def __init__(
self, features_out: int, depth: int, kernel_size: int = 3, normalization: bool = True, is_input: bool = False
):
super().__init__()
norm_first = not is_input
inch, ouch = self._in_out_channels(depth, features_out)
self.conv0 = ConvUnit(
in_channels=inch[0], out_channels=ouch[0], normalization=norm_first, kernel_size=kernel_size
)
self.conv1 = ConvUnit(
in_channels=inch[1], out_channels=ouch[1], normalization=normalization, kernel_size=kernel_size
)
self.pool = MaxPool3d(kernel_size=2, stride=2)
def __init__(self, d_model):
super(Convolutional_Feature_Extractor, self).__init__()
#(N,Cin,D,H,W)
self.conv3d1 = Conv3d(in_channels=3,
out_channels=32,
kernel_size=(5, 5, 5),
padding=(2, 2, 2)
# ,stride=(1,2,2)
)
self.mp3d1 = MaxPool3d(kernel_size=(1, 2, 2))
self.bn1 = BatchNorm3d(32)
self.basicblock1 = BasicBlock3D(inplanes=32, outplanes=32)
self.conv3d2 = Conv3d(in_channels=32,
out_channels=64,
kernel_size=(5, 5, 5),
padding=(2, 2, 2)
# ,stride=(1,2,2)
)
self.mp3d2 = MaxPool3d(kernel_size=(1, 2, 2),
stride=(1, 2, 2)) #(N,Cin,D,H',W')
self.bn2 = BatchNorm3d(64)
self.basicblock2 = BasicBlock3D(inplanes=64, outplanes=64)
self.conv3d3 = Conv3d(in_channels=64,
out_channels=96,
kernel_size=(1, 5, 5),
padding=(0, 2, 2)
# ,stride=(1,2,2)
)
self.mp3d3 = MaxPool3d(kernel_size=(1, 2, 2),
stride=(1, 2, 2)) # (N,Cin,D,H',W')
self.bn3 = BatchNorm3d(96)
self.basicblock3 = BasicBlock3D(inplanes=96, outplanes=96)
self.gap = AdaptiveAvgPool2d((1, 1))
#self.linear=Linear(in_features=96*72,out_features=d_model)
self.linear = Linear(in_features=96, out_features=d_model)
self.bn = BatchNorm1d(d_model)
def __init__(self, pv):
super(ResNet18, self).__init__()
self.conv1_relu = ConvolutionBlock(3, 64, pv)
self.maxpool = MaxPool3d(kernel_size=3,
padding=1,
stride=2,
dilation=1)
self.res2a_relu = ResidualBlock(64, 64, pv)
self.res2b_relu = ResidualBlock(64, 64, pv)
self.res3a_relu = ResidualBlockB(64, 128, pv)
self.res3b_relu = ResidualBlock(128, 128, pv)
self.res4a_relu = ResidualBlockB(128, 256, pv)
self.res4b_relu = ResidualBlock(256, 256, pv)
self.res5a_relu = ResidualBlockB(256, 512, pv)
self.res5b_relu = ResidualBlock(512, 512, pv)
self.avgpool = AdaptiveAvgPool3d(output_size=1)
self.fc = torch.nn.Linear(512, 27)
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride=1,
dilation=1,
padding=0,
activation=ReLU(inplace=True),
pooling=MaxPool3d(kernel_size=2),
same_padding=False):
super(ConvBlock3D, self).__init__()
self.conv = Conv3d(in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
dilation=dilation,
padding=padding)
self.activation = activation
self.pooling = pooling
self.same_padding = same_padding
def __define_timeception_layers(self, input_shape, n_layers, n_groups,
expansion_factor, is_dilated):
'''
Define layers inside the timeception layers. 定义timeception层的内部操作结构
:param input_shape: (32, 1024, 128, 7, 7)：1024输入的通道数
:param n_layers: 2
:param n_groups: 8
:param expansion_factor: 1.25
:param is_dilated: True
:return: n_channels_in :返回该层timeception的输出通道数，作为下一层timeception的输入通道数
'''
n_channels_in = input_shape[1]
# how many layers of timeception
for i in range(n_layers):
# i表示层数
layer_num = i + 1
# get details about grouping
n_channels_per_branch, n_channels_out = self.__get_n_channels_per_branch(
n_groups, expansion_factor, n_channels_in)
# temporal conv per group，在这里将timeception中用到的操作进行了定义
self.__define_grouped_convolutions(input_shape, n_groups,
n_channels_per_branch,
is_dilated, layer_num)
# downsample over time，定义在时间域上的下采样操作，即在时间域上进行最大池化操作，来满足第二个子空间平衡的原则
layer_name = 'maxpool_tc%d' % (layer_num)
layer = MaxPool3d(kernel_size=(2, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
n_channels_in = n_channels_out
input_shape[
1] = n_channels_in #下一层输入时的通道个数[1280, 1600, 2000, 2480]
return n_channels_in
def __init__(self, thresh, radius, setup_params):
super().__init__()
self.thresh = thresh
self.r = radius
self.device = setup_params['device']
self.psize_xy = setup_params['pixel_size_rec']
self.psize_z = setup_params['pixel_size_axial']
self.zmin = setup_params['zmin']
self.upsampling_shift = 0 # 2 due to floor(W/2) affected by upsampling factor of 4
self.maxpool = MaxPool3d(kernel_size=2 * self.r + 1,
stride=1,
padding=self.r)
self.pad = ConstantPad3d(self.r, 0.0)
self.zero = torch.FloatTensor([0.0]).to(self.device)
# construct the local average filters
filt_vec = np.arange(-self.r, self.r + 1)
yfilter, zfilter, xfilter = np.meshgrid(filt_vec, filt_vec, filt_vec)
xfilter = torch.FloatTensor(xfilter).unsqueeze(0).unsqueeze(0)
yfilter = torch.FloatTensor(yfilter).unsqueeze(0).unsqueeze(0)
zfilter = torch.FloatTensor(zfilter).unsqueeze(0).unsqueeze(0)
sfilter = torch.ones_like(xfilter)
self.local_filter = torch.cat((sfilter, xfilter, yfilter, zfilter),
0).to(self.device)
def build_model(self):
num_featmaps_lay1 = self.num_featmaps_base
self.convolution_downlay1_1 = Conv3d(self.num_channels_in,
num_featmaps_lay1,
kernel_size=3,
padding=1)
self.convolution_downlay1_2 = Conv3d(num_featmaps_lay1,
num_featmaps_lay1,
kernel_size=3,
padding=1)
self.pooling_downlay1 = MaxPool3d(kernel_size=2, padding=0)
num_featmaps_lay2 = 2 * num_featmaps_lay1
self.convolution_downlay2_1 = Conv3d(num_featmaps_lay1,
num_featmaps_lay2,
kernel_size=3,
padding=1)
self.convolution_downlay2_2 = Conv3d(num_featmaps_lay2,
num_featmaps_lay2,
kernel_size=3,
padding=1)
self.pooling_downlay2 = MaxPool3d(kernel_size=2, padding=0)
num_featmaps_lay3 = 2 * num_featmaps_lay2
self.convolution_downlay3_1 = Conv3d(num_featmaps_lay2,
num_featmaps_lay3,
kernel_size=3,
padding=1)
self.convolution_downlay3_2 = Conv3d(num_featmaps_lay3,
num_featmaps_lay3,
kernel_size=3,
padding=1)
self.pooling_downlay3 = MaxPool3d(kernel_size=2, padding=0)
num_featmaps_lay4 = 2 * num_featmaps_lay3
self.convolution_downlay4_1 = Conv3d(num_featmaps_lay3,
num_featmaps_lay4,
kernel_size=3,
padding=1)
self.convolution_downlay4_2 = Conv3d(num_featmaps_lay4,
num_featmaps_lay4,
kernel_size=3,
padding=1)
self.pooling_downlay4 = MaxPool3d(kernel_size=2, padding=0)
num_featmaps_lay5 = 2 * num_featmaps_lay4
self.convolution_downlay5_1 = Conv3d(num_featmaps_lay4,
