主干网络的输入通道数从3改为1，同时保证预训练权重输出

在医学图像处理时，B超图像、CT图像一般被处理为灰度图，也就是通道数为1。然而一些优秀的主干特征提取网络例如ResNet、VGGNet、DenseNet，以及近几年推出的Swin Transformer、ConvNeXt等，它们都是基于RGB三通道数据进行预训练。这就导致了我们如果要使用预训练模型参数，就得把原本(H, W, 1)的灰度数据拼接为(H, W, 3)，这很大地降低了显存的利用效率。我们还可以选择对预训练权重进行修改，在保证模型输出与预训练模型相同的情况下，支持单通道图像。

例：ResNet50

我们先从官方给的resnet复现开始

输入以下代码，按住ctrl点击刚输入的“resnet”，即可进入源码。

from torchvision.models import resnet

往下翻可以找到各个模型对应的预训练权重地址，下载resnet50的权重。

model_urls = {
    'resnet18': 'https://download.pytorch.org/models/resnet18-f37072fd.pth',
    'resnet34': 'https://download.pytorch.org/models/resnet34-b627a593.pth',
    'resnet50': 'https://download.pytorch.org/models/resnet50-0676ba61.pth',
    'resnet101': 'https://download.pytorch.org/models/resnet101-63fe2227.pth',
    'resnet152': 'https://download.pytorch.org/models/resnet152-394f9c45.pth',
    'resnext50_32x4d': 'https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth',
    'resnext101_32x8d': 'https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth',
    'wide_resnet50_2': 'https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pth',
    'wide_resnet101_2': 'https://download.pytorch.org/models/wide_resnet101_2-32ee1156.pth',
}

 将权值文件放到项目下的model_weights文件夹下（什么文件夹请随意），用pytorch来load，这些键值就是模块的名称，找到第一个卷积层的weight（bias不用动），发现他的size是(64, 3, 7, 7)。

