5.5.1 Loading and saving
import torch
from torch import nn
from torch.nn import functional as F
x = torch.arange(4)
torch.save(x, 'x-file') # 使用 save 保存
x2 = torch.load('x-file') # 使用 load 读回内存
x2
tensor([0, 1, 2, 3])
y = torch.zeros(4)
torch.save([x, y],'x-files') # 也可以存储张量列表
x2, y2 = torch.load('x-files')
(x2, y2)
(tensor([0, 1, 2, 3]), tensor([0., 0., 0., 0.]))
mydict = {
'x': x, 'y': y} # 存储从字符串映射到张量的字典
torch.save(mydict, 'mydict')
mydict2 = torch.load('mydict')
mydict2
{'x': tensor([0, 1, 2, 3]), 'y': tensor([0., 0., 0., 0.])}
5.5.2 Loading and saving model parameters
class MLP(nn.Module):
def __init__(self):
super().__init__()
self.hidden = nn.Linear(20, 256)
self.output = nn.Linear(256, 10)
def forward(self, x):
return self.output(F.relu(self.hidden(x)))
net = MLP()
X = torch.randn(size=(2, 20))
Y = net(X)
torch.save(net.state_dict(), 'mlp.params') # 保存模型参数
clone = MLP()
clone.load_state_dict(torch.load('mlp.params')) # 加载文件中存储的参数
Y_clone = clone(X) # 参数一致则计算结果也应相同
clone.eval(), Y_clone == Y
(MLP(
(hidden): Linear(in_features=20, out_features=256, bias=True)
(output): Linear(in_features=256, out_features=10, bias=True)
),
tensor([[True, True, True, True, True, True, True, True, True, True],
[True, True, True, True, True, True, True, True, True, True]]))
practise
(1) Even if there is no need to deploy the trained model to different devices, what practical benefit does the saved model parameters have?
It can be used as a backup or for further processing.
(2) Suppose we only want to reuse a part of the network and have merged it into a different network architecture. For example, if you want to use the first two layers of the previous network in a new network, what should you do?
torch.save(net.hidden.state_dict(), 'mlp.hidden.params') # 需要哪里存哪里
clone = MLP()
clone.hidden.load_state_dict(torch.load('mlp.hidden.params')) # 需要哪里加载哪里
clone.eval(), clone.hidden.weight == net.hidden.weight
(MLP(
(hidden): Linear(in_features=20, out_features=256, bias=True)
(output): Linear(in_features=256, out_features=10, bias=True)
),
tensor([[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
...,
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True]]))
(3) How to save network architecture and parameters at the same time? What constraints need to be placed on the architecture?
net = nn.Sequential(nn.Linear(20, 256), nn.ReLU(), nn.Linear(256, 10))
torch.save(net, 'net') # pytorch 本身就支持保存模型
net_new = torch.load('net')
net_new
Sequential(
(0): Linear(in_features=20, out_features=256, bias=True)
(1): ReLU()
(2): Linear(in_features=256, out_features=10, bias=True)
)