PyTorch保存和加载模型方法汇总-马春杰杰

文章目录

看到一篇非常好的文章，介绍pytorch保存和加载模型的，非常详细。备份：

太长不看版：

方法1：

torch.save(modelA.state_dict(), PATH)

model.load_state_dict(torch.load(PATH), strict=False)

方法2：

torch.save(model, PATH)

model = torch.load(PATH)

0 前言

这篇博客主要是对使用PyTorch保存和加载训练模型参数的一个学习记录。第1-5小节是比较常规的模型参数保存操作，第6小是用已经训练好的模型参数来初始化新的模型，包括从一层加载到另一层，某些参数名不匹配的情况，也给出了实验代码和结果，完整实验项目见github。如果对您有所帮助，欢迎关注点赞~

1 state_dict

在PyTorch中，torch.nn.Module的可学习参数（i.e. weights and biases），保存在模型的parameters中，它可以通过model.parameters()进行访问。state_dict是一个从参数名称映射到参数Tensor的字典对象。注意，只有具有可学习参数的层（卷积层、线性层等）和已经注册的缓冲区（bachnorm’s running _mean）才有state_dict中的条目。优化器(optim)也有一个state_dict，其中包含关于优化器状态以及所使用的超参数的信息。由于state_dic对象是Python字典，因此可以轻松地保存、更新、更改和还原它们，从而为PyTorch模型和优化增加了很多模块性。
从训练分类器教程中使用的简单模型看一下state_dict。

# 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

# Initialize model
model = TheModelClass()

# Initialize optimizer
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

# Print model's state_dict
print("Model's state_dict:")
for param_tensor in model.state_dict():
    print(param_tensor, "\t", model.state_dict()[param_tensor].size())

# Print optimizer's state_dict
print("Optimizer's state_dict:")
for var_name in optimizer.state_dict():
    print(var_name, "\t", optimizer.state_dict()[var_name])

# 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

# Initialize model

model = TheModelClass()

# Initialize optimizer

optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

# Print model's state_dict

print("Model's state_dict:")

for param_tensor in model.state_dict():

print(param_tensor, "\t", model.state_dict()[param_tensor].size())

# Print optimizer's state_dict

print("Optimizer's state_dict:")

for var_name in optimizer.state_dict():

print(var_name, "\t", optimizer.state_dict()[var_name])

从输出结果可以看出，每一层的模型参数的名称格式是：层名.参数；如果有它的一层是由另一个类定义的话，那么就把层名往后扩展：层名.层名…参数。下面对上述代码的模型进行重新整理，验证一下。

TheModelClass(
  (conv1): Conv2d(3, 6, kernel_size=(5, 5), stride=(1, 1))
  (pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  (conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
  (fc1): Linear(in_features=400, out_features=120, bias=True)
  (fc2): Linear(in_features=120, out_features=84, bias=True)
  (fc3): Linear(in_features=84, out_features=10, bias=True)
)
Model's state_dict:
conv1.weight 	 torch.Size([6, 3, 5, 5])
conv1.bias 	 torch.Size([6])
conv2.weight 	 torch.Size([16, 6, 5, 5])
conv2.bias 	 torch.Size([16])
fc1.weight 	 torch.Size([120, 400])
fc1.bias 	 torch.Size([120])
fc2.weight 	 torch.Size([84, 120])
fc2.bias 	 torch.Size([84])
fc3.weight 	 torch.Size([10, 84])
fc3.bias 	 torch.Size([10])

TheModelClass(

(conv1): Conv2d(3, 6, kernel_size=(5, 5), stride=(1, 1))

(pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)

(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))

(fc1): Linear(in_features=400, out_features=120, bias=True)

(fc2): Linear(in_features=120, out_features=84, bias=True)

(fc3): Linear(in_features=84, out_features=10, bias=True)

)

