模型的使用及修改、保存与加载
以图像处理中torchvision为例，PyTorch通过torchvision.models模块提供了更多的预训练模型.
在图像分类当中，包括许多模型
import torchvision
import warnings
import torch
warnings.filterwarnings("ignore")
本文将以VGG16为例，展示Pytorch对现有模型的使用及修改的具体操作
VGG16是一个经典的卷积神经网络模型，由牛津大学计算机视觉组（Visual Geometry Group）提出，用于参加2014年的ImageNet图像分类。
VGG 最大的特点就是通过比较彻底地采用 3x3 尺寸的卷积核来堆叠神经网络，这样也加深整个神经网络的深度。这两个重要的改变对于人们重新定义卷积神经网络模型架构也有不小的帮助，至少证明使用更小的卷积核并且增加卷积神经网络的深度，可以更有效地提升模型的性能。
torchvision.models.vgg16(*, weights: Optional[VGG16_Weights] = None, progress: bool = True, **kwargs: Any)
- weights(可选)：指定要加载的预训练权重。可以是None（默认值）表示不加载预训练权重，或是指定为预定义的某个预训练权重标识符。
- progress:指示下载进度条的显示设置，默认为True显示下载进度条。
- kwargs:其它可选参数，传递给VGG-16模型的基类torchvision.models.VGG
vgg16 = torchvision.models.vgg16(weights=True,progress=True)
print(vgg16)
VGG(
(features): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace=True)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace=True)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(6): ReLU(inplace=True)
(7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(8): ReLU(inplace=True)
(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(11): ReLU(inplace=True)
(12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(13): ReLU(inplace=True)
(14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(15): ReLU(inplace=True)
(16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(18): ReLU(inplace=True)
(19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(20): ReLU(inplace=True)
(21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(22): ReLU(inplace=True)
(23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(25): ReLU(inplace=True)
(26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(27): ReLU(inplace=True)
(28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(29): ReLU(inplace=True)
(30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(avgpool): AdaptiveAvgPool2d(output_size=(7, 7))
(classifier): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace=True)
(2): Dropout(p=0.5, inplace=False)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace=True)
(5): Dropout(p=0.5, inplace=False)
(6): Linear(in_features=4096, out_features=1000, bias=True)
)
)
从上述运行结果可知：VGG16网络是由13层卷积层和3层全连接层组成，最后网络输出一共有1000个分类结果。
修改VGG16模型：
-
以CIFAR10为例
-
使用add_module()方法在VGG16模型后增加一个线性层，实现将VGG16的1000个类别输出为类似CIFAR10的10个类别，代码如下：
import torchvision.models as models
from torch import nn
vgg16 = torchvision.models.vgg16(weights=True,progress=True)
vgg16.add_module("add_linear", nn.Linear(1000, 10))
print(vgg16)
VGG(
(features): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace=True)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace=True)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(6): ReLU(inplace=True)
(7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(8): ReLU(inplace=True)
(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(11): ReLU(inplace=True)
(12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(13): ReLU(inplace=True)
(14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(15): ReLU(inplace=True)
(16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(18): ReLU(inplace=True)
(19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(20): ReLU(inplace=True)
(21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(22): ReLU(inplace=True)
(23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(25): ReLU(inplace=True)
(26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(27): ReLU(inplace=True)
(28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(29): ReLU(inplace=True)
(30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(avgpool): AdaptiveAvgPool2d(output_size=(7, 7))
(classifier): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace=True)
(2): Dropout(p=0.5, inplace=False)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace=True)
(5): Dropout(p=0.5, inplace=False)
(6): Linear(in_features=4096, out_features=1000, bias=True)
)
(add_linear): Linear(in_features=1000, out_features=10, bias=True)
)
由上述可以知道，add_linear是在classifier外面的，如果要在classifier里面，可以将
vgg16.add_module("add_linear", nn.Linear(1000, 10))
替换为
vgg16.classifier.add_module("add_linear", nn.Linear(1000, 10))
import torchvision.models as models
from torch import nn
vgg16 = torchvision.models.vgg16(weights=True,progress=True)
vgg16.classifier.add_module("add_linear", nn.Linear(1000, 10))
print(vgg16)
VGG(
(features): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace=True)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace=True)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(6): ReLU(inplace=True)
(7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(8): ReLU(inplace=True)
(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(11): ReLU(inplace=True)
(12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(13): ReLU(inplace=True)
(14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(15): ReLU(inplace=True)
(16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(18): ReLU(inplace=True)
(19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(20): ReLU(inplace=True)
(21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(22): ReLU(inplace=True)
(23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(25): ReLU(inplace=True)
(26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(27): ReLU(inplace=True)
(28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(29): ReLU(inplace=True)
(30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(avgpool): AdaptiveAvgPool2d(output_size=(7, 7))
(classifier): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace=True)
