def __init__(self):
super(Net, self).__init__()
self.cnn_layers = Sequential()
input_channels = configuration.input_channels
width = configuration.image_width
height = configuration.image_height
i = 1
for k, f, a, s, p, mf, ms, mp in zip(
configuration.no_conv_filters, configuration.conv_filters, configuration.activation_list,
configuration.conv_strides, configuration.conv_padding, configuration.maxpool_filters,
configuration.maxpool_strides, configuration.maxpool_padding):
self.cnn_layers.add_module("conv_2d_{}".format(i),
Conv2d(input_channels, k, kernel_size=f, stride=s, padding=p))
self.cnn_layers.add_module("dropout{}".format(i), Dropout(configuration.cnn_dropout))
width = (width - f + 2 * p) / s + 1
height = (height - f + 2 * p) / s + 1
input_channels = k
self.cnn_layers.add_module("activation_{}".format(i), get_activation_from_name(a))
if configuration.batch_norm:
self.cnn_layers.add_module("batch_norm{}".format(i), BatchNorm2d(input_channels))
self.cnn_layers.add_module("max_pool_{}".format(i), MaxPool2d(kernel_size=mf, stride=ms, padding=mp))
width = int((width - mf + 2 * mp) / ms + 1)
height =int((height - mf + 2 * mp) / ms + 1)
i += 1
self.linear_layers = Sequential()
self.linear_layers.add_module("flatten",Flatten())
self.linear_layers.add_module("linear_{}".format(i),Linear(input_channels * width * height, configuration.linear))
self.linear_layers.add_module("dropout_{}".format(i),Dropout(configuration.fcn_dropout))
self.linear_layers.add_module("activation_{}".format(i),get_activation_from_name(configuration.relu))
if configuration.batch_norm:
self.linear_layers.add_module("batch_norm{}".format(i),BatchNorm1d(configuration.linear))
self.linear_layers.add_module("output_{}".format(i),Linear(configuration.linear, configuration.output_classes))
def __init__(self,
depth,
input_size,
output_size,
dropout=0.0,
batch_norm=True):
super().__init__()
non_lin = ReLU()
assert dropout == 0, "Dropout not allowed for ResNet"
assert batch_norm
# Configure ResNet
# number of layers per block
if (depth - 2) % 6 != 0:
raise ValueError(
"Only ResNet-X with X=6N-2 allowed where N natural number.")
layers = (depth - 2) // 6
# number of filters for each block
filters = [16, 32, 64]
# Subsampling at first layer in each block
strides = [1, 2, 2]
# Build ResNet
self.conv1 = conv3x3(int(input_size[0]), filters[0])
self.inplanes = filters[0]
self.bn1 = BatchNorm2d(filters[0])
self.non_lin1 = non_lin
# Each layer makes the input smaller using stride, then adds `layers` layers of two layers each with a skip
self.layer1 = self._make_layer(filters[0], layers, strides[0], non_lin)
self.layer2 = self._make_layer(filters[1], layers, strides[1], non_lin)
self.layer3 = self._make_layer(filters[2], layers, strides[2], non_lin)
# Average each filter
self.avgpool = AvgPool2d(
int(input_size[1]) // reduce(operator.mul, strides, 1))
self.linear_out = Linear(filters[-1], output_size)
self.kind = depth
self.non_linearity = non_lin.__class__
self.batch_norm = batch_norm
def __init__(self, in_channel, depth, stride):
super(bottleneck_IR, self).__init__()
self.stride = stride
if stride == 2:
self.shortcut_layer = Sequential(Conv2d(in_channel, depth, (1, 1), stride ,bias=False),
BatchNorm2d(depth,eps=2e-5,momentum=0.9))
self.res_layer = Sequential(
BatchNorm2d(in_channel,eps=2e-5,momentum=0.9),
Conv2d(in_channel, depth, (3, 3), (1, 1), 1 ,bias=False),BatchNorm2d(depth,eps=2e-5,momentum=0.9),