Model's state_dict:

conv1.weight torch.Size([6, 3, 5, 5])

conv1.bias torch.Size([6])

conv2.weight torch.Size([16, 6, 5, 5])

conv2.bias torch.Size([16])

fc1.weight torch.Size([120, 400])

fc1.bias torch.Size([120])

fc2.weight torch.Size([84, 120])

fc2.bias torch.Size([84])

fc3.weight torch.Size([10, 84])

fc3.bias torch.Size([10])

模型重新整理的代码与结果：

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.fc_the_model_class = FC()

    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 = self.fc_the_model_class(x)
        return x


class FC(nn.Module):
    def __init__(self):
        super(FC, self).__init__()
        self.fc = nn.Sequential(
            nn.Linear(16 * 5 * 5, 120),
            nn.ReLU(),
            nn.Linear(120, 84),
            nn.ReLU(),
            nn.Linear(84, 10)
        )

    def forward(self, x):
        return self.fc(x)


# Initialize model
model = TheModelClass()

print(model)

# Initialize optimizer
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

# Print model's state_dict
print("Model's state_dict:")
for param_tensor in model.state_dict():
    print(param_tensor, "\t", model.state_dict()[param_tensor].size())

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.fc_the_model_class = FC()

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 = self.fc_the_model_class(x)

return x

class FC(nn.Module):

def __init__(self):

super(FC, self).__init__()

self.fc = nn.Sequential(

nn.Linear(16 * 5 * 5, 120),

nn.ReLU(),

nn.Linear(120, 84),

nn.ReLU(),

nn.Linear(84, 10)

)

def forward(self, x):

return self.fc(x)

# Initialize model

model = TheModelClass()

print(model)

# Initialize optimizer

optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

# Print model's state_dict

print("Model's state_dict:")

for param_tensor in model.state_dict():

print(param_tensor, "\t", model.state_dict()[param_tensor].size())

参数名称输出。整个模型的网络结构还是一样的，但是将全连接层重新使用FC类来定义了。从输出的网络结构可以看出，TheModelClass类中定义时，使用的类的结构层次，在网络的结构中会体现。与(conv2)并列的(fc_the_model_class)是在TheModelClass类定义时用的变量名。后接的FC是fc_the_model_class使用的类名，后面的是这个类中定义的层。输出模型时，就是按一种深度优先的方法遍历了整个模型。对于更深层次层的参数，类名是不会出现在参数名中的，然后将参数名按深度组织：fc_the_model_class.fc.0.weight，也就是在打印过程中，：后面的名称会忽略。

TheModelClass(
  (conv1): Conv2d(3, 6, kernel_size=(5, 5), stride=(1, 1))
  (pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  (conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
  (fc_the_model_class): FC(
    (fc): Sequential(
      (0): Linear(in_features=400, out_features=120, bias=True)
      (1): ReLU()
      (2): Linear(in_features=120, out_features=84, bias=True)
      (3): ReLU()
      (4): Linear(in_features=84, out_features=10, bias=True)
    )
  )
)
Model's state_dict:
conv1.weight 	 torch.Size([6, 3, 5, 5])
conv1.bias 	 torch.Size([6])
conv2.weight 	 torch.Size([16, 6, 5, 5])
conv2.bias 	 torch.Size([16])
fc_the_model_class.fc.0.weight 	 torch.Size([120, 400])
fc_the_model_class.fc.0.bias 	 torch.Size([120])
fc_the_model_class.fc.2.weight 	 torch.Size([84, 120])
fc_the_model_class.fc.2.bias 	 torch.Size([84])
fc_the_model_class.fc.4.weight 	 torch.Size([10, 84])
fc_the_model_class.fc.4.bias 	 torch.Size([10])

TheModelClass(

(conv1): Conv2d(3, 6, kernel_size=(5, 5), stride=(1, 1))

(pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)

(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))

(fc_the_model_class): FC(

(fc): Sequential(

(0): Linear(in_features=400, out_features=120, bias=True)

(1): ReLU()

(2): Linear(in_features=120, out_features=84, bias=True)

(3): ReLU()

(4): Linear(in_features=84, out_features=10, bias=True)

)

