[人工智能]PyTorch学习笔记

如何使用PyTorch定义神经网络?

首先是模型层要用到的一些基本函数

#模型层

input_image = torch.rand(3,28,28) #随机生成三个28*28的为输入数据
print(input_image.size())

#nn.Flatten层
#将28*28的数据转化为784的一维数据
flatten = nn.Flatten()#要先初始化flatten层 
flat_image = flatten(input_image)#调用nn.Flatten层

print(flat_image.size())

#nn.Linear层
layer1 = nn.Linear(in_features=28*28, out_features=20)# 要先初始化Linear层
hidden1 = layer1(flat_image)#调用nn.linear层 给了输入，输出默认为上面设置的20
print(hidden1.size())

#nn.relu层
print(f"before relu:{hidden1}\n\n")
hidden1 = nn.ReLU()(hidden1)
print(f"after relu :{hidden1}")

#nn.Sequential  以定义的顺序通过所有的模块 可以使用序列容器快速的创建一个网络

seq_modules = nn.Sequential(
    flatten,
    layer1,
    nn.ReLU(),
    nn.Linear(20,10)
)
input_image = torch.rand(3,28,28)
logits = seq_modules(input_image)

#nn.Softmax 神经网络的最后一个线性层返回logits 传给softmax后 数据被缩放到0-1之间代表每个类别的预测概率,dim参数表示数值必须和为1的维度
softmax = nn.Softmax(dim=1)
pred_probab=softmax(logits)

输出结果：

torch.Size([3, 28, 28])
torch.Size([3, 784])
torch.Size([3, 20])
before relu:tensor([[-0.1643, -0.6735, -0.1328, -0.4160,  0.0519,  0.0962,  0.5110, -0.0060,
          0.4754, -0.1763, -0.1053, -0.1048,  0.3832,  0.4390,  0.2825,  0.1255,
          0.0655, -0.0166,  0.2684,  0.2844],
        [-0.0868, -0.3997,  0.0816, -0.1037,  0.2405, -0.4162,  0.2308, -0.1436,
          0.8886, -0.2782, -0.0973, -0.1159,  0.1795, -0.0238,  0.2312,  0.2655,
          0.8309, -0.2080,  0.1703,  0.6288],
        [-0.3683, -0.4862, -0.1711, -0.1568,  0.2857,  0.3082,  0.1831,  0.2739,
          0.6198, -0.1501,  0.2008, -0.5275,  0.4669,  0.1893,  0.1369,  0.3056,
          0.7309,  0.1656,  0.0215,  0.2588]], grad_fn=<AddmmBackward>)


after relu :tensor([[0.0000, 0.0000, 0.0000, 0.0000, 0.0519, 0.0962, 0.5110, 0.0000, 0.4754,
         0.0000, 0.0000, 0.0000, 0.3832, 0.4390, 0.2825, 0.1255, 0.0655, 0.0000,
         0.2684, 0.2844],
        [0.0000, 0.0000, 0.0816, 0.0000, 0.2405, 0.0000, 0.2308, 0.0000, 0.8886,
         0.0000, 0.0000, 0.0000, 0.1795, 0.0000, 0.2312, 0.2655, 0.8309, 0.0000,
         0.1703, 0.6288],
        [0.0000, 0.0000, 0.0000, 0.0000, 0.2857, 0.3082, 0.1831, 0.2739, 0.6198,
         0.0000, 0.2008, 0.0000, 0.4669, 0.1893, 0.1369, 0.3056, 0.7309, 0.1656,
         0.0215, 0.2588]], grad_fn=<ReluBackward0>)

使用torch.autograd自动计算梯度，举例：

#torch.autograd自动计算梯度
x = torch.ones(5)#输入tensor
y = torch.zeros(3)#期望的输出
#requires_grad表示是否需要求导，默认为不求导
w = torch.randn(5,3,requires_grad = True)
b = torch.randn(3,requires_grad = True)#对于两个参数w,b需要求导
z = torch.matmul(x,w)+b
loss = torch.nn.functional.binary_cross_entropy_with_logits(z,y)#定义损失函数

#grad_fn用来记录变量是怎么来的，方便计算梯度，如y = x*3,grad_fn记录了y由x计算的过程。

print('Gradient function for z= ',z.grad_fn)
print('Gradient funciton for loss = ',loss.grad_fn)

#grad记录的是backward()后的梯度值
loss.backward()
print(w.grad)
print(b.grad)

输出结果：

Gradient function for z=  <AddBackward0 object at 0x000002ABB2BB4040>
Gradient funciton for loss =  <BinaryCrossEntropyWithLogitsBackward object at 0x000002ABB2BB49D0>
tensor([[0.2510, 0.2536, 0.3221],
        [0.2510, 0.2536, 0.3221],
        [0.2510, 0.2536, 0.3221],
        [0.2510, 0.2536, 0.3221],
        [0.2510, 0.2536, 0.3221]])
tensor([0.2510, 0.2536, 0.3221])

