Spaces:
Sleeping
Sleeping
import torch as T | |
import torch.nn as nn | |
from Networks.network import Network | |
class UNet_concat_student(Network): | |
def __init__(self, base, expansion): | |
super(UNet_concat_student, self).__init__() | |
self._build(base, expansion) | |
def forward(self, image, segbox = None, eval_mode = True): | |
layer_0, layer_1, layer_2, layer_3 = self._analysis(image) | |
inter_output, s_3, s_2, s_1 = self._synthesis(layer_0, layer_1, layer_2, layer_3, self._bridge(layer_3)) | |
output_0 = self.output_0(inter_output) | |
u_0 = self._upsample(s_3, s_2, s_1, inter_output) | |
output_1 = self.output_1(u_0) | |
if (eval_mode): | |
return output_0 | |
return output_0, output_1 | |
def train_step(self, image, segment, criterion, segbox = None): | |
seg_out, box_out = self.forward(image, eval_mode=False) | |
seg_loss = criterion(seg_out, segment) | |
box_loss = criterion(box_out, segbox) | |
return seg_loss + box_loss | |
def _analysis(self, x): | |
layer_0 = self.analysis_0(x) | |
layer_1 = self.analysis_1(layer_0) | |
layer_2 = self.analysis_2(layer_1) | |
layer_3 = self.analysis_3(layer_2) | |
return layer_0, layer_1, layer_2, layer_3 | |
def _bridge(self, layer_3): | |
return self.bridge(layer_3) | |
def _synthesis(self, l0, l1, l2, l3, l4): | |
c_3 = T.cat((l3, l4), dim = 1) | |
s_3 = self.synthesis_3(c_3) | |
c_2 = T.cat((l2, self.up_conv_3(s_3)), dim = 1) | |
s_2 = self.synthesis_2(c_2) | |
c_1 = T.cat((l1, self.up_conv_2(s_2)), dim = 1) | |
s_1 = self.synthesis_1(c_1) | |
c_0 = T.cat((l0, self.up_conv_1(s_1)), dim = 1) | |
return self.synthesis_0(c_0), s_3, s_2, s_1 | |
def _upsample(self, s_3, s_2, s_1, inter): | |
u_3 = self.up_sample_3(s_3) | |
u_2 = self.up_sample_2(s_2) | |
u_1 = self.up_sample_1(s_1) | |
return T.cat((u_3, u_2, u_1, inter), dim = 1) | |
def _build(self, base, expansion): | |
fl_i1 = int(base / (expansion ** 2)) | |
fl_i0 = int(base / expansion) | |
fl_0 = int(base) | |
fl_1 = int(base * expansion) | |
fl_2 = int(base * (expansion ** 2)) | |
fl_3 = int(base * (expansion ** 3)) | |
self.analysis_0 = nn.Sequential( | |
nn.Conv3d(1, fl_0, 3, 1, 1), | |
nn.BatchNorm3d(fl_0), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_0, fl_0, 3, 1, 1), | |
nn.BatchNorm3d(fl_0), | |
nn.LeakyReLU(), | |
) | |
self.analysis_1 = nn.Sequential( | |
nn.MaxPool3d(2), | |
nn.Conv3d(fl_0, fl_1, 3, 1, 1), | |
nn.BatchNorm3d(fl_1), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_1, fl_1, 3, 1, 1), | |
nn.BatchNorm3d(fl_1), | |
nn.LeakyReLU(), | |
) | |
self.analysis_2 = nn.Sequential( | |
nn.MaxPool3d(2), | |
nn.Conv3d(fl_1, fl_2, 3, 1, 1), | |
nn.BatchNorm3d(fl_2), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_2, fl_2, 3, 1, 1), | |
nn.BatchNorm3d(fl_2), | |
nn.LeakyReLU(), | |
) | |
self.analysis_3 = nn.Sequential( | |
nn.MaxPool3d(2), | |
nn.Conv3d(fl_2, fl_3, 3, 1, 1), | |
nn.BatchNorm3d(fl_3), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_3, fl_3, 3, 1, 1), | |
nn.BatchNorm3d(fl_3), | |
nn.LeakyReLU(), | |
) | |
self.bridge = nn.Sequential( | |
nn.MaxPool3d(2), | |
nn.Conv3d(fl_3, fl_3, 3, 1, 1), | |
nn.BatchNorm3d(fl_3), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_3, fl_3, 3, 1, 1), | |
nn.BatchNorm3d(fl_3), | |
nn.LeakyReLU(), | |
nn.ConvTranspose3d(fl_3, fl_3, 2, 2, 0), | |
) | |
self.synthesis_3 = nn.Sequential( | |
nn.Conv3d(fl_3 + fl_3, fl_2, 3, 1, 1), | |
nn.BatchNorm3d(fl_2), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_2, fl_2, 3, 1, 1), | |
nn.BatchNorm3d(fl_2), | |
nn.LeakyReLU(), | |
) | |
self.synthesis_2 = nn.Sequential( | |
nn.Conv3d(fl_2 + fl_2, fl_1, 3, 1, 1), | |
nn.BatchNorm3d(fl_1), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_1, fl_1, 3, 1, 1), | |
nn.BatchNorm3d(fl_1), | |
nn.LeakyReLU(), | |
) | |
self.synthesis_1 = nn.Sequential( | |
nn.Conv3d(fl_1 + fl_1, fl_0, 3, 1, 1), | |
nn.BatchNorm3d(fl_0), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_0, fl_0, 3, 1, 1), | |
nn.BatchNorm3d(fl_0), | |
nn.LeakyReLU(), | |
) | |
self.up_conv_3 = nn.Sequential( | |
nn.ConvTranspose3d(fl_2, fl_2, 2, 2, 0), | |
) | |
self.up_conv_2 = nn.Sequential( | |
nn.ConvTranspose3d(fl_1, fl_1, 2, 2, 0), | |
) | |
self.up_conv_1 = nn.Sequential( | |
nn.ConvTranspose3d(fl_0, fl_0, 2, 2, 0), | |
) | |
self.synthesis_0 = nn.Sequential( | |
nn.Conv3d(fl_0 + fl_0, fl_0, 3, 1, 1), | |
nn.BatchNorm3d(fl_0), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_0, fl_0, 3, 1, 1), | |
nn.BatchNorm3d(fl_0), | |
nn.LeakyReLU(), | |
) | |
self.output_0 = nn.Sequential( | |
nn.Conv3d(fl_0, fl_i0, 3, 1, 1), | |
nn.BatchNorm3d(fl_i0), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_i0, 1, 3, 1, 1), | |
nn.Sigmoid() | |
) | |
self.up_sample_3 = nn.Sequential( | |
nn.ConvTranspose3d(fl_2, fl_0, 8, 8, 0), | |
) | |
self.up_sample_2 = nn.Sequential( | |
nn.ConvTranspose3d(fl_1, fl_i0, 4, 4, 0), | |
) | |
self.up_sample_1 = nn.Sequential( | |
nn.ConvTranspose3d(fl_0, fl_i1, 2, 2, 0), | |
) | |
self.output_1 = nn.Sequential( | |
nn.Conv3d(fl_0 + fl_0 + fl_i0 + fl_i1, fl_0, 3, 1, 1), | |
nn.BatchNorm3d(fl_0), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_0, fl_0, 3, 1, 1), | |
nn.BatchNorm3d(fl_0), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_0, fl_i0, 3, 1, 1), | |
nn.BatchNorm3d(fl_i0), | |
nn.LeakyReLU(), | |
nn.Conv3d(fl_i0, 1, 3, 1, 1), | |
nn.Sigmoid() | |
) | |