PReLU(depth),Conv2d(depth, depth, (3, 3), stride, 1 ,bias=False), BatchNorm2d(depth,eps=2e-5,momentum=0.9))
else:
self.res_layer = Sequential(
BatchNorm2d(in_channel,eps=2e-5,momentum=0.9),
Conv2d(in_channel, depth, (3, 3), (1, 1), 1 ,bias=False),BatchNorm2d(depth,eps=2e-5,momentum=0.9),
PReLU(depth), Conv2d(depth, depth, (3, 3), stride, 1 ,bias=False), BatchNorm2d(depth,eps=2e-5,momentum=0.9))
def __init__(self):
super(Net1, self).__init__()
self.features = nn.Sequential(
DSC(5, 64, 3, 2, 1), BatchNorm2d(64), ReLU(inplace=True),
DSC(64, 192, 3, 2, 1), BatchNorm2d(192), ReLU(inplace=True),
DSC(192, 384, 3, 1, 1), BatchNorm2d(384), ReLU(inplace=True),
MaxPool2d(kernel_size=3,
stride=2,
padding=0,
dilation=1,
ceil_mode=False), DSC(384, 512, 3, 1,
1), BatchNorm2d(512),
ReLU(inplace=True),
MaxPool2d(kernel_size=3,
stride=2,
padding=0,
dilation=1,
ceil_mode=False), DSC(512, 512, 3, 1, 1),
BatchNorm2d(512), ReLU(inplace=True), DSC(512, 1024, 3, 1, 1),
BatchNorm2d(1024), ReLU(inplace=True),
MaxPool2d(kernel_size=3,
stride=2,
padding=0,
dilation=1,
ceil_mode=False), DSC(1024, 1024, 3, 1, 1),
BatchNorm2d(1024), ReLU(inplace=True), DSC(1024, 512, 3, 1, 1),
BatchNorm2d(512), ReLU(inplace=True), DSC(512, 256, 3, 1, 1),
BatchNorm2d(256), ReLU(inplace=True), CBAM(256))
self.avgpool = AdaptiveAvgPool2d(output_size=(6, 6))
self.classifier = nn.Sequential(
Dropout(p=0.5),
Linear(in_features=9216, out_features=2048, bias=True),
ReLU(inplace=True), Dropout(p=0.5),
Linear(in_features=2048, out_features=1024, bias=True),
ReLU(inplace=True), Dropout(p=0.5),
Linear(in_features=1024, out_features=3, bias=True))
def __init__(self, depth, widen_factor, dropout_rate, num_classes, beta=1.0, base=16, name=None):
super(Wide_ResNet, self).__init__()
self.in_planes, self.num_classes = base, num_classes
self.beta = beta
assert ((depth - 4) % 6 == 0), 'Wide-resnet depth should be 6n+4'
n = (depth - 4) / 6
k = widen_factor
print('| Wide-Resnet %dx%d' % (depth, k))
nStages = [base, base * k, base * k * 2, base * k * 4]
self.conv1 = conv3x3(3, nStages[0])
self.layer1 = self._wide_layer(wide_basic, nStages[1], n, dropout_rate, stride=1)
self.layer2 = self._wide_layer(wide_basic, nStages[2], n, dropout_rate, stride=2)
self.layer3 = self._wide_layer(wide_basic, nStages[3], n, dropout_rate, stride=2)
self.bn1 = BatchNorm2d(nStages[3])
self.linear = nn.Linear(nStages[3], num_classes, bias=True)
self.name = name
def __init__(self, scale_factor):
upsample_block_num = int(math.log(scale_factor, 2))
super(Generator, self).__init__()
self.block1 = Sequential(
Conv2d(3, 64, kernel_size=(9, 9), padding=(4, 4)), PReLU())
self.block2 = ResidualBlock(64)
self.block3 = ResidualBlock(64)
self.block4 = ResidualBlock(64)
self.block5 = ResidualBlock(64)
self.block6 = ResidualBlock(64)
self.block7 = Sequential(
Conv2d(64, 64, kernel_size=(3, 3), padding=(1, 1)),
BatchNorm2d(64))
# up sampling
block8 = [UpSampleBlock(64, 2) for _ in range(upsample_block_num)]
block8.append(Conv2d(64, 3, kernel_size=(9, 9), padding=(4, 4)))
self.block8 = Sequential(*block8)
def __init__(self,
K_size=2,
stride_size=1,
padding_size=0,
hights=5,
widths=5,
outs=5):
super(Net, self).__init__()
self.K_h = K_size[0] if isinstance(K_size, tuple) else K_size
self.K_w = K_size[1] if isinstance(K_size, tuple) else K_size
self.Stride_h = stride_size[0] if isinstance(stride_size,
tuple) else stride_size
