PyTorch深度学习入门

传送门:B站 刘二大人 《PyTorch深度学习实践》完结合集_哔哩哔哩_bilibili

1.穷举法计算线性模型

知识点:前向传播、损失函数

问题描述:构建线性模型探究x到y的映射关系

主要思路:构建模型 y = w × x y = w \times x y=w×x, 穷举遍历不同的 w w w值,利用 M S E MSE MSE均方差衡量预测值与真实值的大小,绘图寻找使得损失函数最小的权重。

代码实现:

  • 构建数据集(输入+标签)
  • 定义前向计算和损失计算函数
  • 主函数外循环遍历权重,内循环为每次数据前向传播
  • 计算平均损失值,绘制平均损失值随着不同权重变化的曲线
import numpy as np
import matplotlib.pyplot as plt

# 1.数据整理
x_data = [1.0, 2.0, 3.0]
y_data = [2.0, 4.0, 6.0]


# 2.前向传播(算y_pred)
def forward(x):
    return x * w


# 3.损失函数计算(计算预测值与真实值差距)
def loss(x,y):
    y_pred = forward(x)
    return (y_pred - y)2


w_list = []
mse_list = []

# 4.设置不同权重,计算损失值并可视化(最优化问题)
for w in np.arange(0.0, 4.1, 0.1):  # 大循环遍历权重
    print('w=', )
    l_sum = 0
    for x_val,y_val in zip(x_data,y_data):  # 小循环处理所有数据集
        y_pred_val = forward(x_val)
        loss_val = loss(x_val,y_val)
        l_sum += loss_val
        print('\t', x_val, y_val, y_pred_val, loss_val)
    print('MSE=', l_sum/3)
    w_list.append(w)
    mse_list.append(l_sum/3)

# 5.绘制不同w下的数据集平均损失值
plt.plot(w_list,mse_list)
plt.ylabel('Loss')
plt.xlabel('w')
plt.show()
np.arange(a, b, c)生成从a到b的等差序列(公差为c)
zip函数将两个迭代器对应元素进行打包成元组
利用plt绘图记得加show()

2.梯度下降法更新权重

2.1梯度下降法

知识点:梯度下降、训练

问题描述:遍历权重主观性太强,需要制定相应的权重更新法则

主要思路:利用梯度下降法进行权重更新,计算损失沿着链式法则对权重的导数(更新方向),选择合适步长进行更新(更新大小)。引入训练的概念,多次循环对同一数据集进行权重更新。

代码实现:

  • 定义整体平均损失值函数、以及计算平均梯度函数。
  • 主函数构建训练循环,每一轮计算平均损失值和平均梯度,利用平均梯度以及自定义的学习率对权重进行更新
x_data = [1.0, 2.0, 3.0]
y_data = [2.0, 4.0, 6.0]
w = 1.0  # 初始权重猜测


def forward(x):
    # 前馈计算
    return x * w


def cost(xs, ys):
    # 计算平均损失
    cost = 0
    for x, y in zip(xs, ys):
        y_pre = forward(x)
        cost += (y_pre - y)  2
    return cost / len(xs)


def loss(x, y):
    # 损失值
    y_pre = forward(x)
    return (y_pre - y)  2


def gradient(xs, ys):
    # 计算平均梯度
    grad = 0
    for x, y in zip(xs, ys):
        grad += 2 * x * (x * w - y)
    return grad / len(xs)


print('Predict (before training)', 4, forward(4))
# 梯度下降
for epoch in range(100):
    cost_val = cost(x_data, y_data)
    grad_val = gradient(x_data, y_data)
    w -= 0.01 * grad_val  # 权重更新(每个epoch更新一次)
    print('Epoch', epoch, 'w=', w, 'loss=', cost_val)

print('Predict (after training)', 4, forward(4))

2.2 随机梯度下降法

知识点:SGD随机梯度下降法,minibatch思想

问题描述:进行样本级的梯度计算,效果优于计算整体平均梯度进行更新

主要思路:不再计算平均梯度进行更新。(逐样本更新权重会提高时间复杂度,后期引入minibatch的概念,即将整个数据集分成相应小份,每一小份一小份进行计算)

代码实现:主函数构建两层循环,将权重更新代码放入内层循环中

x_data = [1.0, 2.0, 3.0]
y_data = [2.0, 4.0, 6.0]
w = 1.0  # 初始权重猜测


def forward(x):
    return x * w


def loss(x, y):
    # 随机梯度下降
    y_pred = forward(x)
    return (y_pred - y)  2


def gradient(x, y):
    return 2 * x * (x * w - y)


print('Predict (before training)', 4, forward(4))
# 梯度下降
for epoch in range(100):
    for x, y in zip(x_data, y_data):
        grad = gradient(x, y)
        w -= 0.01 * grad  # 权重更新(每个样本更新一次)
        l = loss(x, y)

    print('Epoch', epoch, 'w=', w, 'loss=', l)

print('Predict (after training)', 4, forward(4))

3.反向传播

知识点:反向传播、torch

问题描述:上述方法需要自定义梯度计算函数,也就是要手动求梯度

主要思路:利用torch中自动求梯度的功能,进行反向传播更新梯度

代码实现:

  • 定义数据集、初始化权重w(定义为tensor张量,并将requires_grad设置为True
  • 构建两层循环,在内循环逐数据处理时需要进行:
    • 损失计算l = loss(x, y)
    • 反向传播l.backward()
    • 权重更新w.data = w.data - 0.01 * w.grad.data
    • 梯度清零w.grad.data.zero_()
import torch
# 1.数据集
x_data = [1.0, 2.0, 3.0]
y_data = [2.0, 4.0, 6.0]

w = torch.tensor([1.0])
w.requires_grad = True


def forward(x):
    return x * w


def loss(x, y):
    y_pred = forward(x)
    return (y_pred - y)2


print("predict (before training)", forward(4).item())

for epoch in range(100):
    for x, y in zip(x_data, y_data):
        l = loss(x, y)  # l是一个张量,tensor主要是在建立计算图 forward, compute the loss
        l.backward()
        print('\tgrad:', x, y, w.grad.item())

        w.data = w.data - 0.01 * w.grad.data  # 权重更新时,注意grad也是一个tensor
        w.grad.data.zero_()
    print('process:', epoch, l.item())  # 取出loss使用l.item,不要直接使用l(l是tensor会构建计算图)
print("predict (after training)", 4, forward(4).item())
tensor张量包含data和grad两个属性,且data和grad也是张量,故更新w.data时需使用w.grad.data。
grad初始为None,对需要计算梯度的张量设置require_grad为True后,利用backward后自动计算梯度,l.backward()会把计算图中所有需要梯度(grad)的地方都会求出来,然后把梯度都存在对应的待求的参数中,最终计算图被释放。
在print或者只需要tensor的数值时,记得利用item()转化为标量将数值取出来

