本文介绍在nnunet中引入Depth-Aware Feature Enhancement模块。
注意：nnunet训练需要五折交叉验证，2d/3d/cascade 共3个模型，每个五折交叉验证累计需要训练20个模型，需要很久的时间。
tips：博主有 RTX4090资源，有需要可以联系。
一 、Depth-Aware Feature Enhancement
Depth-Aware Feature Enhancement的网络结构如下图所示：
论文地址：MonoDTR: Monocular 3D Object Detection with Depth-Aware Transformer
代码参考：https://github.com/kuanchihhuang/MonoDTR
Depth-Aware Feature Enhancement结构如下：
Depth-Aware Feature Enhancement代码：
import torch
import torch.nn as nn
import torch.nn.functional as F
# 论文：MonoDTR: Monocular 3D Object Detection with Depth-Aware Transformer
# 论文地址：https://arxiv.org/pdf/2203.13310
class DepthAwareFE(nn.Module):
def __init__(self, output_channel_num):
super(DepthAwareFE, self).__init__()
self.output_channel_num = output_channel_num
self.depth_output = nn.Sequential(
nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True),
nn.Conv2d(self.output_channel_num, int(self.output_channel_num / 2), 3, padding=1),
nn.BatchNorm2d(int(self.output_channel_num / 2)),
nn.ReLU(),
nn.Conv2d(int(self.output_channel_num / 2), 96, 1),
)
self.depth_down = nn.Conv2d(96, 12, 3, stride=1, padding=1, groups=12)
self.acf = dfe_module(256, 256)
def forward(self, x):
depth = self.depth_output(x)
N, C, H, W = x.shape
depth_guide = F.interpolate(depth, size=x.size()[2:], mode='bilinear', align_corners=False)
depth_guide = self.depth_down(depth_guide)
x = x + self.acf(x, depth_guide)
return depth, depth_guide, x
class dfe_module(nn.Module):
def __init__(self, in_channels, out_channels):
super(dfe_module, self).__init__()
self.softmax = nn.Softmax(dim=-1)
self.conv1 = nn.Sequential(nn.Conv2d(in_channels, out_channels, 1, bias=False),
nn.BatchNorm2d(out_channels),
nn.ReLU(True),
nn.Dropout2d(0.2, False))
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=1, stride=1, padding=0)
def forward(self, feat_ffm, coarse_x):
N, D, H, W = coarse_x.size()
# depth prototype
feat_ffm = self.conv1(feat_ffm)
_, C, _, _ = feat_ffm.size()
proj_query = coarse_x.view(N, D, -1)
proj_key = feat_ffm.view(N, C, -1).permute(0, 2, 1)
energy = torch.bmm(proj_query, proj_key)
energy_new = torch.max(energy, -1, keepdim=True)[0].expand_as(energy) - energy
attention = self.softmax(energy_new)
# depth enhancement
attention = attention.permute(0, 2, 1)
proj_value = coarse_x.view(N, D, -1)
out = torch.bmm(attention, proj_value)
out = out.view(N, C, H, W)
out = self.conv2(out)
return out二、nnunet加入Depth-Aware Feature Enhancement
之前的教程已经提到过，nnunet的网络需要在dynamic-network-architectures中修改，并在数据集的plan中修改来实现自己的网络训练。
1、网络结构修改
在dynamic-network-architectures的architectures目录下新建defunet.py，如下图：
代码内容如下：
from typing import Union, Type, List, Tuple
import numpy as np
import torch
import torch.nn.functional as F
import torch.nn as nn
from dynamic_network_architectures.building_blocks.helper import convert_conv_op_to_dim