self.Stride_w = stride_size[1] if isinstance(stride_size,
tuple) else stride_size
self.Padding_h = padding_size[0] if isinstance(padding_size,
tuple) else padding_size
self.Padding_w = padding_size[1] if isinstance(padding_size,
tuple) else padding_size
self.H = hights
self.W = widths
self.out_size = outs
self.cnn_output_h = int((self.H + 2 * self.Padding_h -
(self.K_h - 1) - 1) / self.Stride_h + 1)
self.cnn_output_w = int((self.W + 2 * self.Padding_w -
(self.K_w - 1) - 1) / self.Stride_w + 1)
self.cnn_layers = Sequential(
# Defining a 2D convolution layer
Conv2d(1,
1,
kernel_size=(self.K_h, self.K_w),
stride=(self.Stride_h, self.Stride_w),
padding=(self.Padding_h, self.Padding_w)),
BatchNorm2d(1),
ReLU(inplace=True),
MaxPool2d(kernel_size=(1, self.cnn_output_w),
stride=(1, self.cnn_output_w)))
self.linear_layers = Sequential(
Linear(self.cnn_output_h, int(self.out_size)))
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int,
stride: int,
padding: int,
):
super().__init__()
self.conv = Conv2d(
in_channels,
out_channels,
kernel_size,
stride,
padding,
bias=False,
)
self.bn = BatchNorm2d(out_channels)
self.act = ReLU(num_channels=out_channels, inplace=True)
def get_resnet(encoder_name, pretrained=True):
"""
Returns a resnet18 or resnet50 model, with AdaptiveAvgPool2d
:param adaptive_pool_dim: dimension of feature output from adaptive
average pooling h * w
:param pretrained: if the model should be Imagenet pretrained
:return: model, cnn_features_dim
"""
if encoder_name == "resnet18":
model = models.resnet18(pretrained=pretrained)
latent_dim = 512
else:
model = models.resnet50(pretrained=pretrained)
latent_dim = 2048
children = ([BatchNorm2d(3)] + list(model.children())[:-2])
model = torch.nn.Sequential(*children)
return model, latent_dim
def __init__(self, inplanes, planes, dilation):
super(ASPP_module, self).__init__()
if dilation == 1:
kernel_size = 1
padding = 0
else:
kernel_size = 3
padding = dilation
self.atrous_convolution = nn.Conv2d(inplanes,
planes,
kernel_size=kernel_size,
stride=1,
padding=padding,
dilation=dilation,
bias=False)
self.bn = BatchNorm2d(planes)
self.relu = nn.ReLU()
self._init_weight()
def __init__(self):
super(Discriminator, self).__init__()
self.conv = Sequential(
Conv2d(in_channels=CONFIG["IMAGE_CHANNEL"], out_channels=64, kernel_size=4, stride=2, padding=1),
LeakyReLU(negative_slope=0.2),
Conv2d(in_channels=64, out_channels=128, kernel_size=4, stride=2, padding=1),
BatchNorm2d(num_features=128),
LeakyReLU(negative_slope=0.2)
)
self.fc = Sequential(
Linear(128 * 7 * 7, 1024),
BatchNorm1d(1024),
LeakyReLU(negative_slope=0.2),
Linear(1024, 1),
Sigmoid()
)
initialize_weights(self)
def __init__(self, num_layers, mode='ir'):
super(Backbone, self).__init__()
assert num_layers in [50, 100,
152], 'num_layers should be 50,100, or 152'
assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
blocks = get_blocks(num_layers)
if mode == 'ir':
unit_module = bottleneck_IR
elif mode == 'ir_se':
unit_module = bottleneck_IR_SE
self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False),
BatchNorm2d(64), PReLU(64))
modules = []
for block in blocks:
for bottleneck in block:
modules.append(
unit_module(bottleneck.in_channel, bottleneck.depth,
bottleneck.stride))
self.body = Sequential(*modules)
def __init__(self, features, out_features=512, sizes=(1, 2, 3, 6)):
super(PSPModule, self).__init__()
self.stages = []
self.stages = nn.ModuleList(
