python
复制代码
import torch
import numpy as np
import torch.utils.data as Data
import torchvision.datasets
import torchvision.transforms as transforms
import sys
sys.path.append("路径")
import d2lzh_pytorch as d2l
'''
--------------------------------------------------获取和读取数据
'''
batch_size = 256
train_mnist = torchvision.datasets.FashionMNIST(root='路径',
download=True, train=True, transform=transforms.ToTensor())
test_mnist = torchvision.datasets.FashionMNIST(root='路径',
download=True, train=False, transform=transforms.ToTensor())
train_iter = Data.DataLoader(train_mnist, batch_size=batch_size, shuffle=True)
test_iter = Data.DataLoader(test_mnist, batch_size=batch_size, shuffle=False)
'''
--------------------------------------------------定义模型参数
'''
num_inputs = 784
num_outputs = 10
num_hidden = 256
# 有几个隐藏层就要设置几个参数，简洁实现中，linear网络会自动配置初始参数，自己可以使用init.normal_()设置参数初始值
w1 = torch.tensor(np.random.normal(0, 0.01, (num_inputs, num_hidden)), dtype=torch.float)
b1 = torch.zeros(num_hidden, dtype=torch.float)
w2 = torch.tensor(np.random.normal(0, 0.1, (num_hidden, num_outputs)), dtype=torch.float)
b2 = torch.zeros(num_outputs, dtype=torch.float)
params = [w1, b1, w2, b2]
for param in params:
param.requires_grad_(requires_grad=True)
'''
---------------------------------------------定义激活函数
'''
def relu(X):
return torch.max(input=X, other=torch.tensor(0.0))
'''
---------------------------------------------------定义模型
'''
# 使用view函数将输入的样本转换成inputs特征数大小的图像
def net(X):
X = X.view((-1, num_inputs))
H = relu(torch.matmul(X, w1) + b1) # torch.mm(X, w1) + b1得到隐藏层输出
# 对隐藏层变量进行激活函数变换，然后作为下一个全连接层的输入
# 第一层不是隐藏层，直接线性计算，隐藏层输出作为输出层输入的时候，对隐藏层进行非线性变换，然后传入输入层
return torch.matmul(H, w2) + b2 # 隐藏层作为输出层的输入 n层layer有最多n-2个激活函数
'''
-----------------------------------------------------定义损失函数
'''
loss = torch.nn.CrossEntropyLoss() # 包含了softmax运算和交叉熵运算
'''
------------------------------------------------------softmax操作，用于训练模型中训练集准确率调用
'''
def softmax(X):
X_exp = X.exp() # 幂指数化
partition = X_exp.sum(dim=1, keepdim=True) # 求和每行的元素值
return X_exp / partition # 做比值得预测概率
'''
----------------------------------------------------测试集准确率函数，训练模型中测试集准确率调用
'''
def evaluate_accuracy(test_data):
acc_num, num = 0.0, 0
for X, y in test_data: # X,y分别是一个元组
acc_num += (softmax(net(X)).argmax(dim=1) == y).float().sum().item()
num += y.shape[0]
return acc_num / num
'''
------------------------------------------------------训练模型
'''
num_epochs, lr = 5, 100
def train():
for epoch in range(num_epochs):
train_acc, train_l, test_acc, n, num = 0.0, 0.0, 0.0, 0, 0
for X, y in train_iter: #
l = loss(net(X), y) # CrossEntropyLoss 函数已经是对一个批次内所有样本的平均损失计算了
if params[0].grad is not None: # 第一次训练迭代前是没有梯度产生的，梯度是由于反向传播才产生的
for param in params: # 参数梯度清零
param.grad.data.zero_()
l.backward() # 反向传播
d2l.sgd(params, lr, batch_size) # 梯度下降操作
train_l += l.item()
# net(X)返回每个样本各个类别的预测值，有n个样本返回
train_acc += (softmax(net(X)).argmax(dim=1) == y).float().sum().item() # 累加预测正确个数
n += y.shape[0]
num += 1
test_acc = evaluate_accuracy(test_iter)
print(f'epoch %d, loss %.4f, train_acc %.3f, test_acc %.3f'
% (epoch + 1, train_l / num, train_acc / n, test_acc))
train()