from dynamic_network_architectures.initialization.weight_init import InitWeights_He
from dynamic_network_architectures.initialization.weight_init import init_last_bn_before_add_to_0
from torch import nn
from torch.nn.modules.conv import _ConvNd
from torch.nn.modules.dropout import _DropoutNd
from dynamic_network_architectures.building_blocks.helper import maybe_convert_scalar_to_list, get_matching_pool_op
from dynamic_network_architectures.building_blocks.helper import get_matching_convtransp
class DEFPlainConvUNet(nn.Module):
def __init__(self,
input_channels: int,
n_stages: int,
features_per_stage: Union[int, List[int], Tuple[int, ...]],
conv_op: Type[_ConvNd],
kernel_sizes: Union[int, List[int], Tuple[int, ...]],
strides: Union[int, List[int], Tuple[int, ...]],
n_conv_per_stage: Union[int, List[int], Tuple[int, ...]],
num_classes: int,
n_conv_per_stage_decoder: Union[int, Tuple[int, ...], List[int]],
conv_bias: bool = False,
norm_op: Union[None, Type[nn.Module]] = None,
norm_op_kwargs: dict = None,
dropout_op: Union[None, Type[_DropoutNd]] = None,
dropout_op_kwargs: dict = None,
nonlin: Union[None, Type[torch.nn.Module]] = None,
nonlin_kwargs: dict = None,
deep_supervision: bool = False,
nonlin_first: bool = False
):
"""
nonlin_first: if True you get conv -> nonlin -> norm. Else it's conv -> norm -> nonlin
"""
super().__init__()
if isinstance(n_conv_per_stage, int):
n_conv_per_stage = [n_conv_per_stage] * n_stages
if isinstance(n_conv_per_stage_decoder, int):
n_conv_per_stage_decoder = [n_conv_per_stage_decoder] * (n_stages - 1)
assert len(n_conv_per_stage) == n_stages, "n_conv_per_stage must have as many entries as we have " \
f"resolution stages. here: {n_stages}. " \
f"n_conv_per_stage: {n_conv_per_stage}"
assert len(n_conv_per_stage_decoder) == (n_stages - 1), "n_conv_per_stage_decoder must have one less entries " \
f"as we have resolution stages. here: {n_stages} " \
f"stages, so it should have {n_stages - 1} entries. " \
f"n_conv_per_stage_decoder: {n_conv_per_stage_decoder}"
self.encoder = PlainConvEncoder(input_channels, n_stages, features_per_stage, conv_op, kernel_sizes, strides,
n_conv_per_stage, conv_bias, norm_op, norm_op_kwargs, dropout_op,
dropout_op_kwargs, nonlin, nonlin_kwargs, return_skips=True,
nonlin_first=nonlin_first)
self.decoder = UNetDecoder(self.encoder, num_classes, n_conv_per_stage_decoder, deep_supervision,
nonlin_first=nonlin_first)
print('-------------------------------------DEFPlainConvUNet------------------------------------')
def forward(self, x):
skips = self.encoder(x)
return self.decoder(skips)
def compute_conv_feature_map_size(self, input_size):
assert len(input_size) == convert_conv_op_to_dim(self.encoder.conv_op), "just give the image size without color/feature channels or " \
"batch channel. Do not give input_size=(b, c, x, y(, z)). " \
"Give input_size=(x, y(, z))!"