4.利用PyTorch是实现线性回归

知识点:模型架构

问题描述:利用PyTorch是实现线性回归

代码实现:基本框架包括:数据集准备、模型定义并实例化、损失函数和优化器、模型训练测试

  • 数据集准备
    • 需要为tensor张量矩阵形式,float32类型。分为输入和标签,每组要进行对应,注意维度含义
  • 模型定义并实例化
    • 继承torch.nn.Module类,定义__init__(self)forward(self,x)
    • __init__(self)中利用super调用父类构造函数,并初始化模型所需要的层
    • forward(self,x)进行前向传播
    • 实例化模型对象
  • 损失函数与优化器
    • 损失函数:计算损失值,进行反向传播
    • 优化器:进行梯度清零与更新,需要指定模型总参数、学习率
  • 模型训练测试
    • 计算输出值、计算损失值、梯度清零、反向传播、梯度更新
import torch

# 1.数据准备
# 矩阵x,y为3行1列,有3个数据,每组数据只有1个特征
x_data = torch.tensor([[1.0], [2.0], [3.0]])
y_data = torch.tensor([[2.0], [4.0], [6.0]])


# 2.模型设计
"""
our model class should be inherit from nn.Module, which is base class for all neural network modules.
member methods __init__() and forward() have to be implemented
class nn.linear contain two member Tensors: weight and bias
class nn.Linear has implemented the magic method __call__(),which enable the instance of the class can
be called just like a function.Normally the forward() will be called 
"""
class LinearModel(torch.nn.Module):
    def __init__(self):
        super(LinearModel, self).__init__()
        self.linear = torch.nn.Linear(1, 1)
        # (1,1)是指输入x和输出y的特征维度,这里数据集中的x和y的特征都是1维的
        # 该线性层需要学习的参数是w和b  获取w/b的方式分别是~linear.weight/linear.bias

    def forward(self,x):
        y_pred = self.linear(x)
        return y_pred

# 3.实例化对象
model = LinearModel()

# 4.损失函数和优化器
criterion = torch.nn.MSELoss(size_average=False)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

# 5.训练循环
for epoch in range(500):
    y_pred = model(x_data)
    loss = criterion(y_pred, y_data)
    print(epoch, loss.item())

    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

print('w=', model.linear.weight.item())
print('b=',model.linear.bias.item())

# 5.模型测试
x_test = torch.tensor([[4.0]])
y_test = model(x_test)
print('y_test=', y_test.data)

5.逻辑回归求解二分类问题

知识点:sigmoid激活函数,交叉熵损失函数BCELoss

问题描述:针对于输出为离散的分类问题

主要思路:利用sigmoid函数求解某一类的概率,利用交叉熵损失函数BCELoss衡量不同分布的差异大小

代码实现:在模型定义上最后一层添加sigmoid函数,更改损失函数

import torch
import torch.functional as F
# 1.数据集处理
x_data = torch.tensor([[1.0], [2.0], [3.0]])
y_data = torch.tensor([[0.0], [0.0], [1.0]]) # 输出也要是float类型,不然损失值计算会报错

# 2.模型建立
class LogisticModual(torch.nn.Module):
    def __init__(self):
        super(LogisticModual,self).__init__()
        self.linear = torch.nn.Linear(1, 1)

    def forward(self, x):
        y_pred = torch.sigmoid(self.linear(x))  # 在前馈过程中加入sigmoid函数将输出映射到01区间
        return y_pred


model = LogisticModual()

# 3.损失函数和优化器
criterion = torch.nn.BCELoss()  # 换用BCE交叉熵损失函数(适合二分类任务)
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)

# 4.训练循环
for epoch in range(500):
    y_pred = model(x_data)
    loss = criterion(y_pred, y_data)
    optimizer.zero_grad()  # 梯度清零
    loss.backward()  # 反向传递
    optimizer.step()  # 梯度更新
    print('epoch=', epoch, 'loss=', loss.item())  # 在print时注意是否为张量,取数值加item()


x_test = torch.tensor([[4.0]])
y_test = model(x_test)
print('y_test=', y_test.data.item())


# 绘图
import matplotlib.pyplot as plt
import numpy as np
x = np.linspace(0, 10, 200)  # 0到10取200个点
x_t = torch.tensor(x, dtype=torch.float).view(200, 1)  # 类似于numpy中的reshape
y_t = model(x_t)
plt.plot(x_t, y_t.data.numpy())  # tensor取数值转化到numpy
plt.ylabel('Probability of Pass')
plt.xlabel('Hours')
plt.plot([0,10], [0.5,0.5], c='r')
plt.grid()
plt.show()
注意输入和标签的数据集
绘图时先用np生成矩阵,利用torch.tensor转张量后利用view更改维度后输入网络
绘图时还要采用np形式,注意用tensor的data,直接接numpy()就能转为numpy数组了

6.多维数据分类问题

知识点:多层网络,激活函数连接

问题描述:输入数据存在多维特征,输出为多类别

主要思路:输入维度增加,基于矩阵乘法的行列关系,确定模型层输入输出维度即可。此外可以提高模型的复杂度,由单层网络变成多层网络,连接处需要加入激活函数引入非线性

代码实现:

  • 数据集处理:利用np.loadtxt()确定分隔符,以np.float32数据类型读入,转张量后切片分割数据集
  • 在模型构造时设置多个线性层扩充网络层数,在前向传递的过程中每层接入sigmoid激活函数
import numpy as np
import torch
import matplotlib.pyplot as plt

# 1.数据准备(读取gz文件夹数据,并用逗号分隔)
xy = np.loadtxt('diabetes.csv.gz', delimiter=',', dtype=np.float32)
x_data = torch.from_numpy(xy[:, :-1])  # torch.Size([759, 8])
y_data = torch.from_numpy(xy[:, [-1]]) # torch.Size([759, 1])


# 2.模型构造
class Module(torch.nn.Module):
    def __init__(self):
        super(Module, self).__init__()
        self.linear1 = torch.nn.Linear(8,6)
        self.linear2 = torch.nn.Linear(6, 4)
        self.linear3 = torch.nn.Linear(4, 1)
        self.sigmoid = torch.nn.Sigmoid()

    def forward(self, x):
        x = self.sigmoid(self.linear1(x))
        x = self.sigmoid(self.linear2(x))
        x = self.sigmoid(self.linear3(x))
        return x


model = Module()

# 3.构造损失函数和优化器
criterion = torch.nn.BCELoss(reduction='mean')
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

# 4.训练循环
epochlist = []
losslist = []
for epoch in range(100):
    y_pred = model(x_data)
    loss = criterion(y_pred, y_data)
    epochlist.append(epoch)
    losslist.append(loss.item())
    print('epoch=', epoch, 'loss=', loss.item())

