pytorch 中的 state_dict 是一个简单的python的字典对象,将每一层与它的对应参数建立映射关系.(如model的每一层的weights及偏置等等)
(注意,只有那些参数可以训练的layer才会被保存到模型的state_dict中,如卷积层,线性层等等)
优化器对象Optimizer也有一个state_dict,它包含了优化器的状态以及被使用的超参数(如lr, momentum,weight_decay等)
备注:
1) state_dict是在定义了model或optimizer之后pytorch自动生成的,可以直接调用.常用的保存state_dict的格式是".pt"或'.pth'的文件,即下面命令的 PATH="./***.pt"
torch.save(model.state_dict(), PATH)
2) load_state_dict 也是model或optimizer之后pytorch自动具备的函数,可以直接调用
model = TheModelClass(*args, **kwargs)
model.load_state_dict(torch.load(PATH))
model.eval()
注意:model.eval() 的重要性,在2)中最后用到了model.eval(),是因为,只有在执行该命令后,"dropout层"及"batch normalization层"才会进入 evalution 模态. 而在"训练(training)模态"与"评估(evalution)模态"下,这两层有不同的表现形式.
-------------------------------------------------------------------------------------------------------------------------------
模态字典(state_dict)的保存(model是一个网络结构类的对象)
1.1)仅保存学习到的参数,用以下命令
torch.save(model.state_dict(), PATH)
1.2)加载model.state_dict,用以下命令
model = TheModelClass(*args, **kwargs)
model.load_state_dict(torch.load(PATH))
model.eval()
备注:model.load_state_dict的操作对象是 一个具体的对象,而不能是文件名
-----------
2.1)保存整个model的状态,用以下命令
torch.save(model,PATH)
2.2)加载整个model的状态,用以下命令:
# Model class must be defined somewhere
model = torch.load(PATH)
model.eval()
--------------------------------------------------------------------------------------------------------------------------------------
state_dict 是一个python的字典格式,以字典的格式存储,然后以字典的格式被加载,而且只加载key匹配的项
----------------------------------------------------------------------------------------------------------------------
如何仅加载某一层的训练的到的参数(某一层的state)
If you want to load parameters from one layer to another, but some keys do not match, simply change the name of the parameter keys in the state_dict that you are loading to match the keys in the model that you are loading into.
conv1_weight_state = torch.load('./model_state_dict.pt')['conv1.weight']
--------------------------------------------------------------------------------------------
加载模型参数后,如何设置某层某参数的"是否需要训练"(param.requires_grad)
for param in list(model.pretrained.parameters()):
param.requires_grad = False
注意: requires_grad的操作对象是tensor.
疑问:能否直接对某个层直接之用requires_grad呢?例如:model.conv1.requires_grad=False
回答:经测试,不可以.model.conv1 没有requires_grad属性.
---------------------------------------------------------------------------------------------
全部测试代码:
#-*-coding:utf-8-*-
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
# define model
class TheModelClass(nn.Module):
def __init__(self):
super(TheModelClass,self).__init__()
self.conv1 = nn.Conv2d(3,6,5)
self.pool = nn.MaxPool2d(2,2)
self.conv2 = nn.Conv2d(6,16,5)
self.fc1 = nn.Linear(16*5*5,120)
self.fc2 = nn.Linear(120,84)
self.fc3 = nn.Linear(84,10)
def forward(self,x):
x = self.pool(F.relu(self.conv1(x)))
x = self.pool(F.relu(self.conv2(x)))
x = x.view(-1,16*5*5)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
# initial model
model = TheModelClass()
#initialize the optimizer
optimizer = optim.SGD(model.parameters(),lr=0.001,momentum=0.9)
# print the model's state_dict
print("model's state_dict:")
for param_tensor in model.state_dict():
print(param_tensor,' ',model.state_dict()[param_tensor].size())
print("
optimizer's state_dict")
for var_name in optimizer.state_dict():
print(var_name,' ',optimizer.state_dict()[var_name])
print("
print particular param")
print('
',model.conv1.weight.size())
print('
',model.conv1.weight)
print("------------------------------------")
torch.save(model.state_dict(),'./model_state_dict.pt')
# model_2 = TheModelClass()
# model_2.load_state_dict(torch.load('./model_state_dict'))
# model.eval()
# print('
',model_2.conv1.weight)
# print((model_2.conv1.weight == model.conv1.weight).size())
## 仅仅加载某一层的参数
conv1_weight_state = torch.load('./model_state_dict.pt')['conv1.weight']
print(conv1_weight_state==model.conv1.weight)
model_2 = TheModelClass()
model_2.load_state_dict(torch.load('./model_state_dict.pt'))
model_2.conv1.requires_grad=False
print(model_2.conv1.requires_grad)
print(model_2.conv1.bias.requires_grad)
---------------------
作者：wzg2016
来源：CSDN
原文：https://blog.csdn.net/strive_for_future/article/details/83240081
版权声明：本文为博主原创文章，转载请附上博文链接！