Model's state_dict:

conv1.weight torch.Size([6, 3, 5, 5])

conv1.bias torch.Size([6])

conv2.weight torch.Size([16, 6, 5, 5])

conv2.bias torch.Size([16])

fc_the_model_class.fc.0.weight torch.Size([120, 400])

fc_the_model_class.fc.0.bias torch.Size([120])

fc_the_model_class.fc.2.weight torch.Size([84, 120])

fc_the_model_class.fc.2.bias torch.Size([84])

fc_the_model_class.fc.4.weight torch.Size([10, 84])

fc_the_model_class.fc.4.bias torch.Size([10])

2 保存和加载用于推理的模型参数

保存使用：

torch.save(model.state_dict(), PATH)

1	torch.save(model.state_dict(), PATH)

加载使用：

model = TheModelClass(*args, **kwargs)
model.load_state_dict(torch.load(PATH))
model.eval()

model = TheModelClass(*args, **kwargs)

model.load_state_dict(torch.load(PATH))

model.eval()

保存模型用于推理时，仅需要保存训练后的模型的参数，使用torch.save()函数直接保存模型的state_dict，通常文件的后缀名为.pt或.pth。请记住，在运行推理之前，必须先调用model.eval()，将dropout层和batch normalization层设为关闭状态。否则将会产生不一致的推断结果。
需要注意的是，load_state_dict()函数使用的是字典对象，而不是保存对象的路径，所以需要先进行torch.load(PATH)

3 保存和加载整个模型

保存使用：

torch.save(model, PATH)

1	torch.save(model, PATH)

加载使用：

# Model class must be defined somewhere
model = torch.load(PATH)
model.eval()

# Model class must be defined somewhere

model = torch.load(PATH)

model.eval()

4 保存和加载用于推理或者继续训练的general checkpoing

保存使用：

torch.save({
            'epoch': epoch,
            'model_state_dict': model.state_dict(),
            'optimizer_state_dict': optimizer.state_dict(),
            'loss': loss,
            ...
            }, PATH)

torch.save({

'epoch': epoch,

'model_state_dict': model.state_dict(),

'optimizer_state_dict': optimizer.state_dict(),

'loss': loss,

...

}, PATH)

加载使用：

model = TheModelClass(*args, **kwargs)
optimizer = TheOptimizerClass(*args, **kwargs)

checkpoint = torch.load(PATH)
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
epoch = checkpoint['epoch']
loss = checkpoint['loss']

model.eval()
# - or -
model.train()

model = TheModelClass(*args, **kwargs)

optimizer = TheOptimizerClass(*args, **kwargs)

checkpoint = torch.load(PATH)

model.load_state_dict(checkpoint['model_state_dict'])

optimizer.load_state_dict(checkpoint['optimizer_state_dict'])

epoch = checkpoint['epoch']

loss = checkpoint['loss']

model.eval()

# - or -

model.train()

在保存checkpoint用于继续训练时，保存优化器的state_dict是必要的，因为它包含着随着模型训练而更新的缓冲区和参数，可能也需要保存一些其他的项目，包括epoch和loss。常见的PyTorch约定是使用.tar文件扩展名保存这些检查点。

5 将多个模型参数保存在一个文件中

保存使用本质上还是保存的是一个字典对象，PyTorch约定使用.tar保存这些检查点。：

torch.save({
            'modelA_state_dict': modelA.state_dict(),
            'modelB_state_dict': modelB.state_dict(),
            'optimizerA_state_dict': optimizerA.state_dict(),
            'optimizerB_state_dict': optimizerB.state_dict(),
            ...
            }, PATH)

torch.save({

'modelA_state_dict': modelA.state_dict(),

'modelB_state_dict': modelB.state_dict(),

'optimizerA_state_dict': optimizerA.state_dict(),

'optimizerB_state_dict': optimizerB.state_dict(),

...

