softmax实现

用PyTorch实现linear模型

模拟数据集

模拟数据集这块与上节课一致

num_inputs = 2 #feature number
num_examples = 1000 #训练样本个数
true_w = torch.tensor([[2],[-3.4]]) #真实的权重值
true_b = torch.tensor(4.2) #真实的bias
samples = torch.normal(0,1,(num_examples,num_inputs))
noise = torch.normal(0,0.01,(num_examples,1))
labels = samples.matmul(true_w) + true_b + noise

定义模型

class LinearNet(nn.Module):
	def __init__(self,in_features):
		super().__init__()
		self.fc = nn.Linear(in_features=2,out_features=1)
	def forward(self,t):
		t = self.fc(t)
		return t

加载数据集

import torch.utils.data as Data
dataset = Data.TensorDataset(samples,labels)#类似于zip，把两个张量打包
data_loader = Data.DataLoader(dataset,batch_size=100,shuffle=True)

optimizer

network = LinearNet(2)
optimizer = optim.SGD(network.paramters(),lr=0.05)

模型训练

for epoch in range(10):
    total_loss = 0
    for data,label in data_loader:
        predict = network(data)
        loss = F.mse_loss(predict,label)
        total_loss += loss.item()
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
    print(
        'epoch',epoch,
        'loss',total_loss,
        'weight',network.weight,
        'bias',network.bias
    )

softmax回归模型

sotfmax主要用于分类任务。regression最终得到的是一个scalar，根据input中的feature线性相加得到一个output。分类任务的结果是一个类别，是离散的。
假设现在有一批图片是2 * 2大小的灰度图片，这样图片中的每隔二像素用一个标量表示就行了。这批图片一种是三类小动物，第一类是小狗，第二类是小猫，第三类是小兔子。
每张图片总共4个像素点，我们可以看作是4个feature，假设这三类小动物的图片线性可分，每一类对应一组weight和一个bias。
${\delta }_1=w_{11}{x}_1+w_{12}{x}_2+w_{13}{x}_3+w_{14}{x}_4+b_1$
${\delta }_2=w_{21}{x}_1+w_{22}{x}_2+w_{23}{x}_3+w_{24}{x}_4+b_2$
${\delta }_3=w_{31}{x}_1+w_{32}{x}_2+w_{33}{x}_3+w_{34}{x}_4+b_3$
$\left[\begin{array}{c}{\delta }_1\\ {\delta }_2\\ {\delta }_3\end{array}\right]$ $=\left[\begin{array}{cccc}{w}_{11}& w_{12}& w_{13}& w_{14}\\ w_{21}& w_{22}& w_{23}& w_{24}\\ w_{31}& w_{32}& w_{33}& w_{34}\end{array}\right]$$\left[\begin{array}{c}{x}_1\\ x_2\\ x_3\\ x_4\end{array}\right]$+$\left[\begin{array}{c}{b}_1\\ b_2\\ b_3\end{array}\right]$

可以根据输出值较大的来决定哪一类，可这样有个问题，首先输出值没有明确的意义，且可能是实数范围。其次，不好衡量输出值与真实值之间的差距。所以采用softmax操作，将三个输出值转化成概率值，这样输出结果满足概率分布。label采用one-hot编码，相当于对应类别的概率是1，这样就可以用cross_entropy来计算loss。

Fashion-MNIST

本次学习softmax模型采用torchvision.datasets中的Fashion-MNIST。

import torchvision
import torchvision.transforms as transforms
train_set = torchvision.datasets.FashionMNIST(
	root='./data',
	train=True,
	download=True,
	transform=transforms.ToTensor()
)

transforms.ToTensor()将尺寸为（H x W x C）且数据位于（0，255）的PIL图片或者数据类型为np.uint8的NumPy数组转换为尺寸为C x H x W且数据类型为torch.float32且位于（0.0，1.0）的Tensor

len(train_set),len(test_set)
> (60000,10000)

展示一下数据集中的图片

import matplotlib.pyplot as plt
plt.figure(figsize=(10,10))
for i,(image,lable) in enumerate(train_set,start=1):
	plt.subplot(1,10,i)
	plt.imshow(image.squeeze())
	plt.title(train_set.classes[lable])
	plt.axis('off')
	if i == 10:
		break
plt.show()

train_loader = torch.utils.data.DataLoader(train_set,batch_size=100,shuffle=True,num_workers=4)
test_loader = torch.utils.data.DataLoader(test_set,batch_size=100,shuffle=False,num_workers=1)

cross_entropy

def cross_entropy(predict,labels):
	labels = torch.LongTensor(labels).view(-1,1)
	loss = -torch.log(predict.gather(1,labels)
	return loss.mean()

多分类问题的交叉熵函数如下：

$loss=\frac{1}{n}{\sum }_1^{n}log{p}^i$
i表示label对应的种类，pⁱ为真实种类的预测概率，log是以e为底的对数
这里gather函数的作用，就是在predict上取到对应label的概率值，注意负号不能丢，pytorch中的cross_entropy对输入先进行一次softmax操作，以保证输入都是正的。

模型的实现

def net(samples,w,b):
	samples = samples.flatten(start_dim=1) #将c,h,w三个轴展成一个feature轴，长度为28 * 28
	samples = torch.exp(samples)#全体元素取以e为底的指数
	partial_sum = samples.sum(dim=1,keepdim=True) 
	samples = samples / partial_sum #归一化，得概率，这里还应用了广播机制
	return samples.matmul(w) + b	

利用PyTorch简易实现softmax

import torch
import torchvision
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data as Data
import torchvision.transforms as transforms
import torch.optim as optim
import torch.nn.init as init

class SoftmaxNet(nn.Module):
    def __init__(self,in_features,out_features):
        super().__init__()
        self.fc = nn.Linear(in_features=in_features,out_features=out_features)
    def forward(self,t):
        t = t.flatten(start_dim=1)
        t = self.fc(t)
        return t

train_set = torchvision.datasets.FashionMNIST(
    root='E:\project\python\jupyterbook\data',
    train=True,
    download=True,
    transform=transforms.ToTensor()
)

test_set = torchvision.datasets.FashionMNIST(
    root='E:\project\python\jupyterbook\data',
    train=False,
    download=True,
    transform=transforms.ToTensor()
)

train_loader = Data.DataLoader(
    train_set,
    batch_size=100,
    shuffle=True,
    #num_workers=2
)
test_loader = Data.DataLoader(
    test_set,
    batch_size=100,
    shuffle=False,
    #num_workers=2
)

@torch.no_grad()
def get_correct_nums(predict,labels):
    return predict.argmax(dim=1).eq(labels).sum().item()

@torch.no_grad()
def evaluate(test_loader,net,total_num):
    correct = 0
    for image,label in test_loader:
        predict = net(image)
        correct += get_correct_nums(predict,label)
        pass
    return correct / total_num

network = SoftmaxNet()
optimizer = optim.SGD(network.parameters(),lr=0.05)

for epoch in range(10):
    total_loss = 0
    total_correct = 0
    for image,label in train_loader:
        predict = network(image)
        loss = F.cross_entropy(predict,label)
        total_loss += loss.item()
        total_correct += get_correct_nums(predict,label)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        pass

    print(
        'epoch',epoch,
        'loss',total_loss,
        'train_acc',total_correct / len(train_set),
        'test_acc',evaluate(test_loader,network,len(test_set))
    )