[self._make_stage(features, out_features, size) for size in sizes])
self.bottleneck = nn.Sequential(
nn.Conv2d(
features + len(sizes) * out_features,
out_features,
kernel_size=3,
padding=1,
dilation=1,
bias=False,
),
BatchNorm2d(out_features),
nn.ReLU(),
nn.Dropout2d(0.1),
)
def _make_layer(self, block, planes, blocks, stride=1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False),
BatchNorm2d(planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes))
return nn.Sequential(*layers)
def __init__(self, num_layers, mode='ir', opts=None):
super(GradualStyleEncoder, self).__init__()
assert num_layers in [50, 100,
152], 'num_layers should be 50,100, or 152'
assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
blocks = get_blocks(num_layers)
if mode == 'ir':
unit_module = bottleneck_IR
elif mode == 'ir_se':
unit_module = bottleneck_IR_SE
self.input_layer = Sequential(
Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False),
BatchNorm2d(64), PReLU(64))
modules = []
for block in blocks:
for bottleneck in block:
modules.append(
unit_module(bottleneck.in_channel, bottleneck.depth,
bottleneck.stride))
self.body = Sequential(*modules)
self.styles = nn.ModuleList()
self.style_count = opts.n_styles
self.coarse_ind = 3
self.middle_ind = 7
for i in range(self.style_count):
if i < self.coarse_ind:
style = GradualStyleBlock(512, 512, 16)
elif i < self.middle_ind:
style = GradualStyleBlock(512, 512, 32)
else:
style = GradualStyleBlock(512, 512, 64)
self.styles.append(style)
self.latlayer1 = nn.Conv2d(256,
512,
kernel_size=1,
stride=1,
padding=0)
self.latlayer2 = nn.Conv2d(128,
512,
kernel_size=1,
stride=1,
padding=0)
def _make_layer(self,
block,
planes,
blocks,
stride=1,
dilation=1,
multi_grid=1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(
self.inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False,
),
BatchNorm2d(planes * block.expansion, affine=True),
)
layers = []
generate_multi_grid = (lambda index, grids: grids[index % len(grids)]
if isinstance(grids, tuple) else 1)
layers.append(
block(
self.inplanes,
planes,
stride,
dilation=dilation,
downsample=downsample,
multi_grid=generate_multi_grid(0, multi_grid),
))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(
block(
self.inplanes,
planes,
dilation=dilation,
multi_grid=generate_multi_grid(i, multi_grid),
))
return nn.Sequential(*layers)
def __init__(self,
in_c,
out_c,
kernel=(1, 1),
stride=(1, 1),
padding=(0, 0),
groups=1):
super(Conv_block, self).__init__()
self.conv = Conv2d(
in_c,
out_channels=out_c,
kernel_size=kernel,
groups=groups,
stride=stride,
padding=padding,
bias=False,
)
self.bn = BatchNorm2d(out_c)
self.prelu = PReLU(out_c)
def conv(in_channels, out_channels, kernel_size=3, padding=1, batch_norm=True):
"""
A convolution block with a conv layer and batch norm
:param in_channels: number of input channels
:param out_channels: number of output channels
:param kernel_size: size of the kernel
:param padding: number of pixels to pad on all sides
:param batch_norm: to use batch norm or not
:return: PyTorch Tensor
"""
c = Conv2d(in_channels,
out_channels,
kernel_size=kernel_size,
stride=1,
padding=padding)
if batch_norm:
bn = BatchNorm2d(out_channels)
return Sequential(c, bn)
return c
def __init__(self):
super(Generator, self).__init__()
self.fc = Sequential(
Linear(CONFIG["NOISE_DIM"] + 10, 1024),
BatchNorm1d(1024),
ReLU(),
Linear(1024, 128 * 7 * 7),
BatchNorm1d(128 * 7 * 7),
