import torch
from torch import nn
from d2l import torch as d2l
batch_size = 256 # 保持批量大小为 256
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size) # 仍使用Fashion-MNIST数据集
net = nn.Sequential(nn.Flatten(), nn.Linear(784, 10)) # flatten负责调整网络输入形状，添加一个有 10 个输出的全连接层
def init_weights(m):
if type(m) == nn.Linear:
nn.init.normal_(m.weight, std=0.01) # 使用正态分布中的随机值初始化权重
net.apply(init_weights)
Sequential(
(0): Flatten(start_dim=1, end_dim=-1)
(1): Linear(in_features=784, out_features=10, bias=True)
)
从计算的角度来讲，指数可能会造成数值稳定性问题，即可能会发生溢出。解决上溢问题可在 softmax 运算前先从所有
o
k
o_k
ok 中减去
max
(
o
k
)
\max(o_k)
max(ok)。但如若有些
o
j
−
max
(
o
k
)
o_j-\max(o_k)
oj−max(ok) 为过小的负值时则又可能发生下溢，此时可以尽量避免计算
exp
(
o
j
−
max
(
o
k
)
)
\exp(o_j-\max(o_k))
exp(oj−max(ok))：
log
(
y
^
i
)
=
log
(
exp
(
o
j
−
max
(
o
k
)
)
∑
k
exp
(
o
k
−
max
(
o
k
)
)
)
=
log
(
exp
(
o
j
−
max
(
o
k
)
)
)
−
log
(
∑
k
exp
(
o
k
−
max
(
o
k
)
)
)
=
o
j
−
max
(
o
k
)
−
log
(
∑
k
exp
(
o
k
−
max
(
o
k
)
)
)
\begin{align} \log(\hat{y}_i)&=\log\left(\frac{\exp(o_j-\max(o_k))}{\sum_k\exp(o_k-\max(o_k))}\right)\\ &=\log(\exp(o_j-\max(o_k)))-\log\left(\sum_k\exp(o_k-\max(o_k))\right)\\ &=o_j-\max(o_k)-\log\left(\sum_k\exp(o_k-\max(o_k))\right) \end{align}
log(y^i)=log(∑kexp(ok−max(ok))exp(oj−max(ok)))=log(exp(oj−max(ok)))−log(k∑exp(ok−max(ok)))=oj−max(ok)−log(k∑exp(ok−max(ok)))
loss = nn.CrossEntropyLoss(reduction='none')
trainer = torch.optim.SGD(net.parameters(), lr=0.1) # 使用学习率为 0.1 的小批量随机梯度下降作为优化算法
num_epochs = 10
d2l.train_ch3(net, train_iter, test_iter, loss, num_epochs, trainer) # 调用上节定义的训练函数来训练模型
（1）尝试调增超参数，例如批量大小、轮数和学习率，并查看结果。
batch_size2 = 1024 # 将 batch_size 提高到 1024
train_iter2, test_iter2 = d2l.load_data_fashion_mnist(batch_size2)
net2 = nn.Sequential(nn.Flatten(),nn.Linear(784,10))
net2.apply(init_weights)
num_epochs = 10
trainer = torch.optim.SGD(net2.parameters(), lr=0.1)
d2l.train_ch3(net2, train_iter2, test_iter2, loss, num_epochs, trainer) # 提高 batch size 会使 train loss 提高
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
Cell In[6], line 8
6 num_epochs = 10
7 trainer = torch.optim.SGD(net2.parameters(), lr=0.1)
----> 8 d2l.train_ch3(net2, train_iter2, test_iter2, loss, num_epochs, trainer)
File c:\Software\Miniconda3\envs\d2l\lib\site-packages\d2l\torch.py:340, in train_ch3(net, train_iter, test_iter, loss, num_epochs, updater)
338 animator.add(epoch + 1, train_metrics + (test_acc,))
339 train_loss, train_acc = train_metrics
--> 340 assert train_loss < 0.5, train_loss
341 assert train_acc <= 1 and train_acc > 0.7, train_acc
342 assert test_acc <= 1 and test_acc > 0.7, test_acc
AssertionError: 0.5225925379435221
batch_size3 = 256
train_iter3, test_iter3 = d2l.load_data_fashion_mnist(batch_size3)
net3 = nn.Sequential(nn.Flatten(),nn.Linear(784,10))
net3.apply(init_weights)
num_epochs = 40 # 将轮数提高到 40
trainer = torch.optim.SGD(net3.parameters(), lr=0.1)
d2l.train_ch3(net3, train_iter3, test_iter3, loss, num_epochs, trainer) # 提高轮数会使 test acc 突然下降，应该是过拟合了。
batch_size4 = 256
train_iter4, test_iter4 = d2l.load_data_fashion_mnist(batch_size4)
net4 = nn.Sequential(nn.Flatten(),nn.Linear(784,10))
net4.apply(init_weights)
num_epochs = 10
trainer = torch.optim.SGD(net4.parameters(), lr=0.4) # 将学习率提高到 0.4
d2l.train_ch3(net4, train_iter4, test_iter4, loss, num_epochs, trainer) # 提高学习率会使 test acc 极不稳定，train loss 无法收敛。
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
Cell In[9], line 8
6 num_epochs = 10
7 trainer = torch.optim.SGD(net4.parameters(), lr=0.4) # 将学习率提高到 0.4
----> 8 d2l.train_ch3(net4, train_iter4, test_iter4, loss, num_epochs, trainer)
File c:\Software\Miniconda3\envs\d2l\lib\site-packages\d2l\torch.py:340, in train_ch3(net, train_iter, test_iter, loss, num_epochs, updater)
338 animator.add(epoch + 1, train_metrics + (test_acc,))
339 train_loss, train_acc = train_metrics
--> 340 assert train_loss < 0.5, train_loss
341 assert train_acc <= 1 and train_acc > 0.7, train_acc
342 assert test_acc <= 1 and test_acc > 0.7, test_acc
AssertionError: 0.6684705724080404
（2）增加轮数，为什么测试精度会在一段时间后降低？我们如何解决这个问题？
因为发生了过拟合。可以增加样本数。