(2): Dropout(p=0.5, inplace=False)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace=True)
(5): Dropout(p=0.5, inplace=False)
(6): Linear(in_features=4096, out_features=1000, bias=True)
(add_linear): Linear(in_features=1000, out_features=10, bias=True)
)
)
也可以直接进行修改,例如对classifier中的 6): Linear(in_features=4096, out_features=1000, bias=True)直接修改为out_features=10
vgg16.classifier[6] = nn.Linear(in_features=4096,out_features=10,bias=True)
网络模型的保存与读取
模型的保存
torch.save(obj, f, pickle_protocol=DEFAULT_PROTOCOL)
| 参数 | 描述 |
|---|---|
| obj: | 要保存的对象，可以是模型、张量、字典等 |
| f: | 要保存到的文件路径或文件对象 |
| pickle_protocol: | 序列化协议的版本，默认为DEFAULT_PROTOCOL |
方法一：保存整个模型，包括其相关的所有参数，利用
torch.save()
import torchvision
vgg16 = torchvision.models.vgg16(weights=True, progress=True)
torch.save(vgg16, "vgg16_model_true.pth")
方法二：只保存模型参数，在原有vgg16对象中使用.state_dict()方法即可。
import torchvision
vgg16 = torchvision.models.vgg16(weights=True,progress=True)
torch.save(vgg16.state_dict(), "vgg16_model_true_2.pth")
运行成功后，对应的文件：vgg16_model_true.pth和 vgg16_model_true_2.pth会保存在默认路径当中
网络模型的加载
torch.load(f, map_location=None, pickle_module=pickle, *, weights_only=False, mmap=None, **pickle_load_args)
| 参数 | 描述 |
|---|---|
| f | 要加载的文件路径或文件对象 |
| map_location | 可选参数，用于指定在哪个设备上加载模型。如果不提供该参数，默认会加载到当前设备 |
| pickle_module | 可选参数，用于指定用于反序列化的模块。默认为pickle |
| pickle_load_args | 其他可选的用于反序列化的参数 |
import torch
import torchvision.models as models
from torch import nn
# 因为vgg16_model_true.pth是使用方法一保存的，故输出后是整个模型网络结构
model1 = torch.load("vgg16_model_true.pth")
print(model1)
VGG(
(features): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace=True)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace=True)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(6): ReLU(inplace=True)
(7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(8): ReLU(inplace=True)
(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(11): ReLU(inplace=True)
(12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(13): ReLU(inplace=True)
(14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(15): ReLU(inplace=True)
(16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(18): ReLU(inplace=True)
(19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(20): ReLU(inplace=True)
(21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(22): ReLU(inplace=True)
(23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(25): ReLU(inplace=True)
(26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(27): ReLU(inplace=True)
(28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(29): ReLU(inplace=True)
(30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(avgpool): AdaptiveAvgPool2d(output_size=(7, 7))
(classifier): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace=True)
(2): Dropout(p=0.5, inplace=False)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace=True)
(5): Dropout(p=0.5, inplace=False)
(6): Linear(in_features=4096, out_features=1000, bias=True)
)
)
# 因为vgg16_model_true_2.pth是使用方法二保存的，只保留模型参数，故输出后是整个字典类型
model2 = torch.load("vgg16_model_true_2.pth")
print(model2)
OrderedDict([('features.0.weight', tensor([[[[-5.5373e-01, 1.4270e-01, 5.2896e-01],
[-5.8312e-01, 3.5655e-01, 7.6566e-01],
[-6.9022e-01, -4.8019e-02, 4.8409e-01]],
[[ 1.7548e-01, 9.8630e-03, -8.1413e-02],
[ 4.4089e-02, -7.0323e-02, -2.6035e-01],
[ 1.3239e-01, -1.7279e-01, -1.3226e-01]],
[[ 3.1303e-01, -1.6591e-01, -4.2752e-01],
[ 4.7519e-01, -8.2677e-02, -4.8700e-01],
[ 6.3203e-01, 1.9308e-02, -2.7753e-01]]],
[[[ 2.3254e-01, 1.2666e-01, 1.8605e-01],
[-4.2805e-01, -2.4349e-01, 2.4628e-01],
[-2.5066e-01, 1.4177e-01, -5.4864e-03]],
[[-1.4076e-01, -2.1903e-01, 1.5041e-01],
[-8.4127e-01, -3.5176e-01, 5.6398e-01],
[-2.4194e-01, 5.1928e-01, 5.3915e-01]],
.............省略了........
针对只保存了模型参数的第二种情况，使其显示完整的模型结构，利用
.load_state_dict()
import torch
import torchvision.models as models
from torch import nn
vgg16 = models.vgg16(weights=False)
vgg16.load_state_dict(torch.load("vgg16_model_true_2.pth")) # 针对第二种加载参数的情况，使其显示完整的网络结构
print(vgg16_true)
VGG(
(features): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace=True)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace=True)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(6): ReLU(inplace=True)
(7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(8): ReLU(inplace=True)
(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(11): ReLU(inplace=True)
(12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(13): ReLU(inplace=True)
(14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(15): ReLU(inplace=True)
(16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(18): ReLU(inplace=True)
(19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(20): ReLU(inplace=True)
(21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(22): ReLU(inplace=True)
(23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(25): ReLU(inplace=True)
(26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(27): ReLU(inplace=True)
(28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(29): ReLU(inplace=True)
(30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(avgpool): AdaptiveAvgPool2d(output_size=(7, 7))
(classifier): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace=True)
(2): Dropout(p=0.5, inplace=False)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace=True)
(5): Dropout(p=0.5, inplace=False)
(6): Linear(in_features=4096, out_features=1000, bias=True)
)
)
注意： 加载模型时，要确保当前代码中使用的模型类与之前保存的模型类相同。
总结
-
torch.load()是PyTorch中用于加载保存的对象的函数，可以加载之前使用 -
torch.save()保存的模型、张量、字典等。可以指定要加载的文件路径或文件对象，并可选地指定加载到的设备、反序列化模块等参数。