import torch
resnet50_weights = torch.load('./model_weights/resnet50-0676ba61.pth')
print(list(resnet50_weights.keys()))
# ['conv1.weight', 'bn1.running_mean', 'bn1.running_var', 'bn1.weight', 'bn1.bias', 'layer1.0.conv1.weight', 'layer1.0.bn1.running_mean', 'layer1.0.bn1.running_var', 'layer1.0.bn1.weight', 'layer1.0.bn1.bias', 'layer1.0.conv2.weight', 'layer1.0.bn2.running_mean', 'layer1.0.bn2.running_var', 'layer1.0.bn2.weight', 'layer1.0.bn2.bias', 'layer1.0.conv3.weight', 'layer1.0.bn3.running_mean', 'layer1.0.bn3.running_var', 'layer1.0.bn3.weight', 'layer1.0.bn3.bias', 'layer1.0.downsample.0.weight', 'layer1.0.downsample.1.running_mean', 'layer1.0.downsample.1.running_var', 'layer1.0.downsample.1.weight', 'layer1.0.downsample.1.bias', 'layer1.1.conv1.weight', 'layer1.1.bn1.running_mean', 'layer1.1.bn1.running_var', 'layer1.1.bn1.weight', 'layer1.1.bn1.bias', 'layer1.1.conv2.weight', 'layer1.1.bn2.running_mean', 'layer1.1.bn2.running_var', 'layer1.1.bn2.weight', 'layer1.1.bn2.bias', 'layer1.1.conv3.weight', 'layer1.1.bn3.running_mean', 'layer1.1.bn3.running_var', 'layer1.1.bn3.weight', 'layer1.1.bn3.bias', 'layer1.2.conv1.weight', 'layer1.2.bn1.running_mean', 'layer1.2.bn1.running_var', 'layer1.2.bn1.weight', 'layer1.2.bn1.bias', 'layer1.2.conv2.weight', 'layer1.2.bn2.running_mean', 'layer1.2.bn2.running_var', 'layer1.2.bn2.weight', 'layer1.2.bn2.bias', 'layer1.2.conv3.weight', 'layer1.2.bn3.running_mean', 'layer1.2.bn3.running_var', 'layer1.2.bn3.weight', 'layer1.2.bn3.bias', 'layer2.0.conv1.weight', 'layer2.0.bn1.running_mean', 'layer2.0.bn1.running_var', 'layer2.0.bn1.weight', 'layer2.0.bn1.bias', 'layer2.0.conv2.weight', 'layer2.0.bn2.running_mean', 'layer2.0.bn2.running_var', 'layer2.0.bn2.weight', 'layer2.0.bn2.bias', 'layer2.0.conv3.weight', 'layer2.0.bn3.running_mean', 'layer2.0.bn3.running_var', 'layer2.0.bn3.weight', 'layer2.0.bn3.bias', 'layer2.0.downsample.0.weight', 'layer2.0.downsample.1.running_mean', 'layer2.0.downsample.1.running_var', 'layer2.0.downsample.1.weight', 'layer2.0.downsample.1.bias', 'layer2.1.conv1.weight', 'layer2.1.bn1.running_mean', 'layer2.1.bn1.running_var', 'layer2.1.bn1.weight', 'layer2.1.bn1.bias', 'layer2.1.conv2.weight', 'layer2.1.bn2.running_mean', 'layer2.1.bn2.running_var', 'layer2.1.bn2.weight', 'layer2.1.bn2.bias', 'layer2.1.conv3.weight', 'layer2.1.bn3.running_mean', 'layer2.1.bn3.running_var', 'layer2.1.bn3.weight', 'layer2.1.bn3.bias', 'layer2.2.conv1.weight', 'layer2.2.bn1.running_mean', 'layer2.2.bn1.running_var', 'layer2.2.bn1.weight', 'layer2.2.bn1.bias', 'layer2.2.conv2.weight', 'layer2.2.bn2.running_mean', 'layer2.2.bn2.running_var', 'layer2.2.bn2.weight', 'layer2.2.bn2.bias', 'layer2.2.conv3.weight', 'layer2.2.bn3.running_mean', 'layer2.2.bn3.running_var', 'layer2.2.bn3.weight', 'layer2.2.bn3.bias', 'layer2.3.conv1.weight', 'layer2.3.bn1.running_mean', 'layer2.3.bn1.running_var', 'layer2.3.bn1.weight', 'layer2.3.bn1.bias', 'layer2.3.conv2.weight', 'layer2.3.bn2.running_mean', 'layer2.3.bn2.running_var', 'layer2.3.bn2.weight', 'layer2.3.bn2.bias', 'layer2.3.conv3.weight', 'layer2.3.bn3.running_mean', 'layer2.3.bn3.running_var', 'layer2.3.bn3.weight', 'layer2.3.bn3.bias', 'layer3.0.conv1.weight', 'layer3.0.bn1.running_mean', 'layer3.0.bn1.running_var', 'layer3.0.bn1.weight', 'layer3.0.bn1.bias', 'layer3.0.conv2.weight', 'layer3.0.bn2.running_mean', 'layer3.0.bn2.running_var', 'layer3.0.bn2.weight', 'layer3.0.bn2.bias', 'layer3.0.conv3.weight', 'layer3.0.bn3.running_mean', 'layer3.0.bn3.running_var', 'layer3.0.bn3.weight', 'layer3.0.bn3.bias', 'layer3.0.downsample.0.weight', 'layer3.0.downsample.1.running_mean', 'layer3.0.downsample.1.running_var', 'layer3.0.downsample.1.weight', 'layer3.0.downsample.1.bias', 'layer3.1.conv1.weight', 'layer3.1.bn1.running_mean', 'layer3.1.bn1.running_var', 'layer3.1.bn1.weight', 'layer3.1.bn1.bias', 'layer3.1.conv2.weight', 'layer3.1.bn2.running_mean', 'layer3.1.bn2.running_var', 'layer3.1.bn2.weight', 'layer3.1.bn2.bias', 'layer3.1.conv3.weight', 'layer3.1.bn3.running_mean', 'layer3.1.bn3.running_var', 'layer3.1.bn3.weight', 'layer3.1.bn3.bias', 'layer3.2.conv1.weight', 'layer3.2.bn1.running_mean', 'layer3.2.bn1.running_var', 'layer3.2.bn1.weight', 'layer3.2.bn1.bias', 'layer3.2.conv2.weight', 'layer3.2.bn2.running_mean', 'layer3.2.bn2.running_var', 'layer3.2.bn2.weight', 'layer3.2.bn2.bias', 'layer3.2.conv3.weight', 'layer3.2.bn3.running_mean', 'layer3.2.bn3.running_var', 'layer3.2.bn3.weight', 'layer3.2.bn3.bias', 'layer3.3.conv1.weight', 'layer3.3.bn1.running_mean', 'layer3.3.bn1.running_var', 'layer3.3.bn1.weight', 'layer3.3.bn1.bias', 'layer3.3.conv2.weight', 'layer3.3.bn2.running_mean', 'layer3.3.bn2.running_var', 'layer3.3.bn2.weight', 'layer3.3.bn2.bias', 'layer3.3.conv3.weight', 'layer3.3.bn3.running_mean', 'layer3.3.bn3.running_var', 'layer3.3.bn3.weight', 'layer3.3.bn3.bias', 'layer3.4.conv1.weight', 'layer3.4.bn1.running_mean', 'layer3.4.bn1.running_var', 'layer3.4.bn1.weight', 'layer3.4.bn1.bias', 'layer3.4.conv2.weight', 'layer3.4.bn2.running_mean', 'layer3.4.bn2.running_var', 'layer3.4.bn2.weight', 'layer3.4.bn2.bias', 'layer3.4.conv3.weight', 'layer3.4.bn3.running_mean', 'layer3.4.bn3.running_var', 'layer3.4.bn3.weight', 'layer3.4.bn3.bias', 'layer3.5.conv1.weight', 'layer3.5.bn1.running_mean', 'layer3.5.bn1.running_var', 'layer3.5.bn1.weight', 'layer3.5.bn1.bias', 'layer3.5.conv2.weight', 'layer3.5.bn2.running_mean', 'layer3.5.bn2.running_var', 'layer3.5.bn2.weight', 'layer3.5.bn2.bias', 'layer3.5.conv3.weight', 'layer3.5.bn3.running_mean', 'layer3.5.bn3.running_var', 'layer3.5.bn3.weight', 'layer3.5.bn3.bias', 'layer4.0.conv1.weight', 'layer4.0.bn1.running_mean', 'layer4.0.bn1.running_var', 'layer4.0.bn1.weight', 'layer4.0.bn1.bias', 'layer4.0.conv2.weight', 'layer4.0.bn2.running_mean', 'layer4.0.bn2.running_var', 'layer4.0.bn2.weight', 'layer4.0.bn2.bias', 'layer4.0.conv3.weight', 'layer4.0.bn3.running_mean', 'layer4.0.bn3.running_var', 'layer4.0.bn3.weight', 'layer4.0.bn3.bias', 'layer4.0.downsample.0.weight', 'layer4.0.downsample.1.running_mean', 'layer4.0.downsample.1.running_var', 'layer4.0.downsample.1.weight', 'layer4.0.downsample.1.bias', 'layer4.1.conv1.weight', 'layer4.1.bn1.running_mean', 'layer4.1.bn1.running_var', 'layer4.1.bn1.weight', 'layer4.1.bn1.bias', 'layer4.1.conv2.weight', 'layer4.1.bn2.running_mean', 'layer4.1.bn2.running_var', 'layer4.1.bn2.weight', 'layer4.1.bn2.bias', 'layer4.1.conv3.weight', 'layer4.1.bn3.running_mean', 'layer4.1.bn3.running_var', 'layer4.1.bn3.weight', 'layer4.1.bn3.bias', 'layer4.2.conv1.weight', 'layer4.2.bn1.running_mean', 'layer4.2.bn1.running_var', 'layer4.2.bn1.weight', 'layer4.2.bn1.bias', 'layer4.2.conv2.weight', 'layer4.2.bn2.running_mean', 'layer4.2.bn2.running_var', 'layer4.2.bn2.weight', 'layer4.2.bn2.bias', 'layer4.2.conv3.weight', 'layer4.2.bn3.running_mean', 'layer4.2.bn3.running_var', 'layer4.2.bn3.weight', 'layer4.2.bn3.bias', 'fc.weight', 'fc.bias']

