权重pt文件（.pth后缀）如何查看？

小罗碎碎念

我们在查看源码的时候，如果是预训练的模型，或者我们之前说的基础模型，那么就会涉及到一个内容——权重。但是可能有的人没有具体接触过，只是停留在“听说”。

今天另一篇推文介绍了三种UNet的变形，那么这里就以其中的一种作为分析，看一下作者提供的源码，里面就包含了权重文件.pth。

这篇推文先不教你怎么用这个文件，先教你怎么看这个文件的内容，这是以后也一直会用到的，建议收藏！！。

---

一、总代码

这段代码是使用PyTorch库加载一个预训练模型的权重文件（.pth文件），并打印出该模型的相关信息。

我是用Jupyter运行的代码，并导出的md文档，所以在第三部分会看到结果，并且我做了改动，删除了value值（太多了，也没有必要看，了解模型的架构就可以）。

# 导入PyTorch库
import torch

# 定义.pth文件的路径
pthfile = r'./last_ckpt.pt'

# 使用torch.load()函数加载.pth文件中的模型权重，指定在CPU环境下进行
model = torch.load(pthfile, torch.device('cpu'))

# 打印模型类型
print('type:')
print(type(model))

# 打印模型字典的长度
print('length:')
print(len(model))

# 遍历并打印模型字典中的所有键（key）
print('key:')
for k in model.keys():
    print(k)

# 遍历并打印模型字典中的所有值（value）
print('value:')
for k in model:
    print(k, model[k])

这段代码首先导入了PyTorch库，然后指定了一个.pth文件的路径。接着，使用torch.load()函数加载了这个文件，并指定了在CPU环境下进行操作。之后，代码打印出了模型的类型、字典长度、所有的键和对应的值。

需要注意的是，torch.load()函数返回的是一个包含模型状态的字典，而不是模型本身。如果你想要加载这个状态到一个已经定义好的模型实例中，你需要先创建模型实例，然后使用model.load_state_dict()方法来加载状态字典。例如：

# 假设你已经定义了一个模型类MyModel
model_instance = MyModel()

# 加载状态字典到模型实例中
model_instance.load_state_dict(torch.load(pthfile, torch.device('cpu')))

这样，model_instance就包含了从.pth文件中加载的权重和状态。

---

二、逐步输出

import torch

pthfile = r'/Users/luoxiaoluotongxue/Desktop/硕士课题进展记录/项目复现/TransAttUnet-main/pretrained-model.pth'  # .pth文件的路径
model = torch.load(pthfile, torch.device('cpu'))  # 设置在cpu环境下查询

print('type:')
print(type(model))  # 查看模型字典长度
print('length:')
print(len(model))
print('key:')
for k in model.keys():  # 查看模型字典里面的key
    print(k)
print('value:')
for k in model:  # 查看模型字典里面的value
    print(k, model[k])

---

三、输出结果

这段输出展示了从.pth文件加载的模型权重的详细信息。

1. 类型（type）：模型权重以collections.OrderedDict的形式存储，这是一个有序字典，可以记住元素插入的顺序。

2. 长度（length）：模型权重字典包含229个条目。

3. 键（key）：字典中的键是模型中各个层的名称。这些名称通常由层的类型、层内的操作和层的序号组成。例如，inc.double_conv.0.weight表示输入卷积层（inc）的第一个双卷积层（double_conv）的权重。类似地，down1.maxpool_conv.1.double_conv.0.bias表示第一个下采样（down1）的最大池化卷积层（maxpool_conv）的第一个双卷积层的偏置。

4. 值（value）：字典中的值是各个层的权重和偏置张量。这些张量的形状和值取决于层的类型和参数。

从输出中可以看出，这个模型包含多个卷积层、批量归一化层（running_mean和running_var）和全连接层（conv.weight和conv.bias）。此外，模型还包含一些特殊层，如位置嵌入层（pos.row_embed.weight和pos.col_embed.weight）和自注意力机制层（pam.*）。