return self.encoder.compute_conv_feature_map_size(input_size) + self.decoder.compute_conv_feature_map_size(input_size)
@staticmethod
def initialize(module):
InitWeights_He(1e-2)(module)
class PlainConvEncoder(nn.Module):
def __init__(self,
input_channels: int,
n_stages: int,
features_per_stage: Union[int, List[int], Tuple[int, ...]],
conv_op: Type[_ConvNd],
kernel_sizes: Union[int, List[int], Tuple[int, ...]],
strides: Union[int, List[int], Tuple[int, ...]],
n_conv_per_stage: Union[int, List[int], Tuple[int, ...]],
conv_bias: bool = False,
norm_op: Union[None, Type[nn.Module]] = None,
norm_op_kwargs: dict = None,
dropout_op: Union[None, Type[_DropoutNd]] = None,
dropout_op_kwargs: dict = None,
nonlin: Union[None, Type[torch.nn.Module]] = None,
nonlin_kwargs: dict = None,
return_skips: bool = False,
nonlin_first: bool = False,
pool: str = 'conv'
):
super().__init__()
if isinstance(kernel_sizes, int):
kernel_sizes = [kernel_sizes] * n_stages
if isinstance(features_per_stage, int):
features_per_stage = [features_per_stage] * n_stages
if isinstance(n_conv_per_stage, int):
n_conv_per_stage = [n_conv_per_stage] * n_stages
if isinstance(strides, int):
strides = [strides] * n_stages
assert len(kernel_sizes) == n_stages, "kernel_sizes must have as many entries as we have resolution stages (n_stages)"
assert len(n_conv_per_stage) == n_stages, "n_conv_per_stage must have as many entries as we have resolution stages (n_stages)"
assert len(features_per_stage) == n_stages, "features_per_stage must have as many entries as we have resolution stages (n_stages)"
assert len(strides) == n_stages, "strides must have as many entries as we have resolution stages (n_stages). " \
"Important: first entry is recommended to be 1, else we run strided conv drectly on the input"
stages = []
for s in range(n_stages):
stage_modules = []
if pool == 'max' or pool == 'avg':
if (isinstance(strides[s], int) and strides[s] != 1) or \
isinstance(strides[s], (tuple, list)) and any([i != 1 for i in strides[s]]):
stage_modules.append(get_matching_pool_op(conv_op, pool_type=pool)(kernel_size=strides[s], stride=strides[s]))
conv_stride = 1
elif pool == 'conv':
conv_stride = strides[s]
else:
raise RuntimeError()
stage_modules.append(StackedConvBlocks(
n_conv_per_stage[s], conv_op, input_channels, features_per_stage[s], kernel_sizes[s], conv_stride,
conv_bias, norm_op, norm_op_kwargs, dropout_op, dropout_op_kwargs, nonlin, nonlin_kwargs, nonlin_first
))
stages.append(nn.Sequential(*stage_modules))
input_channels = features_per_stage[s]
self.stages = nn.Sequential(*stages)
self.output_channels = features_per_stage
self.strides = [maybe_convert_scalar_to_list(conv_op, i) for i in strides]
self.return_skips = return_skips
# we store some things that a potential decoder needs
self.conv_op = conv_op
self.norm_op = norm_op
self.norm_op_kwargs = norm_op_kwargs
self.nonlin = nonlin
self.nonlin_kwargs = nonlin_kwargs
self.dropout_op = dropout_op
self.dropout_op_kwargs = dropout_op_kwargs
self.conv_bias = conv_bias
self.kernel_sizes = kernel_sizes
def forward(self, x):
ret = []
for s in self.stages:
x = s(x)
ret.append(x)
if self.return_skips:
return ret
else:
return ret[-1]
def compute_conv_feature_map_size(self, input_size):
output = np.int64(0)
for s in range(len(self.stages)):
if isinstance(self.stages[s], nn.Sequential):
for sq in self.stages[s]:
if hasattr(sq, 'compute_conv_feature_map_size'):
output += self.stages[s][-1].compute_conv_feature_map_size(input_size)
else:
output += self.stages[s].compute_conv_feature_map_size(input_size)
input_size = [i // j for i, j in zip(input_size, self.strides[s])]
return output
class UNetDecoder(nn.Module):
def __init__(self,
encoder: Union[PlainConvEncoder],
num_classes: int,
n_conv_per_stage: Union[int, Tuple[int, ...], List[int]],
deep_supervision,
nonlin_first: bool = False,
norm_op: Union[None, Type[nn.Module]] = None,
norm_op_kwargs: dict = None,
dropout_op: Union[None, Type[_DropoutNd]] = None,
dropout_op_kwargs: dict = None,
nonlin: Union[None, Type[torch.nn.Module]] = None,
nonlin_kwargs: dict = None,
conv_bias: bool = None
):
"""
This class needs the skips of the encoder as input in its forward.
the encoder goes all the way to the bottleneck, so that's where the decoder picks up. stages in the decoder
are sorted by order of computation, so the first stage has the lowest resolution and takes the bottleneck
features and the lowest skip as inputs
the decoder has two (three) parts in each stage:
1) conv transpose to upsample the feature maps of the stage below it (or the bottleneck in case of the first stage)
2) n_conv_per_stage conv blocks to let the two inputs get to know each other and merge
3) (optional if deep_supervision=True) a segmentation output Todo: enable upsample logits?