ReLU(),
)
self.deconv = Sequential(
ConvTranspose2d(128, 64, 4, 2, 1),
BatchNorm2d(64),
ReLU(),
ConvTranspose2d(64, CONFIG["IMAGE_CHANNEL"], 4, 2, 1),
Tanh(),
)
normal_init(self)
def __init__(self):
super(GCPACC2Net, self).__init__()
self.hardnet = hardnet(arch=68)
inplanes = 1024
interplanes = 256
self.fam45 = FAM(640, interplanes, interplanes, interplanes)
self.fam34 = FAM(320, interplanes, interplanes, interplanes)
self.fam23 = FAM(128, interplanes, interplanes, interplanes)
self.linear5 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear4 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear3 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear2 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.conva = nn.Sequential(
nn.Conv2d(inplanes, interplanes, 3, padding=1, bias=False),
BatchNorm2d(interplanes),
nn.ReLU(interplanes),
)
self.long_relation = CrissCrossAttention(interplanes)
self.local_attention_4 = LocalAttenModule(interplanes)
self.local_attention_3 = LocalAttenModule(interplanes)
self.local_attention_2 = LocalAttenModule(interplanes)
def __init__(self, in_channels, out_channels):
super(ASPP, self).__init__()
dilations = [1, 6, 12, 18]
self.aspp0 = _ASPPModule(in_channels=in_channels, out_channels=out_channels, kernel_size=1,
padding=0, dilation=dilations[0])
self.aspp1 = _ASPPModule(in_channels=in_channels, out_channels=out_channels, kernel_size=3,
padding=dilations[1], dilation=dilations[1])
self.aspp2 = _ASPPModule(in_channels=in_channels, out_channels=out_channels, kernel_size=3,
padding=dilations[2], dilation=dilations[2])
self.aspp3 = _ASPPModule(in_channels=in_channels, out_channels=out_channels, kernel_size=3,
padding=dilations[3], dilation=dilations[3])
self.conv = Conv2d(out_channels * 4, out_channels, kernel_size=3, stride=1, padding=1, dilation=1, groups=1,
bias=True)
self.relu = ReLU()
self.batchnorm = BatchNorm2d(out_channels)
self._init_weight()
def generatePathBNs(self, srcLayer):
if self != srcLayer:
# get current BN
currBN = self.orgBNs()[self.currWidthIdx()]
# get current BN num_features
bnFeatures = currBN.num_features
# generate new BNs ModuleList
newBNs = ModuleList([
BatchNorm2d(bnFeatures) for _ in range(self.nWidths())
]).cuda()
# copy weights to new BNs
for bn in newBNs:
bn.load_state_dict(currBN.state_dict())
# set layer BNs
self.bn = newBNs
# update width List
self._widthList = [self.currWidth()] * self.nWidths()
# update width ratio list
self._widthRatioList = [self.currWidthRatio()] * self.nWidths()
def _create_backbone(self):
backbone = Sequential()
section1 = Sequential()
section1.add_module('Conv1', Conv2d(in_channels=3, out_channels=32, kernel_size=3, stride=1, padding=1))
section1.add_module('BatchNorm1', BatchNorm2d(num_features=32))
section1.add_module('Activation1', LeakyReLU(negative_slope=.1, inplace=True))
backbone.add_module('Section1', section1)
backbone.add_module('Section2', YOLOv3Block(index=2, in_channels=32, out_channels=64, convs_residual_count=1))
backbone.add_module('Section3', YOLOv3Block(index=5, in_channels=64, out_channels=128, convs_residual_count=2))
backbone.add_module('Section4',
YOLOv3Block(index=10, in_channels=128, out_channels=256, convs_residual_count=8))
backbone.add_module('Section5',
YOLOv3Block(index=27, in_channels=256, out_channels=512, convs_residual_count=8))
backbone.add_module('Section6',
YOLOv3Block(index=44, in_channels=512, out_channels=1024, convs_residual_count=4))