num_featmaps_lay5,
kernel_size=3,
padding=1)
self.convolution_downlay5_2 = Conv3d(num_featmaps_lay5,
num_featmaps_lay5,
kernel_size=3,
padding=1)
self.upsample_uplay5 = Upsample(scale_factor=2, mode='nearest')
num_featmaps_lay4pl5 = num_featmaps_lay4 + num_featmaps_lay5
self.convolution_uplay4_1 = Conv3d(num_featmaps_lay4pl5,
num_featmaps_lay4,
kernel_size=3,
padding=1)
self.convolution_uplay4_2 = Conv3d(num_featmaps_lay4,
num_featmaps_lay4,
kernel_size=3,
padding=1)
self.upsample_uplay4 = Upsample(scale_factor=2, mode='nearest')
num_featmaps_lay3pl4 = num_featmaps_lay3 + num_featmaps_lay4
self.convolution_uplay3_1 = Conv3d(num_featmaps_lay3pl4,
num_featmaps_lay3,
kernel_size=3,
padding=1)
self.convolution_uplay3_2 = Conv3d(num_featmaps_lay3,
num_featmaps_lay3,
kernel_size=3,
padding=1)
self.upsample_uplay3 = Upsample(scale_factor=2, mode='nearest')
num_featmaps_lay2pl3 = num_featmaps_lay2 + num_featmaps_lay3
self.convolution_uplay2_1 = Conv3d(num_featmaps_lay2pl3,
num_featmaps_lay2,
kernel_size=3,
padding=1)
self.convolution_uplay2_2 = Conv3d(num_featmaps_lay2,
num_featmaps_lay2,
kernel_size=3,
padding=1)
self.upsample_uplay2 = Upsample(scale_factor=2, mode='nearest')
num_featmaps_lay1pl2 = num_featmaps_lay1 + num_featmaps_lay2
self.convolution_uplay1_1 = Conv3d(num_featmaps_lay1pl2,
num_featmaps_lay1,
kernel_size=3,
padding=1)
self.convolution_uplay1_2 = Conv3d(num_featmaps_lay1,
num_featmaps_lay1,
kernel_size=3,
padding=1)
self.classification_layer = Conv3d(num_featmaps_lay1,
self.num_classes_out,
kernel_size=1,
padding=0)
self.activation_layer = Sigmoid()
def build_model(self):
num_featmaps_lay1 = self.num_featmaps_base
self.convolution_downlay1_1 = Conv3d(self.num_channels_in,
num_featmaps_lay1,
kernel_size=3,
padding=1)
self.activation_downlay1_1 = ReLU(inplace=True)
# self.batchnorm_downlay1_1 = BatchNorm3d(num_featmaps_lay1)
self.convolution_downlay1_2 = Conv3d(num_featmaps_lay1,
num_featmaps_lay1,
kernel_size=3,
padding=1)
self.activation_downlay1_2 = ReLU(inplace=True)
# self.batchnorm_downlay1_2 = BatchNorm3d(num_featmaps_lay1)
# self.dropout_downlay1 = Dropout3d(p=self.dropout_rate)
self.pooling_downlay1 = MaxPool3d(kernel_size=2, padding=0)
num_featmaps_lay2 = 2 * num_featmaps_lay1
self.convolution_downlay2_1 = Conv3d(num_featmaps_lay1,
num_featmaps_lay2,
kernel_size=3,
padding=1)
self.activation_downlay2_1 = ReLU(inplace=True)
# self.batchnorm_downlay2_1 = BatchNorm3d(num_featmaps_lay2)
self.convolution_downlay2_2 = Conv3d(num_featmaps_lay2,
num_featmaps_lay2,
kernel_size=3,
padding=1)
self.activation_downlay2_2 = ReLU(inplace=True)
# self.batchnorm_downlay2_2 = BatchNorm3d(num_featmaps_lay2)
# self.dropout_downlay2 = Dropout3d(p =self.dropout_rate)
self.pooling_downlay2 = MaxPool3d(kernel_size=2, padding=0)
num_featmaps_lay3 = 2 * num_featmaps_lay2
self.convolution_downlay3_1 = Conv3d(num_featmaps_lay2,
num_featmaps_lay3,
kernel_size=3,
padding=1)
self.activation_downlay3_1 = ReLU(inplace=True)
# self.batchnorm_downlay3_1 = BatchNorm3d(num_featmaps_lay3)
self.convolution_downlay3_2 = Conv3d(num_featmaps_lay3,
num_featmaps_lay3,
kernel_size=3,
padding=1)
self.activation_downlay3_2 = ReLU(inplace=True)
# self.batchnorm_downlay3_2 = BatchNorm3d(num_featmaps_lay3)
# self.dropout_downlay3 = Dropout3d(p=self.dropout_rate)
self.pooling_downlay3 = MaxPool3d(kernel_size=2, padding=0)
num_featmaps_lay4 = 2 * num_featmaps_lay3
self.convolution_downlay4_1 = Conv3d(num_featmaps_lay3,
num_featmaps_lay4,
kernel_size=3,
padding=1)
self.activation_downlay4_1 = ReLU(inplace=True)
# self.batchnorm_downlay4_1 = BatchNorm3d(num_featmaps_lay4)
self.convolution_downlay4_2 = Conv3d(num_featmaps_lay4,
num_featmaps_lay4,
kernel_size=3,
padding=1)
self.activation_downlay4_2 = ReLU(inplace=True)
# self.batchnorm_downlay4_2 = BatchNorm3d(num_featmaps_lay4)
# self.dropout_downlay4 = Dropout3d(p=self.dropout_rate)
self.pooling_downlay4 = MaxPool3d(kernel_size=(1, 2, 2), padding=0)
num_featmaps_lay5 = 2 * num_featmaps_lay4
self.convolution_downlay5_1 = Conv3d(num_featmaps_lay4,
num_featmaps_lay5,
kernel_size=3,
padding=1)
self.activation_downlay5_1 = ReLU(inplace=True)
# self.batchnorm_downlay5_1 = BatchNorm3d(num_featmaps_lay5)
self.convolution_downlay5_2 = Conv3d(num_featmaps_lay5,
num_featmaps_lay5,
kernel_size=3,
padding=1)
self.activation_downlay5_2 = ReLU(inplace=True)
# self.batchnorm_downlay5_2 = BatchNorm3d(num_featmaps_lay5)
# self.dropout_downlay5 = Dropout3d(p=self.dropout_rate)
self.upsample_downlay5 = Upsample(scale_factor=(1, 2, 2),
mode='nearest')
num_featmaps_lay4pl5 = num_featmaps_lay4 + num_featmaps_lay5
self.convolution_uplay4_1 = Conv3d(num_featmaps_lay4pl5,
num_featmaps_lay4,
kernel_size=3,
padding=1)
self.activation_uplay4_1 = ReLU(inplace=True)
# self.batchnorm_uplay4_1 = BatchNorm3d(num_featmaps_lay4)
self.convolution_uplay4_2 = Conv3d(num_featmaps_lay4,
num_featmaps_lay4,
kernel_size=3,
padding=1)
self.activation_uplay4_2 = ReLU(inplace=True)
# self.batchnorm_uplay4_2 = BatchNorm3d(num_featmaps_lay4)
# self.dropout_uplay4 = Dropout3d(p=self.dropout_rate)
self.upsample_uplay4 = Upsample(scale_factor=2, mode='nearest')
num_featmaps_lay3pl4 = num_featmaps_lay3 + num_featmaps_lay4
self.convolution_uplay3_1 = Conv3d(num_featmaps_lay3pl4,
num_featmaps_lay3,
kernel_size=3,
padding=1)
self.activation_uplay3_1 = ReLU(inplace=True)
# self.batchnorm_uplay3_1 = BatchNorm3d(num_featmaps_lay3)
self.convolution_uplay3_2 = Conv3d(num_featmaps_lay3,
num_featmaps_lay3,
kernel_size=3,
padding=1)
self.activation_uplay3_2 = ReLU(inplace=True)
# self.batchnorm_uplay3_2 = BatchNorm3d(num_featmaps_lay3)
# self.dropout_uplay3 = Dropout3d(p=self.dropout_rate)
self.upsample_uplay3 = Upsample(scale_factor=2, mode='nearest')
num_featmaps_lay2pl3 = num_featmaps_lay2 + num_featmaps_lay3
self.convolution_uplay2_1 = Conv3d(num_featmaps_lay2pl3,
num_featmaps_lay2,
kernel_size=3,
padding=1)
self.activation_uplay2_1 = ReLU(inplace=True)
# self.batchnorm_uplay2_1 = BatchNorm3d(num_featmaps_lay2)
self.convolution_uplay2_2 = Conv3d(num_featmaps_lay2,
num_featmaps_lay2,