总之，这段输出展示了从.pth文件加载的模型权重的结构和内容。这些权重可以用于恢复模型的状态，从而在新数据上进行预测或继续训练。

type:
<class 'collections.OrderedDict'>
length:
229
key:
inc.double_conv.0.weight
inc.double_conv.0.bias
inc.double_conv.1.weight
inc.double_conv.1.bias
inc.double_conv.1.running_mean
inc.double_conv.1.running_var
inc.double_conv.1.num_batches_tracked
inc.double_conv.3.weight
inc.double_conv.3.bias
inc.double_conv.4.weight
inc.double_conv.4.bias
inc.double_conv.4.running_mean
inc.double_conv.4.running_var
inc.double_conv.4.num_batches_tracked
down1.maxpool_conv.1.double_conv.0.weight
down1.maxpool_conv.1.double_conv.0.bias
down1.maxpool_conv.1.double_conv.1.weight
down1.maxpool_conv.1.double_conv.1.bias
down1.maxpool_conv.1.double_conv.1.running_mean
down1.maxpool_conv.1.double_conv.1.running_var
down1.maxpool_conv.1.double_conv.1.num_batches_tracked
down1.maxpool_conv.1.double_conv.3.weight
down1.maxpool_conv.1.double_conv.3.bias
down1.maxpool_conv.1.double_conv.4.weight
down1.maxpool_conv.1.double_conv.4.bias
down1.maxpool_conv.1.double_conv.4.running_mean
down1.maxpool_conv.1.double_conv.4.running_var
down1.maxpool_conv.1.double_conv.4.num_batches_tracked
down2.maxpool_conv.1.double_conv.0.weight
down2.maxpool_conv.1.double_conv.0.bias
down2.maxpool_conv.1.double_conv.1.weight
down2.maxpool_conv.1.double_conv.1.bias
down2.maxpool_conv.1.double_conv.1.running_mean
down2.maxpool_conv.1.double_conv.1.running_var
down2.maxpool_conv.1.double_conv.1.num_batches_tracked
down2.maxpool_conv.1.double_conv.3.weight
down2.maxpool_conv.1.double_conv.3.bias
down2.maxpool_conv.1.double_conv.4.weight
down2.maxpool_conv.1.double_conv.4.bias
down2.maxpool_conv.1.double_conv.4.running_mean
down2.maxpool_conv.1.double_conv.4.running_var
down2.maxpool_conv.1.double_conv.4.num_batches_tracked
down3.maxpool_conv.1.double_conv.0.weight
down3.maxpool_conv.1.double_conv.0.bias
down3.maxpool_conv.1.double_conv.1.weight
down3.maxpool_conv.1.double_conv.1.bias
down3.maxpool_conv.1.double_conv.1.running_mean
down3.maxpool_conv.1.double_conv.1.running_var
down3.maxpool_conv.1.double_conv.1.num_batches_tracked
down3.maxpool_conv.1.double_conv.3.weight
down3.maxpool_conv.1.double_conv.3.bias
down3.maxpool_conv.1.double_conv.4.weight
down3.maxpool_conv.1.double_conv.4.bias
down3.maxpool_conv.1.double_conv.4.running_mean
down3.maxpool_conv.1.double_conv.4.running_var
down3.maxpool_conv.1.double_conv.4.num_batches_tracked
down4.maxpool_conv.1.double_conv.0.weight
down4.maxpool_conv.1.double_conv.0.bias
down4.maxpool_conv.1.double_conv.1.weight
down4.maxpool_conv.1.double_conv.1.bias
down4.maxpool_conv.1.double_conv.1.running_mean
down4.maxpool_conv.1.double_conv.1.running_var
down4.maxpool_conv.1.double_conv.1.num_batches_tracked
down4.maxpool_conv.1.double_conv.3.weight
down4.maxpool_conv.1.double_conv.3.bias
down4.maxpool_conv.1.double_conv.4.weight
down4.maxpool_conv.1.double_conv.4.bias
down4.maxpool_conv.1.double_conv.4.running_mean
down4.maxpool_conv.1.double_conv.4.running_var
down4.maxpool_conv.1.double_conv.4.num_batches_tracked
up1.conv.double_conv.0.weight
up1.conv.double_conv.0.bias
up1.conv.double_conv.1.weight
up1.conv.double_conv.1.bias
up1.conv.double_conv.1.running_mean
up1.conv.double_conv.1.running_var
up1.conv.double_conv.1.num_batches_tracked
up1.conv.double_conv.3.weight
up1.conv.double_conv.3.bias
up1.conv.double_conv.4.weight
up1.conv.double_conv.4.bias
up1.conv.double_conv.4.running_mean
up1.conv.double_conv.4.running_var
up1.conv.double_conv.4.num_batches_tracked
up2.conv.double_conv.0.weight
up2.conv.double_conv.0.bias
up2.conv.double_conv.1.weight
up2.conv.double_conv.1.bias
up2.conv.double_conv.1.running_mean
up2.conv.double_conv.1.running_var
up2.conv.double_conv.1.num_batches_tracked
up2.conv.double_conv.3.weight
up2.conv.double_conv.3.bias
up2.conv.double_conv.4.weight