return backbone
def __init__(self, number_of_classes, anchors_dims):
super(Tail, self).__init__()
self.num_of_yolo_layers = 3
route_streams = [0, 2**3, 2**2]
self.tails = ModuleList([
ModuleList([
Sequential(
Conv2d((2**(5 - i) + route_streams[i]) *
filters_multiplier, 2**(4 - i) *
filters_multiplier, **bottleneck),
BatchNorm2d(2**(4 - i) *
filters_multiplier), LeakyReLU(negative_slope),
Conv2d(2**(4 - i) * filters_multiplier, 2**(5 - i) *
filters_multiplier, **casual),
BatchNorm2d(2**(5 - i) *
filters_multiplier), LeakyReLU(negative_slope),
Conv2d(2**(5 - i) * filters_multiplier, 2**(4 - i) *
filters_multiplier, **bottleneck),
BatchNorm2d(2**(4 - i) *
filters_multiplier), LeakyReLU(negative_slope),
Conv2d(2**(4 - i) * filters_multiplier, 2**(5 - i) *
filters_multiplier, **casual),
BatchNorm2d(2**(5 - i) * filters_multiplier),
LeakyReLU(negative_slope)),
Sequential(
Conv2d(2**(5 - i) * filters_multiplier, 2**(4 - i) *
filters_multiplier, **bottleneck),
BatchNorm2d(2**(4 - i) * filters_multiplier),
LeakyReLU(negative_slope)),
Sequential(
Conv2d(2**(4 - i) * filters_multiplier, 2**(5 - i) *
filters_multiplier, **casual),
BatchNorm2d(2**(5 - i) * filters_multiplier),
LeakyReLU(negative_slope)),
Sequential(
Conv2d(2**(5 - i) * filters_multiplier, anchors_dims[i] *
(number_of_classes + 5), **prelude))
] + [
Sequential(
Conv2d(2**(4 - i) * filters_multiplier, 2**
(3 - i) * filters_multiplier, **bottleneck),
BatchNorm2d(2**(3 - i) *
filters_multiplier), LeakyReLU(negative_slope))
] * (i < 2)) for i in range(self.num_of_yolo_layers)
])
def __init__(self, cfg):
super(ConvResStem, self).__init__()
C1 = cfg.MODEL.RESNETS.CONVRESSTEM.CONV_CHANNEL
C2 = cfg.MODEL.RESNETS.CONVRESSTEM.RES_CHANNEL1
n2 = cfg.MODEL.RESNETS.CONVRESSTEM.NUM_BLOCK_RES1
C3 = cfg.MODEL.RESNETS.STEM_OUT_CHANNELS
n3 = cfg.MODEL.RESNETS.CONVRESSTEM.NUM_BLOCK_RES2
res_module = cfg.MODEL.RESNETS.TRANS_FUNC
block_module = _TRANSFORMATION_MODULES[res_module]
convs = []
convs.append(conv3x3(3, C1))
from torch.nn import BatchNorm2d
convs.append(BatchNorm2d(C1))
convs.append(nn.ReLU(inplace=True))
self.stem_conv = nn.Sequential(*convs)
self.stem_res2 = _make_stage(
block_module,
in_channels=C1,
bottleneck_channels=C2//4,
out_channels=C2,
block_count=n2,
num_groups=1,
stride_in_1x1=cfg.MODEL.RESNETS.STRIDE_IN_1X1,
first_stride=1,
dilation=1,
dcn_config={},
)
self.stem_res3 = _make_stage(
block_module,
in_channels=C2,
bottleneck_channels=C3//4,
out_channels=C3,
block_count=n3,
num_groups=1,
stride_in_1x1=cfg.MODEL.RESNETS.STRIDE_IN_1X1,
first_stride=2,
dilation=1,
dcn_config={},
)
def __init__(self):
super(GCPAPSPSmallNet, self).__init__()
self.hardnet = hardnet(arch=68)
inplanes = 1024
interplanes = 128
self.fam45 = FAM(640, interplanes, interplanes, interplanes)
self.fam34 = FAM(320, interplanes, interplanes, interplanes)
self.fam23 = FAM(128, interplanes, interplanes, interplanes)
self.linear5 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear4 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear3 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear2 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.conva = nn.Sequential(
nn.Conv2d(inplanes, interplanes, 3, padding=1, bias=False),
BatchNorm2d(interplanes),
nn.ReLU(interplanes),
)
# self.long_relation = SpatialCGNL(interplanes, interplanes // 2)
self.long_relation = PSPModule(inplanes, interplanes)
