)
self.bn2 = norm_layer(width)
self.conv3 = conv1x1(width, planes * self.expansion)
self.bn3 = norm_layer(planes * self.expansion)
self.relu1 = ReLUCOB(inplace=True)
self.relu2 = ReLUCOB(inplace=True)
self.relu3 = ReLUCOB(inplace=True)
self.downsample = downsample
self.stride = stride
self.add = Add()
def __init__(self):
super(Net4, self).__init__()
self.conv1 = Conv2dCOB(1, 6, 5)
self.conv2 = Conv2dCOB(6, 3, 5)
self.relu1 = ReLUCOB()
self.relu2 = ReLUCOB()
self.bn1 = BatchNorm2dCOB(6)
self.bn2 = BatchNorm2dCOB(3)
def __init__(self, in_ch, out_ch):
super().__init__()
self.net = nn.Sequential(
Conv2dCOB(in_ch, out_ch, kernel_size=3, padding=1),
BatchNorm2dCOB(out_ch),
ReLUCOB(inplace=False),
Conv2dCOB(out_ch, out_ch, kernel_size=3, padding=1),
BatchNorm2dCOB(out_ch),
ReLUCOB(inplace=False),
)
def __init__(self):
super(Net, self).__init__()
self.conv1 = Conv2dCOB(1, 6, 5)
self.conv2 = Conv2dCOB(6, 16, 5)
self.fc1 = LinearCOB(16 * 20 * 20, 120)
self.fc2 = LinearCOB(120, 84)
self.fc3 = LinearCOB(84, 10)
self.flatten = FlattenCOB()
self.relu1 = ReLUCOB()
self.relu2 = ReLUCOB()
self.tanhcob = TanhCOB()
self.relu3 = ReLUCOB()
def __init__(self):
super(ResidualNet3, self).__init__()
self.conv1 = Conv2dCOB(in_channels=1,
out_channels=3,
kernel_size=3,
padding=1)
self.conv2 = Conv2dCOB(in_channels=3,
out_channels=3,
kernel_size=3,
padding=1)
self.relu1 = ReLUCOB()
self.relu2 = ReLUCOB()
self.add1 = Add()
self.relu3 = ReLUCOB()
def __init__(self):
super(DenseNet4, self).__init__()
self.conv1 = Conv2dCOB(in_channels=1,
out_channels=3,
kernel_size=3,
padding=1)
self.conv2 = Conv2dCOB(in_channels=3,
out_channels=3,
kernel_size=3,
padding=1)
self.conv3 = Conv2dCOB(in_channels=6,
out_channels=3,
kernel_size=3,
padding=1)
self.relu1 = ReLUCOB()
self.relu2 = ReLUCOB()
self.relu3 = ReLUCOB()
self.concat1 = Concat()
def __init__(self):
super(Net2, self).__init__()
self.conv1 = Conv2dCOB(1, 6, 5)
self.pool1 = MaxPool2dCOB(kernel_size=2)
self.pool2 = MaxPool2dCOB(kernel_size=2)
self.conv2 = Conv2dCOB(6, 16, 5)
self.flatten = FlattenCOB()
self.sigmoid = SigmoidCOB()
self.relu1 = ReLUCOB()
def __init__(self, num_input_features, num_output_features):
super(_TransitionCOB, self).__init__()
self.add_module('norm', BatchNorm2dCOB(num_input_features))
self.add_module('relu', ReLUCOB(inplace=True))
self.add_module(
'conv',
Conv2dCOB(num_input_features,
num_output_features,
kernel_size=1,
stride=1,
bias=False))
self.add_module('pool', AvgPool2dCOB(kernel_size=2, stride=2))
def __init__(self):
super(SplitConcatModel, self).__init__()
self.conv1 = Conv2dCOB(in_channels=1,
out_channels=3,
kernel_size=3,
padding=1)
self.conv21 = Conv2dCOB(in_channels=3,
out_channels=3,
kernel_size=3,
padding=1)
self.conv22 = Conv2dCOB(in_channels=3,
out_channels=3,
kernel_size=3,
padding=1)
self.conv3 = Conv2dCOB(in_channels=6,
out_channels=3,
kernel_size=3,
padding=1)
self.relu1 = ReLUCOB()
self.relu21 = ReLUCOB()
self.relu22 = ReLUCOB()
self.concat1 = Concat()
def __init__(self,
input_shape,
num_classes,
hidden_layers=(500, 500, 500, 500, 500),
activation="relu"):
super().__init__()
# Create the input and hidden layers
layers_dim = ((reduce(mul, input_shape), ) + tuple(hidden_layers))
layers = []
for idx in range(len(hidden_layers)):
layers.append(LinearCOB(layers_dim[idx], layers_dim[idx + 1]))
if activation == 'elu':
layers.append(ELUCOB())
elif activation == 'leakyrelu':
layers.append(LeakyReLUCOB())
elif activation == 'tanh':
layers.append(TanhCOB())
else:
layers.append(ReLUCOB())
self.net = torch.nn.Sequential(
FlattenCOB(), *layers, LinearCOB(hidden_layers[-1], num_classes))