kernel_size=3,
padding=1)
self.activation_uplay2_2 = ReLU(inplace=True)
# self.batchnorm_uplay2_2 = BatchNorm3d(num_featmaps_lay2)
# self.dropout_uplay2 = Dropout3d(p=self.dropout_rate)
self.upsample_uplay2 = Upsample(scale_factor=2, mode='nearest')
num_featmaps_lay1pl2 = num_featmaps_lay1 + num_featmaps_lay2
self.convolution_uplay1_1 = Conv3d(num_featmaps_lay1pl2,
num_featmaps_lay1,
kernel_size=3,
padding=1)
self.activation_uplay1_1 = ReLU(inplace=True)
# self.batchnorm_uplay1_1 = BatchNorm3d(num_featmaps_lay1)
self.convolution_uplay1_2 = Conv3d(num_featmaps_lay1,
num_featmaps_lay1,
kernel_size=3,
padding=1)
self.activation_uplay1_2 = ReLU(inplace=True)
# self.batchnorm_uplay1_2 = BatchNorm3d(num_featmaps_lay1)
# self.dropout_uplay1 = Dropout3d(p=self.dropout_rate)
self.classification_layer = Conv3d(num_featmaps_lay1,
self.num_classes_out,
kernel_size=1,
padding=0)
self.activation_output = Sigmoid()
def __init__(self, feat_channels=[32, 64, 128, 256, 512], residual='conv'):
# residual: conv for residual input x through 1*1 conv across every layer for downsampling, None for removal of residuals
super(UNet_David, self).__init__()
class Conv3D_Block(Module):
def __init__(self, inp_feat, out_feat, kernel=3, stride=1, padding=1, residual=None):
super(Conv3D_Block, self).__init__()
self.conv1 = Sequential(
Conv3d(inp_feat, out_feat, kernel_size=kernel,
stride=stride, padding=padding, bias=True),
BatchNorm3d(out_feat),
ReLU())
self.conv2 = Sequential(
Conv3d(out_feat, out_feat, kernel_size=kernel,
stride=stride, padding=padding, bias=True),
BatchNorm3d(out_feat),
ReLU())
self.residual = residual
if self.residual is not None:
self.residual_upsampler = Conv3d(inp_feat, out_feat, kernel_size=1, bias=False)
def forward(self, x):
res = x
if not self.residual:
return self.conv2(self.conv1(x))
else:
return self.conv2(self.conv1(x)) + self.residual_upsampler(res)
class Deconv3D_Block(Module):
def __init__(self, inp_feat, out_feat, kernel=3, stride=2, padding=1):
super(Deconv3D_Block, self).__init__()
self.deconv = Sequential(
ConvTranspose3d(inp_feat, out_feat, kernel_size=(kernel, kernel, kernel),
stride=(stride, stride, stride), padding=(padding, padding, padding),
output_padding=1, bias=True),
ReLU())
def forward(self, x):
return self.deconv(x)
class ChannelPool3d(AvgPool1d):
def __init__(self, kernel_size, stride, padding):
super(ChannelPool3d, self).__init__(kernel_size, stride, padding)
self.pool_1d = AvgPool1d(self.kernel_size, self.stride, self.padding, self.ceil_mode)
def forward(self, inp):
n, c, d, w, h = inp.size()
inp = inp.view(n, c, d * w * h).permute(0, 2, 1)
pooled = self.pool_1d(inp)
c = int(c / self.kernel_size[0])
return inp.view(n, c, d, w, h)
# Encoder downsamplers
self.pool1 = MaxPool3d((2, 2, 2))
self.pool2 = MaxPool3d((2, 2, 2))
self.pool3 = MaxPool3d((2, 2, 2))
self.pool4 = MaxPool3d((2, 2, 2))
# Encoder convolutions
self.conv_blk1 = Conv3D_Block(opt.in_channels, feat_channels[0], residual=residual)
self.conv_blk2 = Conv3D_Block(feat_channels[0], feat_channels[1], residual=residual)
self.conv_blk3 = Conv3D_Block(feat_channels[1], feat_channels[2], residual=residual)
self.conv_blk4 = Conv3D_Block(feat_channels[2], feat_channels[3], residual=residual)
self.conv_blk5 = Conv3D_Block(feat_channels[3], feat_channels[4], residual=residual)
# Decoder convolutions
self.dec_conv_blk4 = Conv3D_Block(2 * feat_channels[3], feat_channels[3], residual=residual)
self.dec_conv_blk3 = Conv3D_Block(2 * feat_channels[2], feat_channels[2], residual=residual)
self.dec_conv_blk2 = Conv3D_Block(2 * feat_channels[1], feat_channels[1], residual=residual)
self.dec_conv_blk1 = Conv3D_Block(2 * feat_channels[0], feat_channels[0], residual=residual)
# Decoder upsamplers
self.deconv_blk4 = Deconv3D_Block(feat_channels[4], feat_channels[3])
self.deconv_blk3 = Deconv3D_Block(feat_channels[3], feat_channels[2])
self.deconv_blk2 = Deconv3D_Block(feat_channels[2], feat_channels[1])
self.deconv_blk1 = Deconv3D_Block(feat_channels[1], feat_channels[0])
# Final 1*1 Conv Segmentation map
self.one_conv = Conv3d(feat_channels[0], opt.out_channels, kernel_size=1, stride=1, padding=0, bias=True)
def __init__(self, in_channels, out_channels=1):
super(SmallUnet, self).__init__()
############### ENCODER ##############################
#### CONV 1 ####
self.conv1 = ConvBlock3D(in_channels=in_channels,
out_channels=16,
kernel_size=3,
stride=1,
dilation=2,
pooling=None,
activation=ReLU(inplace=True),
same_padding=True)
self.pool1 = MaxPool3d(kernel_size=2)
#### CONV 2 ####
self.conv2 = ConvBlock3D(in_channels=16,
out_channels=32,
kernel_size=3,
stride=1,
dilation=2,
pooling=None,
activation=ReLU(inplace=True),
same_padding=True)
self.pool2 = MaxPool3d(kernel_size=2)
############### DECODER ############################
#### UPCONV 1 ####
self.conv3 = ConvBlock3D(in_channels=32,
out_channels=32,
kernel_size=3,
stride=1,
dilation=2,
pooling=None,
activation=ReLU(inplace=True),
same_padding=True)
self.convt1 = ConvTransposeBlock3D(in_channels=32,
out_channels=16,
kernel_size=2,
stride=2,
activation=ReLU(inplace=True))
#### UPCONV 2 ####
self.conv4 = ConvBlock3D(in_channels=16 + 32,
out_channels=16,
kernel_size=3,
stride=1,
dilation=1,
activation=ReLU(inplace=True),
pooling=None,
same_padding=True)
self.convt2 = ConvTransposeBlock3D(in_channels=16,
out_channels=4,
kernel_size=2,
stride=2,
activation=ReLU(inplace=True))
#### FINAL LAYER ####
#self.conv5 = ConvBlock3D(in_channels=4 + 16, out_channels=out_channels, same_padding=True, kernel_size=3,
# pooling=None, activation=Softmax() if out_channels > 2 else Sigmoid())
self.conv5 = ConvBlock3D(in_channels=4 + 16,
out_channels=out_channels,
same_padding=True,
kernel_size=3,
pooling=None,
activation=None)
from torch import tensor, float32
from torch.nn import MaxPool3d, MaxUnpool3d
"""
二维的最大值池化逆转操作
*注意*：该操作是有损操作
"""
pool = MaxPool3d(kernel_size=2, stride=1, return_indices=True)
input = tensor(
[[[[[1, 1, 2], [1, 2, 3], [2, 3, 4]], [[1, 2, 3], [2, 3, 4], [3, 4, 5]],
[[2, 3, 4], [3, 4, 5], [4, 5, 6]]]]],
dtype=float32)
output, indices = pool(input)
print(output)
