pytorch学习笔记（7）--模型保存与导入

系列文章

pytorch学习笔记（1）–QUICKSTART
pytorch学习笔记（2）–Tensor
pytorch学习笔记（3）–数据集与数据导入
pytorch学习笔记（4）–创建模型(Build Model)
pytorch学习笔记（5）–Autograd
pytorch学习笔记（6）–Optimization

---

---

首先导入需要的库

import torch
import torchvision.models as models

一、保存和导入模型权重

1.torch.save(model.state_dict(), ‘model_weights.pth’))

PyTorch models将学习到的参数存储在内部状态字典中，称为 state_dict。这些可以通过 torch.save 方法保存：

model = models.vgg16(weights='IMAGENET1K_V1')
torch.save(model.state_dict(), 'model_weights.pth')

输出：

Downloading: "https://download.pytorch.org/models/vgg16-397923af.pth" to /var/lib/jenkins/.cache/torch/hub/checkpoints/vgg16-397923af.pth

  0%|          | 0.00/528M [00:00<?, ?B/s]
  2%|2         | 11.9M/528M [00:00<00:04, 125MB/s]
  5%|4         | 24.3M/528M [00:00<00:04, 128MB/s]
  7%|6         | 36.5M/528M [00:00<00:04, 128MB/s]
  9%|9         | 48.9M/528M [00:00<00:03, 129MB/s]
 12%|#1        | 61.2M/528M [00:00<00:03, 128MB/s]
 14%|#3        | 73.6M/528M [00:00<00:03, 129MB/s]
 16%|#6        | 85.9M/528M [00:00<00:03, 129MB/s]
 19%|#8        | 98.2M/528M [00:00<00:03, 129MB/s]
 21%|##        | 110M/528M [00:00<00:03, 129MB/s]
 23%|##3       | 123M/528M [00:01<00:03, 129MB/s]
 26%|##5       | 135M/528M [00:01<00:03, 129MB/s]
 28%|##7       | 147M/528M [00:01<00:03, 129MB/s]
 30%|###       | 160M/528M [00:01<00:02, 129MB/s]
 33%|###2      | 172M/528M [00:01<00:02, 129MB/s]
 35%|###4      | 184M/528M [00:01<00:02, 129MB/s]
 37%|###7      | 197M/528M [00:01<00:02, 129MB/s]
 40%|###9      | 209M/528M [00:01<00:02, 129MB/s]
 42%|####1     | 221M/528M [00:01<00:02, 129MB/s]
 44%|####4     | 234M/528M [00:01<00:02, 129MB/s]
 47%|####6     | 246M/528M [00:02<00:02, 129MB/s]
 49%|####8     | 258M/528M [00:02<00:02, 129MB/s]
 51%|#####1    | 270M/528M [00:02<00:02, 129MB/s]
 54%|#####3    | 283M/528M [00:02<00:01, 129MB/s]
 56%|#####5    | 295M/528M [00:02<00:01, 129MB/s]
 58%|#####8    | 308M/528M [00:02<00:01, 129MB/s]
 61%|######    | 320M/528M [00:02<00:01, 130MB/s]
 63%|######2   | 332M/528M [00:02<00:01, 128MB/s]
 65%|######5   | 345M/528M [00:02<00:01, 129MB/s]
 68%|######7   | 357M/528M [00:02<00:01, 128MB/s]
 70%|######9   | 369M/528M [00:03<00:01, 129MB/s]
 72%|#######2  | 382M/528M [00:03<00:01, 129MB/s]
 75%|#######4  | 394M/528M [00:03<00:01, 129MB/s]
 77%|#######7  | 407M/528M [00:03<00:00, 130MB/s]
 79%|#######9  | 419M/528M [00:03<00:00, 130MB/s]
 82%|########1 | 432M/528M [00:03<00:00, 130MB/s]
 84%|########4 | 444M/528M [00:03<00:00, 130MB/s]
 86%|########6 | 456M/528M [00:03<00:00, 130MB/s]
 89%|########8 | 469M/528M [00:03<00:00, 131MB/s]
 91%|#########1| 482M/528M [00:03<00:00, 131MB/s]
 94%|#########3| 494M/528M [00:04<00:00, 131MB/s]
 96%|#########5| 506M/528M [00:04<00:00, 131MB/s]
 98%|#########8| 519M/528M [00:04<00:00, 131MB/s]
100%|##########| 528M/528M [00:04<00:00, 129MB/s]

2. load_state_dict(torch.load(‘model_weights.pth’))

要加载模型权重，您需要首先创建同一模型的实例，然后使用 load_state_dict() 方法加载参数。

model = models.vgg16() # we do not specify ``weights``, i.e. create untrained model
model.load_state_dict(torch.load('model_weights.pth'))
model.eval()

输出：

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)
  )
)

一定要在推理之前调用 model.eval() 方法，将 dropout 和批量归一化层设置为评估模式。如果不这样做将会导致推理结果不一致。

二、保存和导入模型的架构

加载模型权重时，我们需要首先实例化模型类，因为该类定义了网络的结构。们可能希望将此类的结构与模型一起保存，在这种情况下，我们可以将 model （而不是 model.state_dict()）传递给保存函数：

1. torch.save(model, ‘model.pth’)

导入时可以使用

2. model = torch.load(‘model.pth’)