    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

plt.plot(epochlist, losslist)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()
torch.from_numpy()可以将数据转为torch张量(送入网络前要先转为张量)
np.loadtxt()可以读取指定文件,同时支持压缩文件,包括gz,bz类型
注意输入输出的维度大小

7.数据集构建

知识点:DatasetDataload,以及minibatch训练

问题描述:将数据集进行统一的封装

主要思路:利用Dataset划分数据集,转tensor张量,并支持索引;利用Dataload进行批处理并打乱数据集

代码实现:

  • 由于Dataset为抽象类,需要自定义后继承
    • __init__:一般根据传入的文件名去指定文件获取数据集,并定义len(shape后取行数), x, y作为属性
    • __getitem__:利用索引直接返回即可
    • __len__:直接返回数据长度的属性
  • Dataload对其进行实例化即可,需要指定:数据集、是否打乱、多线程等(由于winlinux下多线程调用库不一致,因此主循环需要加入判断语句)
  • 由于采用batch进行训练,因此要进行嵌套for循环,Dataload处理成一批批数据了,每次读入一批
import numpy as np
import torch
from torch.utils.data import Dataset, DataLoader
import matplotlib.pyplot as plt


# 1.数据准备(Dataset为抽象类,需要实例化)
class DiabetesDataset(Dataset):
    def __init__(self, filepath):
        xy = np.loadtxt(filepath, delimiter=',', dtype=np.float32)
        self.len = xy.shape[0]
        self.x = torch.from_numpy(xy[:, :-1])
        self.y = torch.from_numpy(xy[:, [-1]])

    def __getitem__(self, index):
        return self.x[index], self.y[index]

    def __len__(self):
        return self.len


dataset = DiabetesDataset('diabetes.csv.gz')
train_loader = DataLoader(dataset=dataset, batch_size=32, shuffle=True, num_workers=0)


# 2.模型定义
class LogicModule(torch.nn.Module):
    def __init__(self):
        super(LogicModule, self).__init__()
        self.linear1 = torch.nn.Linear(8, 6)
        self.linear2 = torch.nn.Linear(6, 4)
        self.linear3 = torch.nn.Linear(4, 1)
        self.sigmoid = torch.nn.Sigmoid()

    def forward(self, x):
        x = self.sigmoid(self.linear1(x))
        x = self.sigmoid(self.linear2(x))
        x = self.sigmoid(self.linear3(x))
        return x


model = LogicModule()

# 3.损失函数和优化器
criterion = torch.nn.BCELoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

# 4.模型训练(由于new_work在win和linux下不兼容,需要用if)
if __name__ == '__main__':
    loss_list = []
    for epoch in range(500):
        loss_epoch = 0
        for i, (x, y) in enumerate(train_loader):  # 这里只有训练数据
            y_pre = model(x)
            loss = criterion(y_pre, y)
            loss_epoch += loss.item()
            # print('epoch=', epoch, 'loss=', loss.item())
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

        loss_list.append(loss_epoch/i)

    # 绘制每个epoch中每个batch平均损失随epoch的图像
    plt.plot(loss_list)
    plt.show()

8.基于SoftMax的多分类

知识点:SoftMax激活函数,多分类交叉熵CrossEntropyLoss,图像transform预处理,训练测试单独封装,``torchvision`库

问题描述:datasetMINIST数据集读取方式为PIL不能直接输入网络,同时多分类问题输出需要为一个分布

主要思路:利用transform进行预处理,将图片拉直后进入线性网络,利用多分类交叉熵CrossEntropyLoss计算损失值(线性层传入即可,CrossEntropyLoss = SoftMax + NLLLoss

代码实现:

  • 数据处理:transforms构建Compose时用ToTensor转为张量,用Normalize进行标准化(涉及到的函数都来自于torchvision.transforms
    • 利用datasetMINIST数据集指定下载、路径、是否为训练集、transform
    • 利用Dataloader整理成分批数据集
  • 由于用线性网络实现,前向传播时先利用view将图像拉直过线性层(注意最后一层不用过激活函数),采用CrossEntropyLoss损失函数
  • 训练:内层循环单独拿出来(传入epoch)记录batch数,进行输出(loss存的是item(),同时输出后记得置零)
  • 测试:
    • 不用计算梯度,利用with torch.no_grad():
    • 利用正确数除以总数记录精确度
      • 总数:累加每个batchsize(0)
      • 正确数:利用``torch.max找到最大概率值,获取其索引于真实值进行比较,利用sum()`汇总数据
import torch
from torchvision import transforms
from torchvision import datasets
from torch.utils.data import DataLoader
import torch.nn.functional as F

# 1.数据集准备
batch_size = 64
# 转化为tensor张量+归一化处理(均值、方差)
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
train_dataset = datasets.MNIST(root='../dataset/minist/', train=True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_dataset = datasets.MNIST(root='../dataset/minist/', train=False, download=True, transform=transform)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)


# 2.模型建立
class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.l1 = torch.nn.Linear(784, 512)
        self.l2 = torch.nn.Linear(512, 256)
        self.l3 = torch.nn.Linear(256, 128)
        self.l4 = torch.nn.Linear(128, 64)
        self.l5 = torch.nn.Linear(64, 10)

    def forward(self, x):
        # 28*28 图片拉成1*784的向量(view(-1,b)表示在a参数未知情况下自动补齐行向量长度)
        x = x.view(-1, 784)
        x = F.relu(self.l1(x))
        x = F.relu(self.l2(x))
        x = F.relu(self.l3(x))
        x = F.relu(self.l4(x))
        return self.l5(x)  # 后续要用到交叉熵,不用再过激活函数了


model = Net()

# 3.损失函数和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.5)  # 带冲量的SGD


# 4.单独把训练层拿出来
def train(epoch):
    running_loss = 0.0
    for batch_idx, (inputs, target) in enumerate(train_loader):
        outputs = model(inputs)
        loss = criterion(outputs, target)
        running_loss += loss.item()
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        if batch_idx % 300 == 299:
            print('[%d, %5d] loss:%.3f', epoch+1, batch_idx+1, running_loss/300)
            running_loss = 0.0


def test():
    correct = 0
    total = 0
    with torch.no_grad():  # 不计算梯度
        for data in test_loader:
            images, labels = data
            outputs = model(images)
            # 取每行(十类)概率最大项,列为0,行为1
            # (max输出两个张量,第一个是该维度最大值,第二个是序号),在此取第二个就行
            _, predicted = torch.max(outputs.data, dim=1)
            total += labels.size(0)  # size获取了N*1,第一项可以读出总样本数
            correct += (predicted == labels).sum().item()  # 逐个元素比较后求总和
        print('accuracy on test set: %d %%' % (100*correct/total))


