Page MenuHomec4science
Files F59743051

net_v1.py
No OneTemporary
Actions

File Metadata

Created
Wed, Apr 24, 21:47

net_v1.py
View Options

# todo:
# multi-scale for better receptive field (decoder only, bilinear input subsampling)
import torch
import torch.nn as nn
import torch.nn.functional as F
def conv_bn_relu(in_channels, out_channels, kernel_size=3):
stride=1
padding=kernel_size//2
return nn.Sequential(
nn.Conv3d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding),
nn.BatchNorm3d(out_channels),
nn.ReLU(inplace=True),
nn.Conv3d(out_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding),
nn.BatchNorm3d(out_channels),
nn.ReLU(inplace=True),
nn.Dropout3d(p=0.1),
)
def upSkipConvolution(s):
c=nn.Conv3d(s,2*s, kernel_size=1, stride=1, padding=0, bias=False)
c.weight.data.zero_()
for i in range(0,s):
c.weight.data[2*i][i]=1
c.weight.data[2*i+1][i]=1
return c
def dwSkipConvolution(s):
c=nn.Conv3d(s,s//2, kernel_size=1, stride=1, padding=0, bias=False)
c.weight.data.zero_()
for i in range(0,s//2):
c.weight.data[i][2*i]=0.5
c.weight.data[i][2*i+1]=0.5
return c
class upResBlock(nn.Module):
def __init__(self, in_channels, kernel_size=3):
super().__init__()
self.cbr=conv_bn_relu(in_channels,2*in_channels,kernel_size)
self.uc=upSkipConvolution(in_channels)
def forward(self, x):
return self.cbr(x)+self.uc(x)
class dwResBlock(nn.Module):
def __init__(self, in_channels, kernel_size=3):
super().__init__()
self.cbr=conv_bn_relu(in_channels,in_channels//2,kernel_size)
self.dc=dwSkipConvolution(in_channels)
def forward(self, x):
return self.cbr(x)+self.dc(x)
def down_pooling():
return nn.MaxPool3d(2)
def up_pooling(in_channels, out_channels, kernel_size=2, stride=2):
return nn.Sequential(
nn.ConvTranspose3d(in_channels, out_channels, kernel_size=kernel_size, stride=stride),
nn.BatchNorm3d(out_channels),
nn.ReLU(inplace=True)
)
class UNet3d(nn.Module):
def __init__(self):
super().__init__()
input_channels = 1
nclasses = 1
# go down
self.conv0 = conv_bn_relu(input_channels, 64)
self.conv1 = upResBlock(64)
self.conv2 = upResBlock(128)
self.down_pooling = nn.MaxPool3d(2)
# go up
self.up_pool6 = up_pooling(256, 128)
self.conv7 = dwResBlock(256)
self.up_pool8 = up_pooling(128, 64)
self.conv9 = dwResBlock(128)
self.conv10 = nn.Conv3d(64, 1, 1)
def forward(self, x):
# normalize input data
# go down
x0 = self.conv0(x)
p0 = self.down_pooling(x0)
x1 = self.conv1(p0)
p1 = self.down_pooling(x1)
x2 = self.conv2(p1)
x6 = x2
# go up
p7 = self.up_pool6(x6)
x7 = torch.cat([p7, x1], dim=1)
x7 = self.conv7(x7)
p8 = self.up_pool8(x7)
x8 = torch.cat([p8, x0], dim=1)
x8 = self.conv9(x8)
output = F.pad(self.conv10(x8), [0,0, 0,0, 0,0, 1,0])
return output

Event Timeline

c4science · Help