:param encoder:
:param num_classes:
:param n_conv_per_stage:
:param deep_supervision:
"""
super().__init__()
self.deep_supervision = deep_supervision
self.encoder = encoder
self.num_classes = num_classes
n_stages_encoder = len(encoder.output_channels)
if isinstance(n_conv_per_stage, int):
n_conv_per_stage = [n_conv_per_stage] * (n_stages_encoder - 1)
assert len(n_conv_per_stage) == n_stages_encoder - 1, "n_conv_per_stage must have as many entries as we have " \
"resolution stages - 1 (n_stages in encoder - 1), " \
"here: %d" % n_stages_encoder
transpconv_op = get_matching_convtransp(conv_op=encoder.conv_op)
conv_bias = encoder.conv_bias if conv_bias is None else conv_bias
norm_op = encoder.norm_op if norm_op is None else norm_op
norm_op_kwargs = encoder.norm_op_kwargs if norm_op_kwargs is None else norm_op_kwargs
dropout_op = encoder.dropout_op if dropout_op is None else dropout_op
dropout_op_kwargs = encoder.dropout_op_kwargs if dropout_op_kwargs is None else dropout_op_kwargs
nonlin = encoder.nonlin if nonlin is None else nonlin
nonlin_kwargs = encoder.nonlin_kwargs if nonlin_kwargs is None else nonlin_kwargs
# we start with the bottleneck and work out way up
stages = []
transpconvs = []
seg_layers = []
def_layers = []
for s in range(1, n_stages_encoder):
input_features_below = encoder.output_channels[-s]
input_features_skip = encoder.output_channels[-(s + 1)]
stride_for_transpconv = encoder.strides[-s]
transpconvs.append(transpconv_op(
input_features_below, input_features_skip, stride_for_transpconv, stride_for_transpconv,
bias=conv_bias
))
if s < n_stages_encoder - 2:
def_layers.append(
DepthAwareFE2D(2 * input_features_skip) if encoder.conv_op == torch.nn.modules.conv.Conv2d else DepthAwareFE3D(2 * input_features_skip)
)
# input features to conv is 2x input_features_skip (concat input_features_skip with transpconv output)
stages.append(StackedConvBlocks(
n_conv_per_stage[s-1], encoder.conv_op, 2 * input_features_skip, input_features_skip,
encoder.kernel_sizes[-(s + 1)], 1,
conv_bias,
norm_op,
norm_op_kwargs,
dropout_op,
dropout_op_kwargs,
nonlin,
nonlin_kwargs,
nonlin_first
))
# we always build the deep supervision outputs so that we can always load parameters. If we don't do this
# then a model trained with deep_supervision=True could not easily be loaded at inference time where
# deep supervision is not needed. It's just a convenience thing
seg_layers.append(encoder.conv_op(input_features_skip, num_classes, 1, 1, 0, bias=True))
self.stages = nn.ModuleList(stages)
self.transpconvs = nn.ModuleList(transpconvs)
self.seg_layers = nn.ModuleList(seg_layers)
self.def_layers = nn.ModuleList(def_layers)
def forward(self, skips):
"""
we expect to get the skips in the order they were computed, so the bottleneck should be the last entry
:param skips:
:return:
"""
lres_input = skips[-1]
seg_outputs = []
for s in range(len(self.stages)):
x = self.transpconvs[s](lres_input)
x = torch.cat((x, skips[-(s+2)]), 1)
if s < len(self.def_layers):
x = self.def_layers[s](x) # 融合
x = self.stages[s](x)
if self.deep_supervision:
seg_outputs.append(self.seg_layers[s](x))
elif s == (len(self.stages) - 1):
seg_outputs.append(self.seg_layers[-1](x))
lres_input = x
# invert seg outputs so that the largest segmentation prediction is returned first
seg_outputs = seg_outputs[::-1]
if not self.deep_supervision:
r = seg_outputs[0]
else:
r = seg_outputs
return r
def compute_conv_feature_map_size(self, input_size):
"""
IMPORTANT: input_size is the input_size of the encoder!