up2.conv.double_conv.4.bias
up2.conv.double_conv.4.running_mean
up2.conv.double_conv.4.running_var
up2.conv.double_conv.4.num_batches_tracked
up3.conv.double_conv.0.weight
up3.conv.double_conv.0.bias
up3.conv.double_conv.1.weight
up3.conv.double_conv.1.bias
up3.conv.double_conv.1.running_mean
up3.conv.double_conv.1.running_var
up3.conv.double_conv.1.num_batches_tracked
up3.conv.double_conv.3.weight
up3.conv.double_conv.3.bias
up3.conv.double_conv.4.weight
up3.conv.double_conv.4.bias
up3.conv.double_conv.4.running_mean
up3.conv.double_conv.4.running_var
up3.conv.double_conv.4.num_batches_tracked
up4.conv.double_conv.0.weight
up4.conv.double_conv.0.bias
up4.conv.double_conv.1.weight
up4.conv.double_conv.1.bias
up4.conv.double_conv.1.running_mean
up4.conv.double_conv.1.running_var
up4.conv.double_conv.1.num_batches_tracked
up4.conv.double_conv.3.weight
up4.conv.double_conv.3.bias
up4.conv.double_conv.4.weight
up4.conv.double_conv.4.bias
up4.conv.double_conv.4.running_mean
up4.conv.double_conv.4.running_var
up4.conv.double_conv.4.num_batches_tracked
outc.conv.weight
outc.conv.bias
pos.row_embed.weight
pos.col_embed.weight
pam.gamma
pam.query_conv.weight
pam.query_conv.bias
pam.key_conv.weight
pam.key_conv.bias
pam.value_conv.weight
pam.value_conv.bias
fuse1.fuse_attn.0.weight
fuse1.fuse_attn.0.bias
fuse1.fuse_attn.1.weight
fuse1.fuse_attn.1.bias
fuse1.fuse_attn.1.running_mean
fuse1.fuse_attn.1.running_var
fuse1.fuse_attn.1.num_batches_tracked
fuse1.fuse_attn.2.weight
fuse1.fuse_attn.3.weight
fuse1.fuse_attn.3.bias
fuse1.fuse_attn.4.weight
fuse1.fuse_attn.4.bias
fuse1.fuse_attn.4.running_mean
fuse1.fuse_attn.4.running_var
fuse1.fuse_attn.4.num_batches_tracked
fuse1.fuse_attn.5.weight
fuse1.fuse_attn.6.weight
fuse1.fuse_attn.6.bias
fuse1.fuse_attn.7.weight
fuse1.fuse_attn.7.bias
fuse1.fuse_attn.7.running_mean
fuse1.fuse_attn.7.running_var
fuse1.fuse_attn.7.num_batches_tracked
fuse2.fuse_attn.0.weight
fuse2.fuse_attn.0.bias
fuse2.fuse_attn.1.weight
fuse2.fuse_attn.1.bias
fuse2.fuse_attn.1.running_mean
fuse2.fuse_attn.1.running_var
fuse2.fuse_attn.1.num_batches_tracked
fuse2.fuse_attn.2.weight
fuse2.fuse_attn.3.weight
fuse2.fuse_attn.3.bias
fuse2.fuse_attn.4.weight
fuse2.fuse_attn.4.bias
fuse2.fuse_attn.4.running_mean
fuse2.fuse_attn.4.running_var
fuse2.fuse_attn.4.num_batches_tracked
fuse2.fuse_attn.5.weight
fuse2.fuse_attn.6.weight
fuse2.fuse_attn.6.bias
fuse2.fuse_attn.7.weight
fuse2.fuse_attn.7.bias
fuse2.fuse_attn.7.running_mean
fuse2.fuse_attn.7.running_var
fuse2.fuse_attn.7.num_batches_tracked
fuse3.fuse_attn.0.weight
fuse3.fuse_attn.0.bias
fuse3.fuse_attn.1.weight
fuse3.fuse_attn.1.bias
fuse3.fuse_attn.1.running_mean
fuse3.fuse_attn.1.running_var
fuse3.fuse_attn.1.num_batches_tracked
fuse3.fuse_attn.2.weight
fuse3.fuse_attn.3.weight
fuse3.fuse_attn.3.bias
fuse3.fuse_attn.4.weight
fuse3.fuse_attn.4.bias
fuse3.fuse_attn.4.running_mean
fuse3.fuse_attn.4.running_var
fuse3.fuse_attn.4.num_batches_tracked
fuse3.fuse_attn.5.weight
fuse3.fuse_attn.6.weight
fuse3.fuse_attn.6.bias
fuse3.fuse_attn.7.weight
fuse3.fuse_attn.7.bias
fuse3.fuse_attn.7.running_mean
fuse3.fuse_attn.7.running_var
fuse3.fuse_attn.7.num_batches_tracked
fuse4.fuse_attn.0.weight
fuse4.fuse_attn.0.bias
fuse4.fuse_attn.1.weight
fuse4.fuse_attn.1.bias
fuse4.fuse_attn.1.running_mean
fuse4.fuse_attn.1.running_var
fuse4.fuse_attn.1.num_batches_tracked
fuse4.fuse_attn.2.weight
fuse4.fuse_attn.3.weight
fuse4.fuse_attn.3.bias
fuse4.fuse_attn.4.weight
fuse4.fuse_attn.4.bias
fuse4.fuse_attn.4.running_mean
fuse4.fuse_attn.4.running_var
fuse4.fuse_attn.4.num_batches_tracked
fuse4.fuse_attn.5.weight
fuse4.fuse_attn.6.weight
fuse4.fuse_attn.6.bias
fuse4.fuse_attn.7.weight
fuse4.fuse_attn.7.bias
fuse4.fuse_attn.7.running_mean
fuse4.fuse_attn.7.running_var
fuse4.fuse_attn.7.num_batches_tracked