# 为避免多进程错误
if __name__ == '__main__':
    for epoch in range(10):
        train(epoch)
        test()
由于 transforms.Compose() 接受的参数是一个列表,所以需要使用 [] 将变换组织成一个列表,以便传递给这个函数。
输入线性层之前一定要记得拉直成向量
torch.max(dim=1)会返回两个值(数和索引),dim表示计算最值的维度,1表示列,也就是取每行的所有列中的最值
进行计数使用tensor.data

9. CNN基础篇

9.1 CNN

知识点:卷积层、池化层(主要在模型构建上)

问题描述:卷积神经网络保留图像空间结构,采用特征提取+分类的形式更有效处理图像

主要思路:利用卷积层+池化层叠加(池化层只改变图像大小)最后拉直向量过全连接层

代码实现:

  • 在模型初始化定义好卷积层、池化层以及线性层(确定好输入输出维度)
    • 前者只考虑通道数即可(卷积核的通道数要与输入一致,输出的通道数要和卷积核个数一致)
    • 输入线性层前需要算好拉直后的维度,可以预设个输入,查看输出的shape
  • 前向传播
    • 计算输入的batch,方便后续进行拉直x = x.view(batch_size, -1)
    • 最后记得通过线性层+采用CrossEntropyLoss()
import torch
from torchvision import transforms
from torchvision import datasets
from torch.utils.data import DataLoader

# 1.数据集准备
batch_size = 64
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
train_dataset = datasets.MNIST(root='../dataset/minist/', train=True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size, num_workers=2)
test_dataset = datasets.MNIST(root='../dataset/minist/', train=False, download=True, transform=transform)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=batch_size, num_workers=2)


# 2.模型构建
class CNNNet(torch.nn.Module):
    def __init__(self):
        super(CNNNet, self).__init__()
        self.cov1 = torch.nn.Conv2d(1, 10,5)
        self.cov2 = torch.nn.Conv2d(10, 20,5)
        self.pool = torch.nn.MaxPool2d(2)
        self.linear = torch.nn.Linear(320, 10)

    def forward(self, x):
        batch_size = x.size(0) # B,C,W,H(之前的transform已经将图像转为tensor张量了,取第一个维度就是batch)
        x = self.pool(torch.relu(self.cov1(x)))
        x = self.pool(torch.relu(self.cov2(x)))
        x = x.view(batch_size, -1)  # batch是不变的,把CWH拉长
        # print(x.shape)
        return self.linear(x)


model = CNNNet()
# sample_image = torch.randn(1, 1, 28, 28)  # 1张图像,1个通道,28x28大小的图像
# output = model(sample_image)


# 3.损失值和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.5)


# 4.训练循环
def train(epoch):
    l = 0.0
    for batch_index, (x, y) in enumerate(train_loader):
        y_pred = model(x)
        optimizer.zero_grad()
        loss = criterion(y_pred, y)
        l += loss.item()
        loss.backward()
        optimizer.step()

        if batch_index % 300 == 299:
            print(f'[epoch{epoch+1}---------batch={batch_index+1}---------loss={round(100*l/300, 3)}]')
            l = 0.0 # 输出完记得置为0


def test():
    size = 0
    acc = 0
    with torch.no_grad():
        for (x, y) in test_loader:
            y_pred = model(x)
            _, predict = torch.max(y_pred.data, dim=1)  # 0列1行,注意这里取的是data(用到张量的时候要格外小心)
            size += predict.size(0)
            acc += (predict == y).sum().item()  # 与标签进行比较
    print('test accuracy= %.3f %%' % (100 * acc / size))


if __name__ == "__main__":
    for epoch in range(10):
        train(epoch)
        test()

9.2 CNN(GPU加速)

知识点:将模型和数据上载到GPU进行加速

代码实现:设置device,利用to()对模型和数据上载即可

import torch
from torchvision import transforms
from torchvision import datasets
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt

# 1.数据集准备
batch_size = 64
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
train_dataset = datasets.MNIST(root='../dataset/minist/', train=True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size, num_workers=2)
test_dataset = datasets.MNIST(root='../dataset/minist/', train=False, download=True, transform=transform)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=batch_size, num_workers=2)


# 2.模型构建
class CNNNet(torch.nn.Module):
    def __init__(self):
        super(CNNNet, self).__init__()
        self.cov1 = torch.nn.Conv2d(1, 10,5)
        self.cov2 = torch.nn.Conv2d(10, 20,5)
        self.pool = torch.nn.MaxPool2d(2)
        self.linear = torch.nn.Linear(320, 10)

    def forward(self, x):
        batch_size = x.size(0) # B,C,W,H(之前的transform已经将图像转为tensor张量了,取第一个维度就是batch)
        x = self.pool(torch.relu(self.cov1(x)))
        x = self.pool(torch.relu(self.cov2(x)))
        x = x.view(batch_size, -1)  # batch是不变的,把CWH拉长
        return self.linear(x)


model = CNNNet()
device = torch.device('cuda'if torch.cuda.is_available() else 'cpu')
model.to(device)

# 3.损失值和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.5)


# 4.训练循环
def train(epoch):
    l = 0.0
    for batch_index, (x, y) in enumerate(train_loader):
        x, y = x.to(device), y.to(device)
        y_pred = model(x)
        optimizer.zero_grad()
        loss = criterion(y_pred, y)
        l += loss.item()
        loss.backward()
        optimizer.step()

        if batch_index % 300 == 299:
            print(f'[epoch{epoch+1}---------batch={batch_index+1}---------loss={round(100*l/300, 3)}]')
            l = 0.0 # 输出完记得置为0


def test():
    size = 0
    acc = 0
    with torch.no_grad():
        for (x, y) in test_loader:
            x, y = x.to(device), y.to(device)
            y_pred = model(x)
            _, predict = torch.max(y_pred.data, dim=1)  # 0列1行,注意这里取的是data(用到张量的时候要格外小心)
            size += predict.size(0)
            acc += (predict == y).sum().item()  # 与标签进行比较
    print('test accuracy= %.3f %%' % (100 * acc / size))
    return acc / size


if __name__ == "__main__":
    epoch_list = []
    acc_list = []
    for epoch in range(10):
        train(epoch)
        acc = test()
        epoch_list.append(epoch)
        acc_list.append(acc)
    plt.plot(epoch_list, acc_list)
    plt.ylabel('accuracy')
    plt.xlabel('epoch')
    plt.show()

10. CNN(高级篇)

10.1 GoogleNet(Inception层)

知识点:Inception层模块单独定义

问题描述:造神经网络里面有一些超参数比较难选(例如卷积核大小)

主要思路:利用多个分支设置不同的卷积核大小,让模型自己进行选择(哪个好就赋较大权重)

代码实现:

  • 定于InceptionA模块,由于要进行复用,将输入通道作为参数传递进来
  • 涉及不同的分支,整理成列表,在前向通道利用torch.cat()进行拼接即可,指定dim=1,以C维度进行拼接
import torch
from torchvision import datasets
from torch.utils.data import DataLoader
from torchvision import transforms
import torch.nn.functional as F