# tensor([[[[[3., 4.],
# [4., 5.]],
#
# [[4., 5.],
# [5., 6.]]]]])
print(indices)
# tensor([[[[[13, 14],
# [16, 17]],
#
# [[22, 23],
# [25, 26]]]]])
unpool = MaxUnpool3d(kernel_size=2, stride=1)
unpool_out = unpool(output, indices)
print(unpool_out)
# tensor([[[[[0., 0., 0.],
# [0., 0., 0.],
# [0., 0., 0.]],
#
from torch import tensor, float32
from torch.nn import MaxPool3d
"""
@:param kernel_size (int) – 池化器窗口大小
@:param stride (int) – 池化器滑动步长. Default value is kernel_size
@:param padding (int) – 边缘0填充,default 0
@:param dilation (int)– 控制窗口内元素的跨度
@:param return_indices (bool) – if True, 则返回结果是最大值的索引. Useful for torch.nn.MaxUnpool1d later
@:param ceil_mode (bool) – when True, 向上取整
"""
pool = MaxPool3d(kernel_size=2, stride=1)
input = tensor(
[[[[[0, 1, 2], [1, 2, 3], [2, 3, 4]], [[1, 2, 3], [2, 3, 4], [3, 4, 5]],
[[2, 3, 4], [3, 4, 5], [4, 5, 6]]]]],
dtype=float32)
output = pool(input)
print(output)
# tensor([[[[[3., 4.],
# [4., 5.]],
#
# [[4., 5.],
# [5., 6.]]]]])
def __define_temporal_convolutional_block(self, input_shape,
n_channels_per_branch_out,
kernel_sizes, dilation_rates,
layer_num, group_num):
"""
Define 5 branches of convolutions that operate of channels of each group.
"""
n_channels_in = input_shape[1] # 每个group的输入维度
dw_input_shape = list(input_shape)
dw_input_shape[1] = n_channels_per_branch_out # 每个branch的输出维度 = 输入维度
# 下面的五个分支对应着论文中某个group的从右到左的五个分支
# branch 1: dimension reduction only and no temporal conv
layer_name = 'conv_b1_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name,
layer) #setattr(object, name, value),setattr是python的方法
layer_name = 'bn_b1_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
# branch 2: dimension reduction followed by depth-wise temp conv (kernel-size 3)
layer_name = 'conv_b2_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'convdw_b2_g%d_tc%d' % (group_num, layer_num)
layer = DepthwiseConv1DLayer(dw_input_shape, kernel_sizes[0],
dilation_rates[0], layer_name)
setattr(self, layer_name, layer)
layer_name = 'bn_b2_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
# branch 3: dimension reduction followed by depth-wise temp conv (kernel-size 5)
layer_name = 'conv_b3_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'convdw_b3_g%d_tc%d' % (group_num, layer_num)
layer = DepthwiseConv1DLayer(dw_input_shape, kernel_sizes[1],
dilation_rates[1], layer_name)
setattr(self, layer_name, layer)
layer_name = 'bn_b3_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
# branch 4: dimension reduction followed by depth-wise temp conv (kernel-size 7)
layer_name = 'conv_b4_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'convdw_b4_g%d_tc%d' % (group_num, layer_num)
layer = DepthwiseConv1DLayer(dw_input_shape, kernel_sizes[2],
dilation_rates[2], layer_name)
setattr(self, layer_name, layer)
layer_name = 'bn_b4_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
# branch 5: dimension reduction followed by temporal max pooling
layer_name = 'conv_b5_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'maxpool_b5_g%d_tc%d' % (group_num, layer_num)
layer = MaxPool3d(kernel_size=(2, 1, 1), stride=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'padding_b5_g%d_tc%d' % (group_num, layer_num)
layer = torch.nn.ReplicationPad3d(
(0, 0, 0, 0, 1, 0)) # left, right, top, bottom, front, back
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'bn_b5_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
def Maxpool3D_Block():
pool = MaxPool3d(kernel_size=2, stride=2, padding=0)
return pool
def __init__(self):
super(ResNet18Explicit3DConvReduced, self).__init__()
# self.conv1_relu = ConvolutionBlock(3, 64, pv)
self.conv1 = Conv3d(in_channels=3,
out_channels=int(64 / 1.718),
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = BatchNorm3d(int(64 / 1.718))
self.maxpool = MaxPool3d(kernel_size=3,
padding=1,
stride=2,
dilation=1)
# self.res2a_relu = ResidualBlock(int(64/1.718), int(64/1.718), pv)
self.conv2 = Conv3d(in_channels=int(64 / 1.718),
out_channels=int(64 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn2 = BatchNorm3d(int(64 / 1.718))
self.conv3 = Conv3d(in_channels=int(64 / 1.718),
out_channels=int(64 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn3 = BatchNorm3d(int(64 / 1.718))
# self.res2b_relu = ResidualBlock(int(64/1.718), int(64/1.718), pv)
self.conv4 = Conv3d(in_channels=int(64 / 1.718),
out_channels=int(64 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn4 = BatchNorm3d(int(64 / 1.718))
self.conv5 = Conv3d(in_channels=int(64 / 1.718),
out_channels=int(64 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn5 = BatchNorm3d(int(64 / 1.718))
# self.res3a_relu = ResidualBlockB(int(64/1.718), int(128/1.718), pv)
self.conv6 = Conv3d(in_channels=int(64 / 1.718),
out_channels=int(128 / 1.718),
kernel_size=1,
stride=2,
bias=False)
self.bn6 = BatchNorm3d(int(128 / 1.718))
self.conv7 = Conv3d(in_channels=int(64 / 1.718),
out_channels=int(128 / 1.718),
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.bn7 = BatchNorm3d(int(128 / 1.718))
self.conv8 = Conv3d(in_channels=int(128 / 1.718),
out_channels=int(128 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn8 = BatchNorm3d(int(128 / 1.718))
# self.res3b_relu = ResidualBlock(int(128/1.718), int(128/1.718), pv)
self.conv9 = Conv3d(in_channels=int(128 / 1.718),
out_channels=int(128 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn9 = BatchNorm3d(int(128 / 1.718))
self.conv10 = Conv3d(in_channels=int(128 / 1.718),