:param input_size:
:return:
"""
# first we need to compute the skip sizes. Skip bottleneck because all output feature maps of our ops will at
# least have the size of the skip above that (therefore -1)
skip_sizes = []
for s in range(len(self.encoder.strides) - 1):
skip_sizes.append([i // j for i, j in zip(input_size, self.encoder.strides[s])])
input_size = skip_sizes[-1]
# print(skip_sizes)
assert len(skip_sizes) == len(self.stages)
# our ops are the other way around, so let's match things up
output = np.int64(0)
for s in range(len(self.stages)):
# print(skip_sizes[-(s+1)], self.encoder.output_channels[-(s+2)])
# conv blocks
output += self.stages[s].compute_conv_feature_map_size(skip_sizes[-(s+1)])
# trans conv
output += np.prod([self.encoder.output_channels[-(s+2)], *skip_sizes[-(s+1)]], dtype=np.int64)
# segmentation
if self.deep_supervision or (s == (len(self.stages) - 1)):
output += np.prod([self.num_classes, *skip_sizes[-(s+1)]], dtype=np.int64)
return output
class StackedConvBlocks(nn.Module):
def __init__(self,
num_convs: int,
conv_op: Type[_ConvNd],
input_channels: int,
output_channels: Union[int, List[int], Tuple[int, ...]],
kernel_size: Union[int, List[int], Tuple[int, ...]],
initial_stride: Union[int, List[int], Tuple[int, ...]],
conv_bias: bool = False,
norm_op: Union[None, Type[nn.Module]] = None,
norm_op_kwargs: dict = None,
dropout_op: Union[None, Type[_DropoutNd]] = None,
dropout_op_kwargs: dict = None,
nonlin: Union[None, Type[torch.nn.Module]] = None,
nonlin_kwargs: dict = None,
nonlin_first: bool = False
):
"""
:param conv_op:
:param num_convs:
:param input_channels:
:param output_channels: can be int or a list/tuple of int. If list/tuple are provided, each entry is for
one conv. The length of the list/tuple must then naturally be num_convs
:param kernel_size:
:param initial_stride:
:param conv_bias:
:param norm_op:
:param norm_op_kwargs:
:param dropout_op:
:param dropout_op_kwargs:
:param nonlin:
:param nonlin_kwargs:
"""
super().__init__()
if not isinstance(output_channels, (tuple, list)):
output_channels = [output_channels] * num_convs
self.convs = nn.Sequential(
ConvDropoutNormReLU(
conv_op, input_channels, output_channels[0], kernel_size, initial_stride, conv_bias, norm_op,
norm_op_kwargs, dropout_op, dropout_op_kwargs, nonlin, nonlin_kwargs, nonlin_first
),
*[
ConvDropoutNormReLU(
conv_op, output_channels[i - 1], output_channels[i], kernel_size, 1, conv_bias, norm_op,
norm_op_kwargs, dropout_op, dropout_op_kwargs, nonlin, nonlin_kwargs, nonlin_first
)
for i in range(1, num_convs)
]
)
self.output_channels = output_channels[-1]
self.initial_stride = maybe_convert_scalar_to_list(conv_op, initial_stride)
def forward(self, x):
return self.convs(x)
def compute_conv_feature_map_size(self, input_size):
assert len(input_size) == len(self.initial_stride), "just give the image size without color/feature channels or " \
"batch channel. Do not give input_size=(b, c, x, y(, z)). " \
"Give input_size=(x, y(, z))!"