print(resnet50_weights['conv1.weight'].size())
# torch.Size([64, 3, 7, 7])

 再找到源码中第一个卷积层的定义，其中self.inplanes为64，这与conv1.weights的size对应。

self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)

所以要把3通道变成1通道，我们只需要把权重的第一维求和，然后保存新的权重。

resnet50_weights['conv1.weight'] = resnet50_weights['conv1.weight'].sum(1, keepdim=True)
torch.save(resnet50_weights, './model_weights/resnet50_new.pth')

我们测试一下修改后的模型输出是否和修改前一样。

import torch
import torch.nn as nn
from torchvision.models import resnet50

resnet50_weights_old = torch.load('./model_weights/resnet50-0676ba61.pth')
resnet50_weights_new = torch.load('./model_weights/resnet50_new.pth')

model_old = resnet50(pretrained=False)
model_old.load_state_dict(resnet50_weights_old)
model_old.eval()

model_new = resnet50(pretrained=False)
model_new.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
model_new.load_state_dict(resnet50_weights_new)
model_new.eval()

test_img = torch.randn((1, 1, 64, 64))
test_img_ = torch.cat((test_img, test_img, test_img), 1)

print((model_old(test_img_)-model_new(test_img)).sum())
# tensor(2.6241e-05, grad_fn=<SumBackward0>)

差距之和为2.6241e-05，近似为0，说明模型的修改是成功的。