out_channels=int(128 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn10 = BatchNorm3d(int(128 / 1.718))
# self.res4a_relu = ResidualBlockB(int(128/1.718), int(256/1.718), pv)
self.conv11 = Conv3d(in_channels=int(128 / 1.718),
out_channels=int(256 / 1.718),
kernel_size=1,
stride=2,
bias=False)
self.bn11 = BatchNorm3d(int(256 / 1.718))
self.conv12 = Conv3d(in_channels=int(128 / 1.718),
out_channels=int(256 / 1.718),
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.bn12 = BatchNorm3d(int(256 / 1.718))
self.conv13 = Conv3d(in_channels=int(256 / 1.718),
out_channels=int(256 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn13 = BatchNorm3d(int(256 / 1.718))
# self.res4b_relu = ResidualBlock(int(256/1.718), int(256/1.718), pv)
self.conv14 = Conv3d(in_channels=int(256 / 1.718),
out_channels=int(256 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn14 = BatchNorm3d(int(256 / 1.718))
self.conv15 = Conv3d(in_channels=int(256 / 1.718),
out_channels=int(256 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn15 = BatchNorm3d(int(256 / 1.718))
# self.res5a_relu = ResidualBlockB(int(256/1.718), int(512/1.718), pv)
self.conv16 = Conv3d(in_channels=int(256 / 1.718),
out_channels=int(512 / 1.718),
kernel_size=1,
stride=2,
bias=False)
self.bn16 = BatchNorm3d(int(512 / 1.718))
self.conv17 = Conv3d(in_channels=int(256 / 1.718),
out_channels=int(512 / 1.718),
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.bn17 = BatchNorm3d(int(512 / 1.718))
self.conv18 = Conv3d(in_channels=int(512 / 1.718),
out_channels=int(512 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn18 = BatchNorm3d(int(512 / 1.718))
# self.res5b_relu = ResidualBlock(int(512/1.718), int(512/1.718), pv)
self.conv19 = Conv3d(in_channels=int(512 / 1.718),
out_channels=int(512 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn19 = BatchNorm3d(int(512 / 1.718))
self.conv20 = Conv3d(in_channels=int(512 / 1.718),
out_channels=int(512 / 1.718),
kernel_size=3,
padding=1,
bias=False)
self.bn20 = BatchNorm3d(int(512 / 1.718))
self.avgpool = AdaptiveAvgPool3d(output_size=1)
self.fc = torch.nn.Linear(int(512 / 1.718), 27)
def __init__(self, pv):
super(ResNet18Explicit3DConv, self).__init__()
# self.conv1_relu = ConvolutionBlock(3, 64, pv)
self.conv1 = Conv3d(in_channels=3,
out_channels=64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = BatchNorm3d(64)
self.maxpool = MaxPool3d(kernel_size=3,
padding=1,
stride=2,
dilation=1)
# self.res2a_relu = ResidualBlock(64, 64, pv)
self.conv2 = Conv3d(in_channels=64,
out_channels=64,
kernel_size=3,
padding=1,
bias=False)
self.bn2 = BatchNorm3d(64)
self.conv3 = Conv3d(in_channels=64,
out_channels=64,
kernel_size=3,
padding=1,
bias=False)
self.bn3 = BatchNorm3d(64)
# self.res2b_relu = ResidualBlock(64, 64, pv)
self.conv4 = Conv3d(in_channels=64,
out_channels=64,
kernel_size=3,
padding=1,
bias=False)
self.bn4 = BatchNorm3d(64)
self.conv5 = Conv3d(in_channels=64,
out_channels=64,
kernel_size=3,
padding=1,
bias=False)
self.bn5 = BatchNorm3d(64)
# self.res3a_relu = ResidualBlockB(64, 128, pv)
self.conv6 = Conv3d(in_channels=64,
out_channels=128,
kernel_size=1,
stride=2,
bias=False)
self.bn6 = BatchNorm3d(128)
self.conv7 = Conv3d(in_channels=64,
out_channels=128,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.bn7 = BatchNorm3d(128)
self.conv8 = Conv3d(in_channels=128,
out_channels=128,
kernel_size=3,
padding=1,
bias=False)
self.bn8 = BatchNorm3d(128)
# self.res3b_relu = ResidualBlock(128, 128, pv)
self.conv9 = Conv3d(in_channels=128,
out_channels=128,
kernel_size=3,
padding=1,
bias=False)
self.bn9 = BatchNorm3d(128)
self.conv10 = Conv3d(in_channels=128,
out_channels=128,
kernel_size=3,
padding=1,
bias=False)
self.bn10 = BatchNorm3d(128)
# self.res4a_relu = ResidualBlockB(128, 256, pv)
self.conv11 = Conv3d(in_channels=128,
out_channels=256,
kernel_size=1,
stride=2,
bias=False)
self.bn11 = BatchNorm3d(256)
self.conv12 = Conv3d(in_channels=128,
out_channels=256,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.bn12 = BatchNorm3d(256)
self.conv13 = Conv3d(in_channels=256,
out_channels=256,
kernel_size=3,
padding=1,
bias=False)
self.bn13 = BatchNorm3d(256)
# self.res4b_relu = ResidualBlock(256, 256, pv)
self.conv14 = Conv3d(in_channels=256,
out_channels=256,
kernel_size=3,
padding=1,
bias=False)
self.bn14 = BatchNorm3d(256)
self.conv15 = Conv3d(in_channels=256,
out_channels=256,
kernel_size=3,
padding=1,
bias=False)
self.bn15 = BatchNorm3d(256)
# self.res5a_relu = ResidualBlockB(256, 512, pv)
self.conv16 = Conv3d(in_channels=256,
out_channels=512,
kernel_size=1,
stride=2,
bias=False)
self.bn16 = BatchNorm3d(512)
self.conv17 = Conv3d(in_channels=256,
out_channels=512,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.bn17 = BatchNorm3d(512)
self.conv18 = Conv3d(in_channels=512,
out_channels=512,
kernel_size=3,
padding=1,
bias=False)
self.bn18 = BatchNorm3d(512)
# self.res5b_relu = ResidualBlock(512, 512, pv)
self.conv19 = Conv3d(in_channels=512,
out_channels=512,
kernel_size=3,
padding=1,
bias=False)
self.bn19 = BatchNorm3d(512)
self.conv20 = Conv3d(in_channels=512,
out_channels=512,
kernel_size=3,
padding=1,
bias=False)
self.bn20 = BatchNorm3d(512)
self.avgpool = AdaptiveAvgPool3d(output_size=1)
self.fc = torch.nn.Linear(512, pv.label_size)
def __define_temporal_convolutional_block(self, input_shape,
n_channels_per_branch_out,
kernel_sizes, dilation_rates,
layer_num, group_num):
'''
Define 5 branches of convolutions that operate of channels of each group.
