Pytroch网络参数的快速增删改查（实例讲解）

Relu>Prelu

网络模型代码

class TumorNet(nn.Module):
	def __init__(self):
		super(TumorNet, self).__init__()
		self. nff = [1, 8]#NumFeature
		# forward1
		self.forward1 = nn.Sequential(
			nn.Conv3d(self.nff[0], self.nff[1], kernel_size=3, padding=1),
			nn.InstanceNorm3d(self.nff[1]),
			nn.ReLU(inplace=True),
			# nn.PReLU(),
			nn.Conv3d(self.nff[1], self.nff[1], kernel_size=3, padding=1),
			nn.InstanceNorm3d(self.nff[1]),
			nn.ReLU(inplace=True)
			# nn.PReLU()
		)
	def forward(self, input):
		#encoder
		layer1 = self.forward1(input)
		return layer1

Step1: 打印网络修改前后的参数名和shape

此步骤帮助你对模型改动一目了然，然后确定参数修改方案。

net = TumorNet().cuda()
for name, param in net.named_parameters():
	print(name,param.shape)

其中红色为“增加”的网络参数。
注：模型的参数表是orderDict类型（有序字典），旧模型中的relu虽然没有参数，但是也会占用编号。

Step2: 确定并实施修改方案

import torch
dict = torch.load("./results/model_0-9-8222.pth")
default = torch.Tensor([0.25])# this step is of great importance, not 0.25 or[0.25] 
dict['forward1.2.weight'] = default
dict["forward1.5.weight"] = default
torch.save(dict, "./results/Prelu_model_0-9-8222.pth")

Pool>Stride Convolution

网络模型代码

from torch import nn
class TumorNet(nn.Module):
	def __init__(self):
		super(TumorNet, self).__init__()
		self. nff = [1, 8, 16]#NumFeature
		# forward1
		self.forward1 = nn.Sequential(
			nn.Conv3d(self.nff[0], self.nff[1], kernel_size=3, padding=1),
			nn.InstanceNorm3d(self.nff[1]),
			# nn.ReLU(inplace=True),
			nn.PReLU(),
			nn.Conv3d(self.nff[1], self.nff[1], kernel_size=3, padding=1),
			nn.InstanceNorm3d(self.nff[1]),
			# nn.ReLU(inplace=True)
			nn.PReLU()
		)
		self.maxpool = nn.MaxPool3d(kernel_size=2, stride=2)
		# self.downsamp1 = nn.Conv3d(self.nff[1], self.nff[2], kernel_size=3, stride=2, padding=1)

	def forward(self, input):
		#encoder
		layer1 = self.forward1(input)
		down1 = self.maxpool(layer1)
		# down1 = self.downsamp1(layer1)
		return down1

Step1: 打印网络修改前后的参数名和shape

此步骤帮助你对模型改动一目了然，然后确定参数修改方案。

net = TumorNet().cuda()
for name, param in net.named_parameters():
	print(name,param.shape)

其中红色为“增加”的网络参数。

Step2: 确定并实施修改方案

import torch
dict = torch.load("./results/model_0-9-8222.pth")
# layer absent
for name, param in net.named_parameters():
    if name.startswith('downsamp'):
        print(name,param.shape)
        if name.endswith('weight'):
            weight = torch.empty(param.shape)
            dict[name] = nn.init.kaiming_uniform_(weight, 0.25)
        if name.endswith('bias'):
            bias = torch.empty(param.shape)
            dict[name] = nn.init.constant_(bias, 0)
torch.save(dict, "./results/Prelu_model_0-9-8222.pth")

神经网络的权重初始化：
针对ReLU的初始化方法：torch.nn.init.kaiming_uniform_(), torch.nn.init.kaiming_normal_()
Pytroch网络模型：修改参数值，修改参数名，添加参数层，删除参数层