# 1.数据集准备
batch_size = 64
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
train_dataset = datasets.MNIST(root='../dataset/minist', train = True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size, num_workers=2)
test_dataset = datasets.MNIST(root='../dataset/minist', train = False, download=True, transform=transform)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=batch_size, num_workers=2)


# 2.模型构建
class InceptionA(torch.nn.Module):
    def __init__(self, in_channel):
        super(InceptionA,self).__init__()
        # 1*1卷积分支
        self.branch1x1 = torch.nn.Conv2d(in_channel, 16, 1)
        # 5*5卷积分支
        self.branch5x5_1 = torch.nn.Conv2d(in_channel, 16, 1)
        self.branch5x5_2 = torch.nn.Conv2d(16, 24, 5, padding=2)
        # 3*3卷积分支
        self.branch3x3_1 = torch.nn.Conv2d(in_channel, 16, 1)
        self.branch3x3_2 = torch.nn.Conv2d(16,24, 3, padding=1)
        self.branch3x3_3 = torch.nn.Conv2d(24, 24, 3, padding=1)
        # 平均池化分支
        self.pool = torch.nn.Conv2d(in_channel, 24, 1)

    def forward(self, x):
        branch1x1 = self.branch1x1(x)

        branch5x5 = self.branch5x5_1(x)
        branch5x5 = self.branch5x5_2(branch5x5)

        branch3x3 = self.branch3x3_1(x)
        branch3x3 = self.branch3x3_2(branch3x3)
        branch3x3 = self.branch3x3_3(branch3x3)

        branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)
        branch_pool = self.pool(branch_pool)

        outputs = [branch1x1, branch5x5, branch3x3, branch_pool]
        return torch.cat(outputs, dim=1)  # b,c,w,h  c对应的是dim=1


class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.c1 = torch.nn.Conv2d(1, 10, 5)
        self.c2 = torch.nn.Conv2d(88, 20, 5)

        self.inception1 = InceptionA(10)  # 与conv1 中的10对应
        self.inception2 = InceptionA(20)  # 与conv2 中的20对应

        self.mp = torch.nn.MaxPool2d(2)  # 图像缩小一半 12  (不要改步长啊)
        self.linear = torch.nn.Linear(1408, 10)

    def forward(self, x):
        batch = x.size(0)
        x = torch.relu(self.mp(self.c1(x)))  # b*10*12*12
        x = self.inception1(x)
        x = torch.relu(self.mp(self.c2(x)))  # b*20*4*4
        x = self.inception2(x)
        x = x.view(batch, -1)
        # print(x.shape)
        return self.linear(x)


device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = Net().to(device)
# # 创建一个示例图像查看模型输出shape(好给全连接层赋值)----->输出:torch.Size([1, 1408])
# sample_image = torch.randn(1, 1, 28, 28)  # 1张图像,1个通道,28x28大小的图像
# output = model(sample_image)

# 3.损失函数和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

# 4.训练
def train(epoch):
    running_loss = 0.0
    for batch_idex, (x, y) in enumerate(train_loader):
        x, y = x.to(device), y.to(device)
        y_pred = model(x)
        optimizer.zero_grad()
        loss = criterion(y_pred, y)
        running_loss += loss.item()
        loss.backward()
        optimizer.step()

        if batch_idex % 300 == 299:
            print(f'epoch{epoch+1}--------batch{batch_idex+1}-------loss={round(running_loss/300, 3)}')
            running_loss = 0.0


def test():
    total = 0
    acc = 0
    with torch.no_grad():
        for (x, y) in test_loader:
            x, y = x.to(device), y.to(device)
            y_pred = model(x)
            total += y_pred.size(0)
            _, predicted = torch.max(y_pred, dim=1)
            acc += (predicted == y).sum().item()
    print('test= %.3f %%' % (100 * acc/total))


if __name__ == '__main__':
    for epoch in range(10):
        train(epoch)
        test()

10.2 Residual Net

知识点:残差层定义

问题描述:卷积核层数不是越深越好,可能存在梯度消失

主要思路:引入残差连接,拼接后再激活,计算梯度的时候就能有所保留,要求输入输出大小相同

代码实现:定义残差块类,指定输入通道数,跳转拼接后再激活。模型构建时再定义相关层

import torch
from torchvision import datasets
from torch.utils.data import DataLoader
from torchvision import transforms
import torch.nn.functional as F

# 1.数据集准备
batch_size = 64
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
train_dataset = datasets.MNIST(root='../dataset/minist', train = True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size, num_workers=2)
test_dataset = datasets.MNIST(root='../dataset/minist', train = False, download=True, transform=transform)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=batch_size, num_workers=2)

# 2.模型构建
class ResidualBlock(torch.nn.Module):
    def __init__(self, channels):
        super(ResidualBlock,self).__init__()
        self.channels = channels  # 过残差连接输入输出通道不变
        self.conv1 = torch.nn.Conv2d(channels, channels, 3,padding=1)
        self.conv2 = torch.nn.Conv2d(channels, channels, 3,padding=1)

    def forward(self, x):
        y = F.relu(self.conv1(x))
        y = self.conv2(y)
        return F.relu(x + y)  # 先求和再激活

class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.c1 = torch.nn.Conv2d(1, 16, 5)
        self.c2 = torch.nn.Conv2d(16, 32, 5)

        self.rblock1 = ResidualBlock(16)  # 与conv1 中的16对应
        self.rblock2 = ResidualBlock(32)  # 与conv2 中的32对应

        self.mp = torch.nn.MaxPool2d(2)  # 图像缩小一半 12  (不要改步长啊)
        self.l = torch.nn.Linear(512, 10)

    def forward(self, x):
        batch = x.size(0)
        x = torch.relu(self.mp(self.c1(x)))  # b*10*12*12
        x = self.rblock1(x)
        x = torch.relu(self.mp(self.c2(x)))  # b*20*4*4
        x = self.rblock2(x)
        x = x.view(batch, -1)
        # print(x.shape)
        return self.l(x)


device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = Net().to(device)
# # 创建一个示例图像查看模型输出shape(好给全连接层赋值)----->输出:torch.Size([1, 1408])
# sample_image = torch.randn(1, 1, 28, 28)  # 1张图像,1个通道,28x28大小的图像
# output = model(sample_image)

# 3.损失函数和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

# 4.训练
def train(epoch):
    running_loss = 0.0
    for batch_idex, (x, y) in enumerate(train_loader):
        x, y = x.to(device), y.to(device)
        y_pred = model(x)
        optimizer.zero_grad()
        loss = criterion(y_pred, y)
        running_loss += loss.item()
        loss.backward()
        optimizer.step()

        if batch_idex % 300 == 299:
            print(f'epoch{epoch+1}--------batch{batch_idex+1}-------loss={round(running_loss/300, 3)}')
            running_loss = 0.0


def test():
    total = 0
    acc = 0
    with torch.no_grad():
        for (x, y) in test_loader:
            x, y = x.to(device), y.to(device)
            y_pred = model(x)
            total += y_pred.size(0)
            _, predicted = torch.max(y_pred, dim=1)
            acc += (predicted == y).sum().item()
    print('test= %.3f %%' % (100 * acc/total))


if __name__ == '__main__':
    for epoch in range(10):
        train(epoch)
        test()