定义每组通道的具体处理操作
:param input_shape: (32, 128, 128, 7, 7)：1024输入的通道数
:param n_channels_per_branch_out: [32,40,50,62]
:param kernel_sizes: dilated-[3,3,3],no dilated-[3,5,7]
:param dilation_rates: dilated-[1,2,3],no dilated-[1,1,1]
:param layer_num: 当前的层数
:param group_num: 当前的group数
:return: 定义temporal卷积的block
'''
n_channels_in = input_shape[1]
dw_input_shape = list(input_shape)
dw_input_shape[1] = n_channels_per_branch_out #[32,40,50,62]
# setattr()函数对应函数getattr()，用于设置属性值，该属性不一定是存在的
'''
示例：
layer_name = 'conv_b1_g%d_tc%d' % (1, 1)
layer_name
Out[13]: 'conv_b1_g1_tc1'，表示第1个timeception层的第1个group的第1个branch
'''
# branch 1: dimension reduction only and no temporal conv (kernel-size 1)
layer_name = 'conv_b1_g%d_tc%d' % (
group_num, layer_num) #不同timeception层的不同group在同一个branch上采用的是同一个操作
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'bn_b1_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
# branch 2: dimension reduction followed by depth-wise temp conv (kernel-size 3)
# 缩减通道
layer_name = 'conv_b2_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
# layer = Conv3d(n_channels_in = 128, n_channels_per_branch_out = 32, kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
#卷积
layer_name = 'convdw_b2_g%d_tc%d' % (group_num, layer_num)
layer = DepthwiseConv1DLayer(dw_input_shape, kernel_sizes[0],
dilation_rates[0], layer_name)
# layer = DepthwiseConv1DLayer(dw_input_shape=(32, 32, 128, 7, 7) , kernel_sizes[0], dilation_rates[0], layer_name)
setattr(self, layer_name, layer)
#BN操作
layer_name = 'bn_b2_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
# branch 3: dimension reduction followed by depth-wise temp conv (kernel-size 5)
layer_name = 'conv_b3_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'convdw_b3_g%d_tc%d' % (group_num, layer_num)
layer = DepthwiseConv1DLayer(dw_input_shape, kernel_sizes[1],
dilation_rates[1], layer_name)
setattr(self, layer_name, layer)
layer_name = 'bn_b3_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
# branch 4: dimension reduction followed by depth-wise temp conv (kernel-size 7)
layer_name = 'conv_b4_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'convdw_b4_g%d_tc%d' % (group_num, layer_num)
layer = DepthwiseConv1DLayer(dw_input_shape, kernel_sizes[2],
dilation_rates[2], layer_name)
setattr(self, layer_name, layer)
layer_name = 'bn_b4_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
# branch 5: dimension reduction followed by temporal max pooling
layer_name = 'conv_b5_g%d_tc%d' % (group_num, layer_num)
layer = Conv3d(n_channels_in,
n_channels_per_branch_out,
kernel_size=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'maxpool_b5_g%d_tc%d' % (group_num, layer_num)
layer = MaxPool3d(kernel_size=(2, 1, 1), stride=(1, 1, 1))
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'padding_b5_g%d_tc%d' % (group_num, layer_num)
layer = torch.nn.ReplicationPad3d(
(0, 0, 0, 0, 1, 0)) # left, right, top, bottom, front, back
layer._name = layer_name
setattr(self, layer_name, layer)
layer_name = 'bn_b5_g%d_tc%d' % (group_num, layer_num)
layer = BatchNorm3d(n_channels_per_branch_out)
layer._name = layer_name
setattr(self, layer_name, layer)
def __init__(self, pv):
super(Googlenet3TConv_explicit, self).__init__()
self.conv1 = ConvTTN3d(in_channels=3, out_channels=64, kernel_size=7, padding=3, stride=2, project_variable=pv, bias=False)
self.bn1 = BatchNorm3d(64)
self.maxpool1 = MaxPool3d(kernel_size=(1, 3, 3), padding=0, stride=(1, 2, 2))
self.conv2 = Conv3d(in_channels=64, out_channels=64, kernel_size=1, padding=0, stride=1, bias=False)
self.bn2 = BatchNorm3d(64)
self.conv3 = ConvTTN3d(in_channels=64, out_channels=192, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn3 = BatchNorm3d(192)
self.maxpool2 = MaxPool3d(kernel_size=(1, 3, 3), padding=0, stride=(1, 2, 2))
# inception 3a
self.conv4 = Conv3d(in_channels=192, out_channels=64, kernel_size=1, padding=0, stride=1, bias=False)
self.bn4 = BatchNorm3d(64)
self.conv5 = Conv3d(in_channels=192, out_channels=96, kernel_size=1, padding=0, stride=1, bias=False)
self.bn5 = BatchNorm3d(96)
self.conv6 = ConvTTN3d(in_channels=96, out_channels=128, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn6 = BatchNorm3d(128)
self.conv7 = Conv3d(in_channels=192, out_channels=16, kernel_size=1, padding=0, stride=1, bias=False)
self.bn7 = BatchNorm3d(16)
self.conv8 = ConvTTN3d(in_channels=16, out_channels=32, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn8 = BatchNorm3d(32)
self.maxpool3 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv9 = Conv3d(in_channels=192, out_channels=32, kernel_size=1, padding=0, stride=1, bias=False)
self.bn9 = BatchNorm3d(32)
# inception 3b
self.conv10 = Conv3d(in_channels=256, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn10 = BatchNorm3d(128)
self.conv11 = Conv3d(in_channels=256, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn11 = BatchNorm3d(128)
self.conv12 = ConvTTN3d(in_channels=128, out_channels=192, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn12 = BatchNorm3d(192)
self.conv13 = Conv3d(in_channels=256, out_channels=32, kernel_size=1, padding=0, stride=1, bias=False)
self.bn13 = BatchNorm3d(32)
self.conv14 = ConvTTN3d(in_channels=32, out_channels=96, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn14 = BatchNorm3d(96)
self.maxpool4 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv15 = Conv3d(in_channels=256, out_channels=64, kernel_size=1, padding=0, stride=1, bias=False)
self.bn15 = BatchNorm3d(64)
self.maxpool5 = MaxPool3d(kernel_size=3, padding=0, stride=2)
# inception 4a
self.conv16 = Conv3d(in_channels=480, out_channels=192, kernel_size=1, padding=0, stride=1, bias=False)
self.bn16 = BatchNorm3d(192)
self.conv17 = Conv3d(in_channels=480, out_channels=96, kernel_size=1, padding=0, stride=1, bias=False)
self.bn17 = BatchNorm3d(96)
self.conv18 = ConvTTN3d(in_channels=96, out_channels=208, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn18 = BatchNorm3d(208)
self.conv19 = Conv3d(in_channels=480, out_channels=16, kernel_size=1, padding=0, stride=1, bias=False)
self.bn19 = BatchNorm3d(16)
self.conv20 = ConvTTN3d(in_channels=16, out_channels=48, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn20 = BatchNorm3d(48)
self.maxpool6 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv21 = Conv3d(in_channels=480, out_channels=64, kernel_size=1, padding=0, stride=1, bias=False)
self.bn21 = BatchNorm3d(64)
# inception 4b
self.conv22 = Conv3d(in_channels=512, out_channels=160, kernel_size=1, padding=0, stride=1, bias=False)
self.bn22 = BatchNorm3d(160)
self.conv23 = Conv3d(in_channels=512, out_channels=112, kernel_size=1, padding=0, stride=1, bias=False)
self.bn23 = BatchNorm3d(112)
self.conv24 = ConvTTN3d(in_channels=112, out_channels=224, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn24 = BatchNorm3d(224)