}, PATH)

加载使用加载还是加载的是字典对象，然后取字典对象。：

modelA = TheModelAClass(*args, **kwargs)
modelB = TheModelBClass(*args, **kwargs)
optimizerA = TheOptimizerAClass(*args, **kwargs)
optimizerB = TheOptimizerBClass(*args, **kwargs)

checkpoint = torch.load(PATH)
modelA.load_state_dict(checkpoint['modelA_state_dict'])
modelB.load_state_dict(checkpoint['modelB_state_dict'])
optimizerA.load_state_dict(checkpoint['optimizerA_state_dict'])
optimizerB.load_state_dict(checkpoint['optimizerB_state_dict'])

modelA.eval()
modelB.eval()
# - or -
modelA.train()
modelB.train()

modelA = TheModelAClass(*args, **kwargs)

modelB = TheModelBClass(*args, **kwargs)

optimizerA = TheOptimizerAClass(*args, **kwargs)

optimizerB = TheOptimizerBClass(*args, **kwargs)

checkpoint = torch.load(PATH)

modelA.load_state_dict(checkpoint['modelA_state_dict'])

modelB.load_state_dict(checkpoint['modelB_state_dict'])

optimizerA.load_state_dict(checkpoint['optimizerA_state_dict'])

optimizerB.load_state_dict(checkpoint['optimizerB_state_dict'])

modelA.eval()

modelB.eval()

# - or -

modelA.train()

modelB.train()

6 使用来自不同模型的参数进行 Warmstarting Model

保存使用：

torch.save(modelA.state_dict(), PATH)

1	torch.save(modelA.state_dict(), PATH)

加载使用：对于不同的模型，设置strict=False是必要的。

modelB = TheModelBClass(*args, **kwargs)
modelB.load_state_dict(torch.load(PATH), strict=False)

1 2	modelB = TheModelBClass(args, *kwargs) modelB.load_state_dict(torch.load(PATH), strict=False)

在迁移学习或者训练新的复杂模型时，部分加载模型或加载部分模型是常见的方案。利用训练过的参数，即使是只有一小部分可以使用，也会对warmstart训练过程有所帮助，而且有望比从头开始训练模型更快地收敛。所谓的warmstart，我理解的就是在参数初始化时，将待训练模型的参数使用已经训练好的模型的部分参数进行初始化，然后接着训练，这种参数初始化方案会大大提高收敛的速度。
无论是从缺少某些键的部分state_dict加载，还是要加载比待加载模型更多的键的state_dic，都可以在lod_state_dict()中将strict参数设置为False，这样可以忽略不匹配的键。
如果要将参数从一层加载到另一层，但是某些键不匹配，只需要加载的state_dict中参数键的名称，来匹配到要加载到的模型中的键。实验代码如下。

targetModel = TheModelClass()

cifar_net = torch.load('./cifar_net.pth')

for item in cifar_net:
    print('cifar_net \t', item, '\t')

targetModel.load_state_dict(cifar_net, strict=False)

for item in targetModel.state_dict():
    print('targetModel \t', item, '\t')

print('cifar_net \t', cifar_net["fc3.bias"], '\t', cifar_net["fc3.bias"].data)
print('targetModel \t', targetModel.state_dict()["fc_the_model_class.fc.4.bias"], '\t', targetModel.state_dict()["fc_the_model_class.fc.4.bias"].data)

# 更新层的名称
cifar_net["fc_the_model_class.fc.0.weight"] = cifar_net.pop("fc1.weight")
cifar_net["fc_the_model_class.fc.0.bias"] = cifar_net.pop("fc1.bias")
cifar_net["fc_the_model_class.fc.2.weight"] = cifar_net.pop("fc2.weight")
cifar_net["fc_the_model_class.fc.2.bias"] = cifar_net.pop("fc2.bias")
cifar_net["fc_the_model_class.fc.4.weight"] = cifar_net.pop("fc3.weight")
cifar_net["fc_the_model_class.fc.4.bias"] = cifar_net.pop("fc3.bias")

