|
|
import torch
|
|
|
from rscd.models.backbones import Decompose
|
|
|
|
|
|
|
|
|
class ResNet3D(torch.nn.Module):
|
|
|
def __init__(self, resnet2d):
|
|
|
super(ResNet3D, self).__init__()
|
|
|
self.conv1 = Decompose.Decompose_conv(resnet2d.conv1, time_dim=3, time_padding=1, center=True)
|
|
|
self.bn1 = Decompose.Decompose_norm(resnet2d.bn1)
|
|
|
self.relu = torch.nn.ReLU(inplace=True)
|
|
|
self.maxpool = Decompose.Decompose_pool(resnet2d.maxpool, time_dim=1, time_padding=0, time_stride=1)
|
|
|
|
|
|
self.layer1 = Decompose_layer(resnet2d.layer1)
|
|
|
self.layer2 = Decompose_layer(resnet2d.layer2)
|
|
|
self.layer3 = Decompose_layer(resnet2d.layer3)
|
|
|
self.layer4 = Decompose_layer(resnet2d.layer4)
|
|
|
|
|
|
def forward(self, x):
|
|
|
x = self.conv1(x)
|
|
|
x = self.bn1(x)
|
|
|
x = self.relu(x)
|
|
|
|
|
|
x = self.maxpool(x)
|
|
|
|
|
|
x = self.layer1(x)
|
|
|
x = self.layer2(x)
|
|
|
x = self.layer3(x)
|
|
|
x = self.layer4(x)
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def Decompose_layer(reslayer2d):
|
|
|
reslayers3d = []
|
|
|
for layer2d in reslayer2d:
|
|
|
layer3d = Bottleneck3d(layer2d)
|
|
|
reslayers3d.append(layer3d)
|
|
|
return torch.nn.Sequential(*reslayers3d)
|
|
|
|
|
|
|
|
|
class Bottleneck3d(torch.nn.Module):
|
|
|
def __init__(self, bottleneck2d):
|
|
|
super(Bottleneck3d, self).__init__()
|
|
|
|
|
|
self.conv1 = Decompose.Decompose_conv(bottleneck2d.conv1, time_dim=3, time_padding=1,
|
|
|
time_stride=1, center=True)
|
|
|
self.bn1 = Decompose.Decompose_norm(bottleneck2d.bn1)
|
|
|
|
|
|
self.conv2 = Decompose.Decompose_conv(bottleneck2d.conv2, time_dim=3, time_padding=1,
|
|
|
time_stride=1, center=True)
|
|
|
self.bn2 = Decompose.Decompose_norm(bottleneck2d.bn2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
self.relu = torch.nn.ReLU(inplace=True)
|
|
|
|
|
|
if bottleneck2d.downsample is not None:
|
|
|
self.downsample = Decompose_downsample(bottleneck2d.downsample, time_stride=1)
|
|
|
else:
|
|
|
self.downsample = None
|
|
|
|
|
|
self.stride = bottleneck2d.stride
|
|
|
|
|
|
def forward(self, x):
|
|
|
residual = x
|
|
|
out = self.conv1(x)
|
|
|
out = self.bn1(out)
|
|
|
out = self.relu(out)
|
|
|
|
|
|
out = self.conv2(out)
|
|
|
out = self.bn2(out)
|
|
|
out = self.relu(out)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if self.downsample is not None:
|
|
|
residual = self.downsample(x)
|
|
|
|
|
|
|
|
|
out = out + residual
|
|
|
out = self.relu(out)
|
|
|
return out
|
|
|
|
|
|
|
|
|
def Decompose_downsample(downsample2d, time_stride=1):
|
|
|
downsample3d = torch.nn.Sequential(
|
|
|
Decompose.inflate_conv(downsample2d[0], time_dim=1, time_stride=time_stride, center=True),
|
|
|
Decompose.Decompose_norm(downsample2d[1]))
|
|
|
return downsample3d
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