11. RNN循环神经网络

知识点:RNN处理数据维度,两种构建RNN的方法

问题描述:对于序列数据采用循环神经网络

11.1 RNNCell

主要内容:随机生成seq_size个句子,通过h0为全零的RNN网络,探究输入输出以及隐层的维度

代码实现:利用RNNCell处理单个运算,RNNCell只是RNN的一个单元,用于处理一个时间步的输入数据,需要在循环中手动处理时间步。

seqLen:每个句子所含词的个数
batchSize:每次处理词的批数
InputSize:每个词嵌入向量的位数
dataset: (seq_size, batch_size, input_size)——>整个数据集维度
input: (batch_size, input_size)——>每个时间步处理的输入(input in dataset)
hidden: (batch_size, output_size)——>每次处理的隐层(即为output)
import torch

batch_size = 1
seq_len = 3
input_size = 4
hidden_size = 2

# Construction of RNNCell
cell = torch.nn.RNNCell(input_size=input_size, hidden_size=hidden_size)
# Wrapping the sequence into:(seqLen,batchSize,InputSize)
dataset = torch.randn(seq_len, batch_size, input_size)  # (3,1,4)
# Initializing the hidden to zero
hidden = torch.zeros(batch_size, hidden_size)  # (1,2)

for idx, input in enumerate(dataset):
    print('=' * 20, idx, '=' * 20)
    print('Input size:', input.shape)  # (batch_size, input_size)
    # 按序列依次输入到cell中,seq_len=3,故循环3次
    hidden = cell(input, hidden)  # 返回的hidden是下一次的输入之一,循环使用同一个cell

    print('output size:', hidden.shape)  # (batch_size, hidden_size)
    print(hidden.data)

11.2 RNNCell

主要内容:随机生成seq_size个句子,通过h0为全零的RNN网络,探究输入输出以及隐层的维度

代码实现:

输入dataset:整个序列(seq_size, batch_size, input_size)   取每一个词input(batch_size, input_size)
隐层:(num_layers, batch_size, hidden_size)
输出(有两个): output:h1~hn  (seq_size, batch_size, hidden_size)
              hidden: hn (num_layers, batch_size, hidden_size)
import torch

batch_size = 1
seq_len = 3
input_size = 4
hidden_size = 2
num_layers = 1  # RNN层数

# 构建RNNCell
rnn = torch.nn.RNN(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers)
# 构建inputs, 需要整理成(seq_len, batch_size, input_size)
inputs = torch.randn(seq_len, batch_size, input_size)
# 全零初始化h0,需要整理成(num_layers, batch_size, hidden_size)
hidden = torch.zeros(num_layers, batch_size, hidden_size)

output, hidden = rnn(inputs, hidden)
print('Output size:', output.shape)  # (seq_len, batch_size, hidden_size)
print('Output:', output.data)
print('Hidden size:', hidden.shape)  # (num_layers, batch_size, hidden_size)
print('Hidden:', hidden.data)
1.与Cell相比多了一个num_layers,同时不用自己写循环
2.输出output是历史hidden记录, hidden为最后一个隐层输出
3.可以指定batch_first参数,使得以(batch_size, seq_len, input_size)进行输入

两者的区别
1.RNNCell只是RNN的一个单元,用于处理一个时间步的输入数据,需要在循环中手动处理时间步。
2.完整的RNN模型,它内部包含了循环结构,可以一次性处理整个序列的输入,从而避免了手动处理时间步的繁琐过程。

11.3 基于RNNCell处理seq->seq任务

import torch

# 1.参数设置
input_size = 4
hidden_size = 4  # 输出维度要与输入一致,才能在同一个表上查询
batch_size = 1

# 2.数据准备
index2char = ['e', 'h', 'l', 'o']  # 字典
x_data = [1, 0, 2, 2, 3]  # hello
y_data = [3, 1, 2, 3, 2]  # 标签:ohlol
# one-hot 查找表(维度为字典中词的个数)
one_hot_lookup = [[1, 0, 0, 0],  # 用来将x_data转换为one-hot向量的参照表
                  [0, 1, 0, 0],
                  [0, 0, 1, 0],
                  [0, 0, 0, 1]]
x_one_hot = [one_hot_lookup[x] for x in x_data]  # 转化为独热向量,维度(seq, batch, input)
dataset = torch.Tensor(x_one_hot).view(-1, batch_size, input_size)  # 转tensor后整理成(𝒔𝒆𝒒𝑳𝒆𝒏,𝒃𝒂𝒕𝒄𝒉𝑺𝒊𝒛𝒆,𝒊𝒏𝒑𝒖𝒕𝑺𝒊𝒛𝒆)
labels = torch.LongTensor(y_data).view(-1, 1)  # 每层对应一个类(seqLen,𝟏)计算交叉熵损失时标签不需要我们进行one-hot编码,其内部会自动进行处理


# 3.模型构建
class Model(torch.nn.Module):
    def __init__(self, input_size, hidden_size, batch_size):  # 需要指定输入,隐层,批
        super(Model, self).__init__()
        self.batch_size = batch_size
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.rnncell = torch.nn.RNNCell(input_size=self.input_size, hidden_size=hidden_size)

    def forward(self, input, hidden):
        hidden =self.rnncell(input, hidden)
        return hidden

    def init_hidden(self):  # 只有在初始化隐藏层需要batch_size
        return torch.zeros(self.batch_size, self.hidden_size)


net = Model(input_size, hidden_size, batch_size)

# 4.损失和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.1)

# 5.训练
for epoch in range(15):
    loss = 0
    optimizer.zero_grad()
    hidden = net.init_hidden()  # 初始化h0
    print('Predicted string:', end='')
    for input, label in zip(dataset, labels):
        hidden = net(input, hidden)
        loss += criterion(hidden, label)  # 这里要求总损失进行梯度下降(构建计算图),因此不能取item()
        _, idx = hidden.max(dim=1)  # 取出概率最大的索引
        print(index2char[idx.item()], end='')
    loss.backward()
    optimizer.step()
    print(',Epoch [%d / 15] loss:%.4f' % (epoch+1, loss.item()))
对于输入序列进行独热编码,构建RNNCell,对每次输入的字符依次进入循环
针对于分类任务的输出要使用LongTensor类型
求总损失进行梯度下降(构建计算图),因此不能取item()