targetModel.load_state_dict(cifar_net, strict=False)
print('cifar_net \t', cifar_net["fc_the_model_class.fc.4.bias"], '\t', cifar_net["fc_the_model_class.fc.4.bias"].data)
print('targetModel \t', targetModel.state_dict()["fc_the_model_class.fc.4.bias"], '\t', targetModel.state_dict()["fc_the_model_class.fc.4.bias"].data)

targetModel = TheModelClass()

cifar_net = torch.load('./cifar_net.pth')

for item in cifar_net:

print('cifar_net \t', item, '\t')

targetModel.load_state_dict(cifar_net, strict=False)

for item in targetModel.state_dict():

print('targetModel \t', item, '\t')

print('cifar_net \t', cifar_net["fc3.bias"], '\t', cifar_net["fc3.bias"].data)

print('targetModel \t', targetModel.state_dict()["fc_the_model_class.fc.4.bias"], '\t', targetModel.state_dict()["fc_the_model_class.fc.4.bias"].data)

# 更新层的名称

cifar_net["fc_the_model_class.fc.0.weight"] = cifar_net.pop("fc1.weight")

cifar_net["fc_the_model_class.fc.0.bias"] = cifar_net.pop("fc1.bias")

cifar_net["fc_the_model_class.fc.2.weight"] = cifar_net.pop("fc2.weight")

cifar_net["fc_the_model_class.fc.2.bias"] = cifar_net.pop("fc2.bias")

cifar_net["fc_the_model_class.fc.4.weight"] = cifar_net.pop("fc3.weight")

cifar_net["fc_the_model_class.fc.4.bias"] = cifar_net.pop("fc3.bias")

targetModel.load_state_dict(cifar_net, strict=False)

print('cifar_net \t', cifar_net["fc_the_model_class.fc.4.bias"], '\t', cifar_net["fc_the_model_class.fc.4.bias"].data)

print('targetModel \t', targetModel.state_dict()["fc_the_model_class.fc.4.bias"], '\t', targetModel.state_dict()["fc_the_model_class.fc.4.bias"].data)

输出结果，可以看出fc_the_model_class.fc.4.bias的参数由随机初始化，变成从cifar_net模型中初始化。

cifar_net 	 tensor([-3.1628e-01, -9.6105e-01, -7.1377e-04,  4.6838e-01,  1.1072e+00,
        -2.2960e-01,  1.9044e-01, -5.1352e-02,  1.8365e-01, -3.4669e-01],
       device='cuda:0') 	 tensor([-3.1628e-01, -9.6105e-01, -7.1377e-04,  4.6838e-01,  1.1072e+00,
        -2.2960e-01,  1.9044e-01, -5.1352e-02,  1.8365e-01, -3.4669e-01],
       device='cuda:0')
targetModel 	 tensor([-0.0878, -0.1059, -0.0949,  0.0353,  0.0164, -0.1002, -0.0126, -0.1012,
        -0.0115, -0.1006]) 	 tensor([-0.0878, -0.1059, -0.0949,  0.0353,  0.0164, -0.1002, -0.0126, -0.1012,
        -0.0115, -0.1006])
cifar_net 	 tensor([-3.1628e-01, -9.6105e-01, -7.1377e-04,  4.6838e-01,  1.1072e+00,
        -2.2960e-01,  1.9044e-01, -5.1352e-02,  1.8365e-01, -3.4669e-01],
       device='cuda:0') 	 tensor([-3.1628e-01, -9.6105e-01, -7.1377e-04,  4.6838e-01,  1.1072e+00,
        -2.2960e-01,  1.9044e-01, -5.1352e-02,  1.8365e-01, -3.4669e-01],
       device='cuda:0')
targetModel 	 tensor([-3.1628e-01, -9.6105e-01, -7.1377e-04,  4.6838e-01,  1.1072e+00,
        -2.2960e-01,  1.9044e-01, -5.1352e-02,  1.8365e-01, -3.4669e-01]) 	 tensor([-3.1628e-01, -9.6105e-01, -7.1377e-04,  4.6838e-01,  1.1072e+00,
        -2.2960e-01,  1.9044e-01, -5.1352e-02,  1.8365e-01, -3.4669e-01])