output = self.convs[0].compute_conv_feature_map_size(input_size)
size_after_stride = [i // j for i, j in zip(input_size, self.initial_stride)]
for b in self.convs[1:]:
output += b.compute_conv_feature_map_size(size_after_stride)
return output
class ConvDropoutNormReLU(nn.Module):
def __init__(self,
conv_op: Type[_ConvNd],
input_channels: int,
output_channels: int,
kernel_size: Union[int, List[int], Tuple[int, ...]],
stride: Union[int, List[int], Tuple[int, ...]],
conv_bias: bool = False,
norm_op: Union[None, Type[nn.Module]] = None,
norm_op_kwargs: dict = None,
dropout_op: Union[None, Type[_DropoutNd]] = None,
dropout_op_kwargs: dict = None,
nonlin: Union[None, Type[torch.nn.Module]] = None,
nonlin_kwargs: dict = None,
nonlin_first: bool = False
):
super(ConvDropoutNormReLU, self).__init__()
self.input_channels = input_channels
self.output_channels = output_channels
stride = maybe_convert_scalar_to_list(conv_op, stride)
self.stride = stride
kernel_size = maybe_convert_scalar_to_list(conv_op, kernel_size)
if norm_op_kwargs is None:
norm_op_kwargs = {}
if nonlin_kwargs is None:
nonlin_kwargs = {}
ops = []
self.conv = conv_op(
input_channels,
output_channels,
kernel_size,
stride,
padding=[(i - 1) // 2 for i in kernel_size],
dilation=1,
bias=conv_bias,
)
ops.append(self.conv)
if dropout_op is not None:
self.dropout = dropout_op(**dropout_op_kwargs)
ops.append(self.dropout)
if norm_op is not None:
self.norm = norm_op(output_channels, **norm_op_kwargs)
ops.append(self.norm)
if nonlin is not None:
self.nonlin = nonlin(**nonlin_kwargs)
ops.append(self.nonlin)
if nonlin_first and (norm_op is not None and nonlin is not None):
ops[-1], ops[-2] = ops[-2], ops[-1]
self.all_modules = nn.Sequential(*ops)
def forward(self, x):
return self.all_modules(x)
def compute_conv_feature_map_size(self, input_size):
assert len(input_size) == len(self.stride), "just give the image size without color/feature channels or " \
"batch channel. Do not give input_size=(b, c, x, y(, z)). " \
"Give input_size=(x, y(, z))!"
output_size = [i // j for i, j in zip(input_size, self.stride)] # we always do same padding
return np.prod([self.output_channels, *output_size], dtype=np.int64)
class DepthAwareFE2D(nn.Module):
def __init__(self, output_channel_num):
super(DepthAwareFE2D, self).__init__()
self.output_channel_num = output_channel_num
self.depth_output = nn.Sequential(
nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True),
nn.Conv2d(self.output_channel_num, int(self.output_channel_num / 2), 3, padding=1),
nn.BatchNorm2d(int(self.output_channel_num / 2)),
nn.ReLU(),
nn.Conv2d(int(self.output_channel_num / 2), 96, 1),
)
self.depth_down = nn.Conv2d(96, 12, 3, stride=1, padding=1, groups=12)
self.acf = dfe_module2D(output_channel_num, output_channel_num)
def forward(self, x):
depth = self.depth_output(x)
N, C, H, W = x.shape
depth_guide = F.interpolate(depth, size=x.size()[2:], mode='bilinear', align_corners=False)
depth_guide = self.depth_down(depth_guide)
x = x + self.acf(x, depth_guide)
# return depth, depth_guide, x
return x
class DepthAwareFE3D(nn.Module):
def __init__(self, output_channel_num):
super(DepthAwareFE3D, self).__init__()
self.output_channel_num = output_channel_num
self.depth_output = nn.Sequential(
nn.Upsample(scale_factor=2, mode='trilinear', align_corners=True),