self.local_attention_4 = SmallLocalAttenModule(interplanes)
self.local_attention_3 = SmallLocalAttenModule(interplanes)
self.local_attention_2 = SmallLocalAttenModule(interplanes)
def __init__(self, k=1, n=28, drop_rate=0):
super(WideResnetBackbone, self).__init__()
self.k, self.n = k, n
assert (self.n - 4) % 6 == 0
n_blocks = (self.n - 4) // 6
n_layers = [
16,
] + [self.k * 16 * (2**i) for i in range(3)]
self.conv1 = nn.Conv2d(3,
n_layers[0],
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.layer1 = self.create_layer(
n_layers[0],
n_layers[1],
bnum=n_blocks,
stride=1,
drop_rate=drop_rate,
pre_res_act=True,
)
self.layer2 = self.create_layer(
n_layers[1],
n_layers[2],
bnum=n_blocks,
stride=2,
drop_rate=drop_rate,
pre_res_act=False,
)
self.layer3 = self.create_layer(
n_layers[2],
n_layers[3],
bnum=n_blocks,
stride=2,
drop_rate=drop_rate,
pre_res_act=False,
)
self.bn_last = BatchNorm2d(n_layers[3], momentum=0.001)
self.relu_last = nn.LeakyReLU(inplace=True, negative_slope=0.1)
self.init_weight()
def __init__(self):
super(GCPACGNLResNet, self).__init__()
self.bkbone = ResNet()
inplanes = 2048
interplanes = 256
self.fam45 = FAM(1024, interplanes, interplanes, interplanes)
self.fam34 = FAM(512, interplanes, interplanes, interplanes)
self.fam23 = FAM(256, interplanes, interplanes, interplanes)
self.linear5 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear4 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear3 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.linear2 = nn.Conv2d(interplanes,
1,
kernel_size=3,
stride=1,
padding=1)
self.conva = nn.Sequential(
nn.Conv2d(inplanes, interplanes, 3, padding=1, bias=False),
BatchNorm2d(interplanes),
nn.ReLU(interplanes),
)
self.long_relation = SpatialCGNL(interplanes, interplanes // 2)
self.local_attention_4 = LocalAttenModule(interplanes)
self.local_attention_3 = LocalAttenModule(interplanes)
self.local_attention_2 = LocalAttenModule(interplanes)
self.local_attention = LocalAttenModule(interplanes)
def __init__(self, num_classes=10):
super().__init__()
self.in_channels = 64
# First - Conv2D 7x7 with 2 stride -> BatchNorm2D -> ReLU -> MaxPool2D
self.conv1 = Conv2d(1,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = BatchNorm2d(64)
self.relu = ReLU(inplace=True)
self.sigmoid = Sigmoid()
self.max_pool = MaxPool2d(kernel_size=3, stride=2, padding=1)
# Setup residual parts block groups
self.layer64 = self._make_residual_layer(out_channels=64, num_blocks=3)
self.layer128 = self._make_residual_layer(out_channels=128,
num_blocks=4,
stride=2)
self.layer256 = self._make_residual_layer(out_channels=256,
num_blocks=6,
stride=2)
self.layer512 = self._make_residual_layer(out_channels=512,
num_blocks=3,
stride=2)
self.avg_pool = AvgPool2d(kernel_size=(1, 1))
self.fc_downscale = Linear(32768, 512)
self.fc_downscale_2 = Linear(512, 128)
self.fc = Linear(128, num_classes)
# Initialize weights
for m in self.modules():
if isinstance(m, Conv2d):
nn.init.kaiming_normal_(m.weight,
mode="fan_out",
nonlinearity="relu")
elif isinstance(m, BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def __init__(self,
in_chan,
out_chan,
ks=3,
stride=1,
padding=1,
*args,
**kwargs):
super(ConvBNReLU, self).__init__()
self.conv = nn.Conv2d(
in_chan,
out_chan,
kernel_size=ks,
stride=stride,
padding=padding,
bias=False,
)
self.bn = BatchNorm2d(out_chan)
self.relu = nn.ReLU6(inplace=True)
self.init_weight()