11.4 基于RNN处理seq->seq任务

import torch

# 1.参数设置
seq_len = 5
input_size = 4
hidden_size = 4
batch_size = 1

# 2.数据准备
index2char = ['e', 'h', 'l', 'o']
x_data = [1, 0, 2, 2, 3]
y_data = [3, 1, 2, 3, 2]
one_hot_lookup = [[1, 0, 0, 0],
                  [0, 1, 0, 0],
                  [0, 0, 1, 0],
                  [0, 0, 0, 1]]
x_one_hot = [one_hot_lookup[x] for x in x_data]
inputs = torch.Tensor(x_one_hot).view(-1, batch_size, input_size)
labels = torch.LongTensor(y_data)

# 3.模型构建
class Model(torch.nn.Module):
    def __init__(self, input_size, hidden_size, batch_size, num_layers=1):  # 需要指定输入,隐层,批
        super(Model, self).__init__()
        self.batch_size = batch_size
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.rnn = torch.nn.RNN(input_size=self.input_size, hidden_size=self.hidden_size, num_layers=self.num_layers)

    def forward(self, input):
        hidden = torch.zeros(self.num_layers,
                             self.batch_size,
                             self.hidden_size)
        out, _ = self.rnn(input, hidden)  # out: tensor of shape (seq_len, batch, hidden_size)
        return out.view(-1, self.hidden_size)  # 将输出的三维张量转换为二维张量,(𝒔𝒆𝒒𝑳𝒆𝒏×𝒃𝒂𝒕𝒄𝒉𝑺𝒊𝒛𝒆,𝒉𝒊𝒅𝒅𝒆𝒏𝑺𝒊𝒛𝒆)


net = Model(input_size, hidden_size, batch_size)

# 4.损失和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.5)

# 5.训练
for epoch in range(15):
    optimizer.zero_grad()
    outputs = net(inputs)
    print(outputs.shape)
    loss = criterion(outputs, labels)
    loss.backward()
    optimizer.step()

    _, idx = outputs.max(dim=1)
    idx = idx.data.numpy()
    print('Predicted string: ', ''.join([index2char[x] for x in idx]), end='')
    print(',Epoch [%d / 15] loss:%.4f' % (epoch+1, loss.item()))
输入:seq_size * batch_size * input_size
输出:seq_size * batch_size * hidden_size, 利用view变成(seq_size * batch_size, hidden_size)作为一个矩阵
labels: seq_size * batch_size * 1  (序列中每一个样本的分类)

最后的out和labels是针对整个seq进行比较计算损失的

11.5 含嵌入层的RNN网络

import torch

# 1、确定参数
num_class = 4
input_size = 4
hidden_size = 8
embedding_size = 10
num_layers = 2
batch_size = 1
seq_len = 5

# 2、准备数据
index2char = ['e', 'h', 'l', 'o']  # 字典
x_data = [[1, 0, 2, 2, 3]]  # (batch_size, seq_len) 用字典中的索引(数字)表示来表示hello
y_data = [3, 1, 2, 3, 2]  # (batch_size * seq_len) 标签:ohlol

inputs = torch.LongTensor(x_data)  # (batch_size, seq_len)
labels = torch.LongTensor(y_data)  # (batch_size * seq_len)


# 3、构建模型
class Model(torch.nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.emb = torch.nn.Embedding(num_class, embedding_size)
        self.rnn = torch.nn.RNN(input_size=embedding_size, hidden_size=hidden_size, num_layers=num_layers,
                                batch_first=True)
        self.fc = torch.nn.Linear(hidden_size, num_class)

    def forward(self, x):
        hidden = torch.zeros(num_layers, x.size(0), hidden_size)  # (num_layers, batch_size, hidden_size)
        x = self.emb(x)  # 返回(batch_size, seq_len, embedding_size)
        x, _ = self.rnn(x, hidden)  # 返回(batch_size, seq_len, hidden_size)
        x = self.fc(x)  # 返回(batch_size, seq_len, num_class)
        return x.view(-1, num_class)  # (batch_size * seq_len, num_class)


net = Model()

# 4、损失和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.05)  # Adam优化器

# 5、训练
for epoch in range(15):
    optimizer.zero_grad()
    outputs = net(inputs)
    loss = criterion(outputs, labels)
    loss.backward()
    optimizer.step()

    _, idx = outputs.max(dim=1)
    idx = idx.data.numpy()
    print('Predicted string: ', ''.join([index2char[x] for x in idx]), end='')
    print(', Epoch [%d/15] loss: %.4f' % (epoch + 1, loss.item()))

11.6 RNN分类器实现人名国家对应

import csv
import time
import matplotlib.pyplot as plt
import numpy as np
import math
import gzip  # 用于读取压缩文件
import torch
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from torch.nn.utils.rnn import pack_padded_sequence

# 1.超参数设置
HIDDEN_SIZE = 100
BATCH_SIZE = 256
N_LAYER = 2  # RNN的层数
N_EPOCHS = 100
N_CHARS = 128  # ASCII码的个数
USE_GPU = False


# 工具类函数
# 把名字转换成ASCII码,返回ASCII码值列表和名字的长度
def name2list(name):
    arr = [ord(c) for c in name]
    return arr, len(arr)


# 是否把数据放到GPU上
def create_tensor(tensor):
    if USE_GPU:
        device = torch.device('cuda:0')
        tensor = tensor.to(device)
    return tensor


def timesince(since):
    now = time.time()
    s = now - since
    m = math.floor(s / 60)  # math.floor()向下取整
    s -= m * 60
    return '%dmin %ds' % (m, s)  # 多少分钟多少秒


# 2.构建数据集
class NameDataset(Dataset):
    def __init__(self, is_train=True):
        # 文件读取
        filename = '../dataset/names_train.csv.gz' if is_train else '../dataset/names_test.csv.gz'
        with gzip.open(filename, 'rt') as f:  # rt表示以只读模式打开文件,并将文件内容解析为文本形式
            reader = csv.reader(f)
            rows =list(reader)  # 每个元素由一个名字和国家组成
        # 提取属性
        self.names = [row[0] for row in rows]
        self.len = len(self.names)
        self.countries = [row[1] for row in rows]
        # 编码处理
        self.country_list = list(sorted(set(self.countries)))  # 列表,按字母表顺序排序,去重后有18个国家名
        self.country_dict = self.get_countries_dict()  # 字典,国家名对应序号标签
        self.country_num = len(self.country_list)

    def __getitem__(self, item):
        # 索引获取
        return self.names[item], self.country_dict[self.countries[item]]  # 根据国家去字典查找索引

    def __len__(self):
        # 获取个数
        return self.len

    def get_countries_dict(self):
        # 根据国家名对应序号
        country_dict = dict()
        for idx, country_name in enumerate(self.country_list):
            country_dict[country_name] = idx
        return country_dict

    def idx2country(self, index):
        # 根据索引返回国家名字
        return self.country_list[index]

    def get_countries_num(self):
        # 返回国家名个数(分类的总个数)
        return self.country_num