self.conv25 = Conv3d(in_channels=512, out_channels=24, kernel_size=1, padding=0, stride=1, bias=False)
self.bn25 = BatchNorm3d(24)
self.conv26 = ConvTTN3d(in_channels=24, out_channels=64, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn26 = BatchNorm3d(64)
self.maxpool7 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv27 = Conv3d(in_channels=512, out_channels=64, kernel_size=1, padding=0, stride=1, bias=False)
self.bn27 = BatchNorm3d(64)
self.avgpool1 = AvgPool3d(kernel_size=5, padding=0, stride=3)
self.conv28 = Conv3d(in_channels=512, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn28 = BatchNorm3d(128)
# self.fc1 = Linear(in_features=2304, out_features=1024)
self.fc1 = Linear(in_features=768, out_features=1024) # 768
self.dropout1 = Dropout3d(p=0.7)
self.fc2 = Linear(in_features=1024, out_features=pv.label_size)
# inception 4c
self.conv29 = Conv3d(in_channels=512, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn29 = BatchNorm3d(128)
self.conv30 = Conv3d(in_channels=512, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn30 = BatchNorm3d(128)
self.conv31 = ConvTTN3d(in_channels=128, out_channels=256, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn31 = BatchNorm3d(256)
self.conv32 = Conv3d(in_channels=512, out_channels=24, kernel_size=1, padding=0, stride=1, bias=False)
self.bn32 = BatchNorm3d(24)
self.conv33 = ConvTTN3d(in_channels=24, out_channels=64, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn33 = BatchNorm3d(64)
self.maxpool8 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv34 = Conv3d(in_channels=512, out_channels=64, kernel_size=1, padding=0, stride=1, bias=False)
self.bn34 = BatchNorm3d(64)
# inception 4d
self.conv35 = Conv3d(in_channels=512, out_channels=112, kernel_size=1, padding=0, stride=1, bias=False)
self.bn35 = BatchNorm3d(112)
self.conv36 = Conv3d(in_channels=512, out_channels=144, kernel_size=1, padding=0, stride=1, bias=False)
self.bn36 = BatchNorm3d(144)
self.conv37 = ConvTTN3d(in_channels=144, out_channels=288, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn37 = BatchNorm3d(288)
self.conv38 = Conv3d(in_channels=512, out_channels=32, kernel_size=1, padding=0, stride=1, bias=False)
self.bn38 = BatchNorm3d(32)
self.conv39 = ConvTTN3d(in_channels=32, out_channels=64, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn39 = BatchNorm3d(64)
self.maxpool9 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv40 = Conv3d(in_channels=512, out_channels=64, kernel_size=1, padding=0, stride=1, bias=False)
self.bn40 = BatchNorm3d(64)
# inception 4e
self.conv41 = Conv3d(in_channels=528, out_channels=256, kernel_size=1, padding=0, stride=1, bias=False)
self.bn41 = BatchNorm3d(256)
self.conv42 = Conv3d(in_channels=528, out_channels=160, kernel_size=1, padding=0, stride=1, bias=False)
self.bn42 = BatchNorm3d(160)
self.conv43 = ConvTTN3d(in_channels=160, out_channels=320, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn43 = BatchNorm3d(320)
self.conv44 = Conv3d(in_channels=528, out_channels=32, kernel_size=1, padding=0, stride=1, bias=False)
self.bn44 = BatchNorm3d(32)
self.conv45 = ConvTTN3d(in_channels=32, out_channels=128, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn45 = BatchNorm3d(128)
self.maxpool10 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv46 = Conv3d(in_channels=528, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn46 = BatchNorm3d(128)
self.avgpool2 = AvgPool3d(kernel_size=5, padding=0, stride=3)
self.conv47 = Conv3d(in_channels=528, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn47 = BatchNorm3d(128)
# self.fc3 = Linear(in_features=2304, out_features=1024)
self.fc3 = Linear(in_features=768, out_features=1024)
self.dropout2 = Dropout3d(p=0.7)
self.fc4 = Linear(in_features=1024, out_features=pv.label_size)
self.maxpool11 = MaxPool3d(kernel_size=3, padding=0, stride=2)
# inception 5a
self.conv48 = Conv3d(in_channels=832, out_channels=256, kernel_size=1, padding=0, stride=1, bias=False)
self.bn48 = BatchNorm3d(256)
self.conv49 = Conv3d(in_channels=832, out_channels=160, kernel_size=1, padding=0, stride=1, bias=False)
self.bn49 = BatchNorm3d(160)
self.conv50 = ConvTTN3d(in_channels=160, out_channels=320, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn50 = BatchNorm3d(320)
self.conv51 = Conv3d(in_channels=832, out_channels=32, kernel_size=1, padding=0, stride=1, bias=False)
self.bn51 = BatchNorm3d(32)
self.conv52 = ConvTTN3d(in_channels=32, out_channels=128, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn52 = BatchNorm3d(128)
self.maxpool12 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv53 = Conv3d(in_channels=832, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn53 = BatchNorm3d(128)
# inception 5b
self.conv54 = Conv3d(in_channels=832, out_channels=384, kernel_size=1, padding=0, stride=1, bias=False)
self.bn54 = BatchNorm3d(384)
self.conv55 = Conv3d(in_channels=832, out_channels=192, kernel_size=1, padding=0, stride=1, bias=False)
self.bn55 = BatchNorm3d(192)
self.conv56 = ConvTTN3d(in_channels=192, out_channels=384, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn56 = BatchNorm3d(384)
self.conv57 = Conv3d(in_channels=832, out_channels=48, kernel_size=1, padding=0, stride=1, bias=False)
self.bn57 = BatchNorm3d(48)
self.conv58 = ConvTTN3d(in_channels=48, out_channels=128, kernel_size=3, padding=1, stride=1, project_variable=pv, bias=False)
self.bn58 = BatchNorm3d(128)
self.maxpool13 = MaxPool3d(kernel_size=3, padding=1, stride=1)
self.conv59 = Conv3d(in_channels=832, out_channels=128, kernel_size=1, padding=0, stride=1, bias=False)
self.bn59 = BatchNorm3d(128)
self.avgpool3 = AdaptiveAvgPool3d(1)
self.dropout3 = Dropout3d(p=0.4)
self.fc5 = Linear(in_features=1024, out_features=pv.label_size)
def __init__(self, input_features, input_dim):
super().__init__()
# First Convolutional Block
self.conv11 = Conv3d(in_channels=input_features,
out_channels=32,
kernel_size=(1, 3, 3),
stride=(1, 1, 1),
padding=(0, 0, 0),
bias=True)
self.conv12 = Conv3d(in_channels=32,
out_channels=32,
kernel_size=(1, 3, 3),
stride=(1, 1, 1),
padding=0,
bias=True)
self.conv13 = Conv3d(in_channels=32,
out_channels=32,
kernel_size=(1, 3, 3),
stride=(1, 1, 1),
padding=0,
bias=True)
self.conv14 = Conv3d(in_channels=32,
out_channels=32,
kernel_size=(1, 3, 3),
stride=(1, 1, 1),
padding=0,
bias=True)
self.maxPool1 = MaxPool3d(kernel_size=(1, 2, 2),
stride=(1, 2, 2),
padding=0)
self.batchN1 = BatchNorm3d(num_features=32)
# Compute output dimensions
dimconv11 = compute_conv_dim(im_dim=input_dim,
padding=0,
dilation=1,
kernel_size=3,
stride=1)
dimconv12 = compute_conv_dim(im_dim=dimconv11,
padding=0,
dilation=1,
kernel_size=3,
stride=1)
dimconv13 = compute_conv_dim(im_dim=dimconv12,
padding=0,
dilation=1,
kernel_size=3,
stride=1)
dimconv14 = compute_conv_dim(im_dim=dimconv13,
padding=0,
dilation=1,
kernel_size=3,