def __init__(self,
growth_rate,
block_config,
num_init_features,
num_classes,
bn_size=4,
drop_rate=0,
input_channels=3,
memory_efficient=False):
super(DenseNetCOB, self).__init__()
# First convolution
self.features = nn.Sequential(
OrderedDict([
('conv0',
Conv2dCOB(input_channels,
num_init_features,
kernel_size=7,
stride=2,
padding=3,
bias=False)),
('norm0', BatchNorm2dCOB(num_init_features)),
('relu0', ReLUCOB(inplace=True)),
('pool0', MaxPool2dCOB(kernel_size=3, stride=2, padding=1)),
]))
# Each denseblock
num_features = num_init_features
for i, num_layers in enumerate(block_config):
block = _DenseBlockCOB(num_layers=num_layers,
num_input_features=num_features,
bn_size=bn_size,
growth_rate=growth_rate,
drop_rate=drop_rate,
memory_efficient=memory_efficient)
self.features.add_module('denseblock%d' % (i + 1), block)
num_features = num_features + num_layers * growth_rate
if i != len(block_config) - 1:
trans = _TransitionCOB(num_input_features=num_features,
num_output_features=num_features // 2)
self.features.add_module('transition%d' % (i + 1), trans)
num_features = num_features // 2
# Final batch norm
self.features.add_module('norm5', BatchNorm2dCOB(num_features))
# Flatten layer
self.relu = ReLUCOB(inplace=True)
self.adaptive_avg_pool2d = AdaptiveAvgPool2dCOB((1, 1))
self.flatten = FlattenCOB()
# Linear layer
self.classifier = LinearCOB(num_features, num_classes)
# Official init from torch repo.
for m in self.modules():
if isinstance(m, Conv2dCOB):
nn.init.kaiming_normal_(m.weight)
elif isinstance(m, BatchNorm2dCOB):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, LinearCOB):
nn.init.constant_(m.bias, 0)
def __init__(self):
super().__init__()
self.net = torch.nn.Sequential(FlattenCOB(), LinearCOB(784, 128),
BatchNorm1dCOB(128), ReLUCOB(),
LinearCOB(128, 10))
def __init__(self):
super(ConvTransposeNet, self).__init__()
self.conv1 = Conv2dCOB(1, 6, 5)
self.conv2 = ConvTranspose2dCOB(6, 3 // 2, kernel_size=2, stride=2)
self.relu1 = ReLUCOB(inplace=True)
self.relu2 = ReLUCOB(inplace=True)
def __init__(self,
block,
layers,
num_classes,
zero_init_residual=False,
groups=1,
width_per_group=64,
replace_stride_with_dilation=None,
norm_layer=None,
input_channels=3,
for_dataset=None):
super(ResNetCOB, self).__init__()
if norm_layer is None:
norm_layer = BatchNorm2dCOB
self._norm_layer = norm_layer
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 = Conv2dCOB(input_channels,
self.inplanes,
kernel_size=7,
stride=2,
padding=3,
bias=False)
if for_dataset == "cifar":
# CIFAR10: kernel_size 7 -> 3, stride 2 -> 1, padding 3->1
self.conv1 = Conv2dCOB(input_channels,
self.inplanes,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn1 = norm_layer(self.inplanes)
self.relu = ReLUCOB(inplace=True)
self.maxpool = MaxPool2dCOB(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block,
128,
layers[1],
stride=2,
dilate=replace_stride_with_dilation[0])
self.layer3 = self._make_layer(block,
256,
layers[2],
stride=2,
dilate=replace_stride_with_dilation[1])
self.layer4 = self._make_layer(block,
512,
layers[3],
stride=2,
dilate=replace_stride_with_dilation[2])
self.avgpool = AdaptiveAvgPool2dCOB((1, 1))
self.fc = LinearCOB(512 * block.expansion, num_classes)
self.flatten = FlattenCOB()
for m in self.modules():
if isinstance(m, Conv2dCOB):
nn.init.kaiming_normal_(m.weight,
mode='fan_out',
nonlinearity='relu')
elif isinstance(m, (BatchNorm2dCOB, nn.GroupNorm)):
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, BottleneckCOB):
nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlockCOB):
nn.init.constant_(m.bn2.weight, 0)