nn.Conv3d(self.output_channel_num, int(self.output_channel_num / 2), 3, padding=1),
nn.BatchNorm3d(int(self.output_channel_num / 2)),
nn.ReLU(),
nn.Conv3d(int(self.output_channel_num / 2), 96, 1),
)
self.depth_down = nn.Conv3d(96, 12, 3, stride=1, padding=1, groups=12)
self.acf = dfe_module3D(output_channel_num, output_channel_num)
def forward(self, x):
depth = self.depth_output(x)
N, C, H, W, Z = x.shape
depth_guide = F.interpolate(depth, size=x.size()[2:], mode='trilinear', align_corners=False)
depth_guide = self.depth_down(depth_guide)
x = x + self.acf(x, depth_guide)
# return depth, depth_guide, x
return x
class dfe_module2D(nn.Module):
def __init__(self, in_channels, out_channels):
super(dfe_module2D, self).__init__()
self.softmax = nn.Softmax(dim=-1)
self.conv1 = nn.Sequential(nn.Conv2d(in_channels, out_channels, 1, bias=False),
nn.BatchNorm2d(out_channels),
nn.ReLU(True),
nn.Dropout2d(0.2, False))
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=1, stride=1, padding=0)
def forward(self, feat_ffm, coarse_x):
N, D, H, W = coarse_x.size()
# depth prototype
feat_ffm = self.conv1(feat_ffm)
_, C, _, _ = feat_ffm.size()
proj_query = coarse_x.view(N, D, -1)
proj_key = feat_ffm.view(N, C, -1).permute(0, 2, 1)
energy = torch.bmm(proj_query, proj_key)
energy_new = torch.max(energy, -1, keepdim=True)[0].expand_as(energy) - energy
attention = self.softmax(energy_new)
# depth enhancement
attention = attention.permute(0, 2, 1)
proj_value = coarse_x.view(N, D, -1)
out = torch.bmm(attention, proj_value)
out = out.view(N, C, H, W)
out = self.conv2(out)
return out
class dfe_module3D(nn.Module):
def __init__(self, in_channels, out_channels):
super(dfe_module3D, self).__init__()
self.softmax = nn.Softmax(dim=-1)
self.conv1 = nn.Sequential(nn.Conv3d(in_channels, out_channels, 1, bias=False),
nn.BatchNorm3d(out_channels),
nn.ReLU(True),
nn.Dropout3d(0.2, False))
self.conv2 = nn.Conv3d(out_channels, out_channels, kernel_size=1, stride=1, padding=0)
def forward(self, feat_ffm, coarse_x):
N, D, H, W, Z = coarse_x.size()
# depth prototype
feat_ffm = self.conv1(feat_ffm)
_, C, _, _, _ = feat_ffm.size()
proj_query = coarse_x.view(N, D, -1)
proj_key = feat_ffm.view(N, C, -1).permute(0, 2, 1)
energy = torch.bmm(proj_query, proj_key)
energy_new = torch.max(energy, -1, keepdim=True)[0].expand_as(energy) - energy
attention = self.softmax(energy_new)
# depth enhancement
attention = attention.permute(0, 2, 1)
proj_value = coarse_x.view(N, D, -1)
out = torch.bmm(attention, proj_value)
out = out.view(N, C, H, W, Z)
out = self.conv2(out)
return out
2、配置文件修改
在完成了模型修改后，还是用上个教程的Task04_Hippocampus数据集来验证（如果没做上个教程的，自行完成数据处理），编辑nnUNet\nnUNet_preprocessed\Dataset004_Hippocampus\nnUNetPlans.json这个配置文件，进行以下改动，把network_class_name改成dynamic_network_architectures.architectures.defunet.DEFPlainConvUNet，如下图：
三、模型训练
完成了模型和数据集配置文件的修改后，开始训练模型，使用的数据集还是Task04_Hippocampus，以上的代码支持2d和3d模型，可以使用以下的训练命令：
nnUNetv2_train 4 2d 0
nnUNetv2_train 4 2d 1
nnUNetv2_train 4 2d 2
nnUNetv2_train 4 2d 3
nnUNetv2_train 4 2d 4
nnUNetv2_train 4 3d_fullres 0
nnUNetv2_train 4 3d_fullres 1
nnUNetv2_train 4 3d_fullres 2
nnUNetv2_train 4 3d_fullres 3
nnUNetv2_train 4 3d_fullres 4 可以看到，2d模型训练起来了：
3d_fullres也训练一下（注意：这个模块在3d训练时经常出现损失为nan，不建议用于3d训练）：
因为nnunet训练非常的久，实验资源有限，没有完成全部训练，只完成了代码修改及跑通。