一个神经网络的完整实例 （使用FashionMNIST数据集）

# 一个神经网络完整实例
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensor, Lambda

device = 'cuda' if torch.cuda.is_available() else 'cpu'
print('Using {} device'.format(device))

# 设置超参数
learing_rate = 1e-3
batch_size = 64
epoches = 5

# 训练数据
training_data = datasets.FashionMNIST(
    root='data',
    train=True,
    download=True,
    transform=ToTensor()
)
# 测试数据
test_data = datasets.FashionMNIST(
    root='data',
    train=False,
    download=True,
    transform=ToTensor()
)
# 定义数据加载器
train_dataloader = DataLoader(training_data, batch_size=batch_size)
test_dataloader = DataLoader(test_data, batch_size=batch_size)

#for X, Y in test_dataloader:
#    print("shape of X [N,C,H,W]:", X.shape)  # N为数量 C为通道数 ，H,W为长和宽
#    print("Shape of Y: ", Y.shape, Y.type)  # y为标签
#    break


# 创建模型
class NeuralNetwork(nn.Module):
    def __init__(self):
        super(NeuralNetwork, self).__init__()
        self.flatten = nn.Flatten()
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(28 * 28, 512),
            nn.ReLU(),
            nn.Linear(512, 512),
            nn.ReLU()
        )

    def forward(self, x):  # 定义前向传播函数
        x = self.flatten(x)
        logits = self.linear_relu_stack(x)
        return logits


model = NeuralNetwork().to(device)  # 调用刚才创建的模型，如果cude可用 转到GPU上

# 选择损失函数
loss_fn = nn.CrossEntropyLoss()
# 选择优化器
optimizer = torch.optim.SGD(model.parameters(), lr=learing_rate)


# 每个epoch由两部分组成
# 训练loop：在训练集上进行迭代 试图收敛到最佳参数 
# 测试loop ：在测试集上进行迭代，检查模型的性能是否提高

# 训练loop三步：
# 1.调用optimizer.zero_grad()重置模型参数的梯度。因为梯度模型累加，为了防止重复计算，每次迭代都要归零
# 2.调用loss.backward()对预测值进行反向传播
# 3.调用optimizer.step()根据向后传递中收集的梯度调整参数(更新优化的参数)

# 定义训练loop：负责循环优化代码
def train_loop(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)  # 查看dataloader有多少张图片
    for batch, (X, y) in enumerate(dataloader):  # 从数据加载器中读取batch(一次读取多少张，即批次数),X(图片数据)，y(标签)
        X, y = X.to(device), y.to(device)
        pred = model(X)  # 得到的预测结果
        loss = loss_fn(pred, y)  # 计算预测和真实的误差

        # 反向传播
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        # 每训练100次 输出一次当前的信息
        if batch % 100 == 0:
            loss, current = loss.item(), batch * len(X)
            print(f"loss:{loss:>7f} [{current:>5d}/{size:>5d}]")


# 定义测试loop：负责根据测试数据评估模型的性能
def test_loop(dataloader, model, loss_fn):
    size = len(dataloader.dataset)
    num_batches = len(dataloader)
    model.eval()  # 将模型转换为验证模式
    test_loss, corrent = 0, 0
    with torch.no_grad():  # 测试时参数不用更新，只需要正向传播预测结果，这个函数范围内的数停止计算梯度
        for X, y in dataloader:
            X, y = X.to(device), y.to(device)
            pred = model(X)
            test_loss += loss_fn(pred, y).item()
            corrent += (pred.argmax(1) == y).type(torch.float).sum().item()  # 统计正确的个数

    test_loss /= num_batches
    corrent /= size
    print(f"Test ERROR :\n Accuracy:{(100 * corrent):>0.1f}%,Avg loss:{test_loss:>8f}\n")

#开始训练
epoch = 10
for t in range(epoch):
    print(f"Epoch {t + 1}\n--------------------")
    train_loop(train_dataloader,model,loss_fn,optimizer)
    test_loop(test_dataloader,model,loss_fn)
print("Done")

#保存模型
#Pytorch模型将学到的参数全都保存在state_dice内部状态字典中
torch.save(model.state_dict(),"model.pth")
print("Save Pytorch Model state to model.pth")

#加载模型
model = NeuralNetwork() #创建一个相同框架的模型
model.load_state_dict(torch.load("model.pth"))#torch.load读取模型pth load_state_dict读取参数

#获取预测结果
classes = [
    "T-shirt/top",
    "Trouser",
    "Pollover",
    "Dress",
    "coat",
    "Sandal",
    "Shirt",
    "Sneaker",
    "Bag",
    "Ankle boot",
]
model.eval() #进入验证阶段，不会梯度更新，只会正向传播
x, y = test_data[0][0], test_data[0][1]#拿出来测试集中的一对数据测试
with torch.no_grad():
    pred = model(x)
    print(pred.size())
    predicted, actual = classes[pred[0].argmax(0)], classes[y]
    print(f'Predicted: "{predicted}", Actual: "{actual}"')