# 3.实例化数据集
train_set = NameDataset(is_train=True)
train_loader = DataLoader(train_set, shuffle=True, batch_size=BATCH_SIZE, num_workers=2)
test_set = NameDataset(is_train=False)
test_loder = DataLoader(test_set, shuffle=False, batch_size=BATCH_SIZE, num_workers=2)
N_COUNTRY = train_set.get_countries_num()  # 18个国家名,即18个类别


# 4.模型构建
class GRUClassifier(torch.nn.Module):
    def __init__(self, input_size, hidden_size, output_size, n_layers=1, bidirectional=True):
        super(GRUClassifier, self).__init__()
        self.hidden_size = hidden_size
        self.n_layers = n_layers
        self.n_directions = 2 if bidirectional else 1

        # 词嵌入层,将词语映射到hidden维度
        self.embedding = torch.nn.Embedding(input_size, hidden_size)
        # GRU层(输入为特征数,这里是embedding_size,其大小等于hidden_size))
        self.gru = torch.nn.GRU(hidden_size, hidden_size, num_layers=n_layers, bidirectional=bidirectional)
        # 线性层
        self.fc = torch.nn.Linear(hidden_size * self.n_directions, output_size)

    def _init_hidden(self, bath_size):
        # 初始化权重,(n_layers * num_directions 双向, batch_size, hidden_size)
        hidden = torch.zeros(self.n_layers * self.n_directions, bath_size, self.hidden_size)
        return create_tensor(hidden)

    def forward(self, input, seq_lengths):
        # 转置 B X S -> S X B
        input = input.t()  # 此时的维度为seq_len, batch_size
        batch_size = input.size(1)
        hidden = self._init_hidden(batch_size)

        # 嵌入层处理 input:(seq_len,batch_size) -> embedding:(seq_len,batch_size,embedding_size)
        embedding = self.embedding(input)

        # 进行打包(不考虑0元素,提高运行速度)需要将嵌入数据按长度排好
        gru_input = pack_padded_sequence(embedding, seq_lengths)

        # output:(*, hidden_size * num_directions),*表示输入的形状(seq_len,batch_size)
        # hidden:(num_layers * num_directions, batch, hidden_size)
        output, hidden = self.gru(gru_input, hidden)
        if self.n_directions == 2:
            hidden_cat = torch.cat([hidden[-1], hidden[-2]], dim=1)  # hidden[-1]的形状是(1,256,100),hidden[-2]的形状是(1,256,100),拼接后的形状是(1,256,200)
        else:
            hidden_cat = hidden[-1]  # (1,256,100)
        fc_output = self.fc(hidden_cat)
        return fc_output


# 3.数据处理(姓名->数字)
def make_tensors(names, countries):
    # 获取嵌入长度从大到小排序的seq_tensor(嵌入向量)、seq_lengths(对应长度)、countries(对应顺序的国家序号)-> 便于pack_padded_sequence处理
    name_len_list = [name2list(name) for name in names]  # 每个名字对应的1列表
    name_seq = [sl[0] for sl in name_len_list]  # 姓名列表
    seq_lengths = torch.LongTensor([sl[1] for sl in name_len_list])  # 名字对应的字符个数
    countries = countries.long()   # PyTorch 中,张量的默认数据类型是浮点型 (float),这里转换成整型,可以避免浮点数比较时的精度误差,从而提高模型的训练效果

    # 创建全零张量,再依次进行填充
    # 创建了一个 len(name_seq) * seq_length.max()维的张量
    seq_tensor = torch.zeros(len(name_seq), seq_lengths.max()).long()
    for idx, (seq, seq_len) in enumerate(zip(name_seq, seq_lengths)):
        seq_tensor[idx, :seq_len] = torch.LongTensor(seq)

    # 为了使用pack_padded_sequence,需要按照长度排序
    # perm_idx是排序后的数据在原数据中的索引,seq_tensor是排序后的数据,seq_lengths是排序后的数据的长度,countries是排序后的国家
    seq_lengths, perm_idx = seq_lengths.sort(dim=0, descending=True)  # descending=True 表示按降序进行排序,即从最长的序列到最短的序列。
    seq_tensor = seq_tensor[perm_idx]
    countries = countries[perm_idx]

    return create_tensor(seq_tensor), create_tensor(seq_lengths), create_tensor(countries)


# 训练循环
def train(epoch, start):
    total_loss = 0
    for i, (names, countries) in enumerate(train_loader, 1):
        inputs, seq_lengths, target = make_tensors(names, countries)  # 输入、每个序列长度、输出
        output = model(inputs, seq_lengths)
        loss = criterion(output, target)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        total_loss += loss.item()
        if i % 10 == 0:
            print(f'[{timesince(start)}] Epoch {epoch} ', end='')
            print(f'[{i * len(inputs)}/{len(train_set)}] ', end='')
            print(f'loss={total_loss / (i * len(inputs))}')  # 打印每个样本的平均损失

    return total_loss


# 测试循环
def test():
    correct = 0
    total = len(test_set)
    print('evaluating trained model ...')
    with torch.no_grad():
        for i, (names, countries) in enumerate(test_loder, 1):
            inputs, seq_lengths, target = make_tensors(names, countries)
            output = model(inputs, seq_lengths)
            pred = output.max(dim=1, keepdim=True)[1]  # 返回每一行中最大值的那个元素的索引,且keepdim=True,表示保持输出的二维特性
            correct += pred.eq(target.view_as(pred)).sum().item()  # 计算正确的个数
        percent = '%.2f' % (100 * correct / total)
        print(f'Test set: Accuracy {correct}/{total} {percent}%')

    return correct / total  # 返回的是准确率,0.几几的格式,用来画图


if __name__ == '__main__':
    model = GRUClassifier(N_CHARS, HIDDEN_SIZE, N_COUNTRY, N_LAYER)
    criterion = torch.nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(), lr=0.001)
    device = 'cuda:0' if USE_GPU else 'cpu'
    model.to(device)
    start = time.time()
    print('Training for %d epochs...' % N_EPOCHS)
    acc_list = []
    # 在每个epoch中,训练完一次就测试一次
    for epoch in range(1, N_EPOCHS + 1):
        # Train cycle
        train(epoch, start)
        acc = test()
        acc_list.append(acc)

    # 绘制在测试集上的准确率
    epoch = np.arange(1, len(acc_list) + 1)
    acc_list = np.array(acc_list)
    plt.plot(epoch, acc_list)
    plt.xlabel('Epoch')
    plt.ylabel('Accuracy')
    plt.grid()
    plt.show()
  • 41
    点赞
  • 77
    收藏
    觉得还不错? 一键收藏
  • 打赏
    打赏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

helloworld大王

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值