stride=1)
self.dim_max = compute_conv_dim(im_dim=dimconv14,
padding=0,
dilation=1,
kernel_size=2,
stride=2)
# Second Convolutional Block
self.conv21 = Conv3d(in_channels=32,
out_channels=64,
kernel_size=(1, 3, 3),
stride=(1, 1, 1),
padding=(0, 0, 0),
bias=True)
self.conv22 = Conv3d(in_channels=64,
out_channels=64,
kernel_size=(1, 3, 3),
stride=(1, 1, 1),
padding=0,
bias=True)
self.maxPool2 = MaxPool3d(kernel_size=(1, 2, 2),
stride=(1, 2, 2),
padding=0)
self.batchN2 = BatchNorm3d(num_features=64)
# Cocmpute output dimensions:
dimconv21 = compute_conv_dim(im_dim=self.dim_max,
padding=0,
dilation=1,
kernel_size=3,
stride=1)
dimconv22 = compute_conv_dim(im_dim=dimconv21,
padding=0,
dilation=1,
kernel_size=3,
stride=1)
self.dim_max2 = compute_conv_dim(im_dim=dimconv22,
padding=0,
dilation=1,
kernel_size=2,
stride=2)
# Third block
self.conv31 = Conv3d(in_channels=64,
out_channels=128,
kernel_size=(1, 3, 3),
stride=(1, 1, 1),
padding=(0, 0, 0),
bias=True)
self.maxPool3 = MaxPool3d(kernel_size=(1, 2, 2),
stride=(1, 2, 2),
padding=0)
self.batchN3 = BatchNorm3d(num_features=128)
# Compute output dimensions
dimconv31 = compute_conv_dim(im_dim=self.dim_max2,
padding=0,
dilation=1,
kernel_size=3,
stride=1)
self.dim_max3 = compute_conv_dim(im_dim=dimconv31,
padding=0,
dilation=1,
kernel_size=2,
stride=2)
# LSTM
self.LSTM = LSTM(input_size=10368, hidden_size=128, num_layers=1)
self.linear = Linear(in_features=3712, out_features=24, bias=True)
self.l_out = Linear(in_features=24, out_features=3, bias=True)
def __init__(self,
num_channels=32,
output_channel=32,
feat_channels=[64, 128, 256, 512, 1024],
residual='conv'):
# residual: conv for residual input x through 1*1 conv across every layer for downsampling, None for removal of residuals
super(UNet3D, self).__init__()
# Encoder downsamplers
self.pool1 = MaxPool3d((2, 2, 2))
self.pool2 = MaxPool3d((2, 2, 2))
self.pool3 = MaxPool3d((2, 2, 2))
self.pool4 = MaxPool3d((2, 2, 2))
# Encoder convolutions
self.conv_blk1 = Conv3D_Block(num_channels,
feat_channels[0],
residual=residual)
self.conv_blk2 = Conv3D_Block(feat_channels[0],
feat_channels[1],
residual=residual)
self.conv_blk3 = Conv3D_Block(feat_channels[1],
feat_channels[2],
residual=residual)
self.conv_blk4 = Conv3D_Block(feat_channels[2],
feat_channels[3],
residual=residual)
self.conv_blk5 = Conv3D_Block(feat_channels[3],
feat_channels[4],
residual=residual)
# Decoder convolutions
self.dec_conv_blk4 = Conv3D_Block(2 * feat_channels[3],
feat_channels[3],
residual=residual)
self.dec_conv_blk3 = Conv3D_Block(2 * feat_channels[2],
feat_channels[2],
residual=residual)
self.dec_conv_blk2 = Conv3D_Block(2 * feat_channels[1],
feat_channels[1],
residual=residual)
self.dec_conv_blk1 = Conv3D_Block(2 * feat_channels[0],
feat_channels[0],
residual=residual)
# Decoder upsamplers
self.deconv_blk4 = Deconv3D_Block(feat_channels[4], feat_channels[3])
self.deconv_blk3 = Deconv3D_Block(feat_channels[3], feat_channels[2])
self.deconv_blk2 = Deconv3D_Block(feat_channels[2], feat_channels[1])
self.deconv_blk1 = Deconv3D_Block(feat_channels[1], feat_channels[0])
# Final 1*1 Conv Segmentation map
#self.one_conv = Conv3d(feat_channels[0], num_channels, kernel_size=1, stride=1, padding=0, bias=True)
self.one_conv = Conv3d(feat_channels[0],
output_channel,
kernel_size=1,
stride=1,
padding=0,
bias=True)
# Activation function
self.sigmoid = Sigmoid()
def __init__(self,
num_channels=3,
feat_channels=[32, 64, 128, 256, 256],
residual='conv',
is_dilated=False,
is_off_unit=False,
is_2d1d=False):
super(unet3d, self).__init__()
self.is_off_unit = is_off_unit
# Encoder downsamplers
self.pool1 = MaxPool3d((1, 2, 2))
self.pool2 = MaxPool3d((1, 2, 2))
self.pool3 = MaxPool3d((1, 2, 2))
self.pool4 = MaxPool3d((1, 2, 2))
# Encoder convolutions
if is_dilated:
self.conv_blk1 = Conv3D_Block_adv(num_channels,
feat_channels[0],
residual=residual,
is_2d1d=is_2d1d)
self.conv_blk2 = Conv3D_Block_adv(feat_channels[0],
feat_channels[1],
residual=residual,
is_2d1d=is_2d1d)
self.conv_blk3 = Conv3D_Block_adv(feat_channels[1],
feat_channels[2],
residual=residual,
is_2d1d=is_2d1d)
self.conv_blk4 = Conv3D_Block_adv(feat_channels[2],
feat_channels[3],
residual=residual,
is_2d1d=is_2d1d)
self.conv_blk5 = Conv3D_Block_adv(feat_channels[3],
feat_channels[4],
residual=residual,
is_2d1d=is_2d1d)
# Decoder convolutions
self.dec_conv_blk4 = Conv3D_Block_adv(2 * feat_channels[3],
feat_channels[3],
residual=residual,
is_2d1d=is_2d1d)
self.dec_conv_blk3 = Conv3D_Block_adv(2 * feat_channels[2],
feat_channels[2],
residual=residual,
is_2d1d=is_2d1d)
self.dec_conv_blk2 = Conv3D_Block_adv(2 * feat_channels[1],
feat_channels[1],
residual=residual,
is_2d1d=is_2d1d)
self.dec_conv_blk1 = Conv3D_Block_adv(2 * feat_channels[0],
feat_channels[0],
residual=residual,
is_2d1d=is_2d1d)
else:
# Encoder convolutions
self.conv_blk1 = Conv3D_Block(num_channels,
feat_channels[0],
residual=residual,
is_2d1d=is_2d1d)
self.conv_blk2 = Conv3D_Block(feat_channels[0],
feat_channels[1],
residual=residual,
is_2d1d=is_2d1d)
self.conv_blk3 = Conv3D_Block(feat_channels[1],
feat_channels[2],
residual=residual,
is_2d1d=is_2d1d)
self.conv_blk4 = Conv3D_Block(feat_channels[2],
feat_channels[3],
residual=residual,
is_2d1d=is_2d1d)
self.conv_blk5 = Conv3D_Block(feat_channels[3],
feat_channels[4],
residual=residual,
is_2d1d=is_2d1d)
# Decoder convolutions
self.dec_conv_blk4 = Conv3D_Block(2 * feat_channels[3],
feat_channels[3],
residual=residual,
is_2d1d=is_2d1d)
self.dec_conv_blk3 = Conv3D_Block(2 * feat_channels[2],
feat_channels[2],
residual=residual,
is_2d1d=is_2d1d)
self.dec_conv_blk2 = Conv3D_Block(2 * feat_channels[1],
feat_channels[1],
residual=residual,
is_2d1d=is_2d1d)
self.dec_conv_blk1 = Conv3D_Block(2 * feat_channels[0],
feat_channels[0],
residual=residual,
is_2d1d=is_2d1d)
if self.is_off_unit:
self.off_unit_enc1 = OFF_Unit(feat_channels[0], feat_channels[0])
self.off_unit_enc2 = OFF_Unit(feat_channels[1], feat_channels[1])
self.off_unit_enc3 = OFF_Unit(feat_channels[2], feat_channels[2])
self.off_unit_enc4 = OFF_Unit(feat_channels[3], feat_channels[3])
self.off_unit_enc5 = OFF_Unit(feat_channels[4], feat_channels[4])
self.off_unit_dec4 = OFF_Unit(feat_channels[3], feat_channels[3])
self.off_unit_dec3 = OFF_Unit(feat_channels[2], feat_channels[2])
self.off_unit_dec2 = OFF_Unit(feat_channels[1], feat_channels[1])
self.off_unit_dec1 = OFF_Unit(feat_channels[0], feat_channels[0])
# Decoder upsamplers
self.deconv_blk4 = Deconv3D_Block(feat_channels[4], feat_channels[3])
self.deconv_blk3 = Deconv3D_Block(feat_channels[3], feat_channels[2])
self.deconv_blk2 = Deconv3D_Block(feat_channels[2], feat_channels[1])
self.deconv_blk1 = Deconv3D_Block(feat_channels[1], feat_channels[0])
# Final 1*1 Conv Segmentation map
self.one_conv = Conv3d(feat_channels[0],
1,
kernel_size=1,
stride=1,
padding=0,
bias=True)