Pytorch 线性回归（下）【Pytorch 基础第四话】

数据归一化

假如我们不对数据进行归一化，数据会在损失曲线下降到一般的时候就不再快速下降了。因为再快速下降的时候，是因为某个特征带来的红利，但一旦红利耗尽之后，再次迭代就不再那么明显了，会十分缓慢。
如图：

对数据进行归一化之后再进行线性回归：
归一化手段，示例使用均值归一化，公式如下：
$$X:=\frac{X-avg(X)}{range}$$
这样可以把X控制在 $-0.5<X<0.5$

import torch
from debug import ptf_tensor

x = torch.tensor([[1, 1, 1], [2, 3, 1], [3, 5, 1],
        [4, 2, 1], [5, 4, 1]], dtype=torch.float32)
y = torch.tensor([-10, 12, 14, 16, 18],
        dtype=torch.float32)

x_mean, x_std = torch.mean(x, dim=0), torch.std(x, dim=0)
x_mean[-1], x_std[-1] = 0, 1 #因为所有组的x的最后一个元素都是1，为了避免出现nan
x_norm = (x - x_mean) / x_std

ptf_tensor(x_mean,'mean')
ptf_tensor(x_std,'std')
ptf_tensor(x_norm,'norm')

输出：

The info of mean:
#############################
  @dims: 1
  @size: torch.Size([3])
  @ele_sum: 3
  @dtype: torch.float32
  @data:
tensor([3., 3., 0.])
#############################


The info of std:
#############################
  @dims: 1
  @size: torch.Size([3])
  @ele_sum: 3
  @dtype: torch.float32
  @data:
tensor([1.5811, 1.5811, 1.0000])
#############################


The info of norm:
#############################
  @dims: 2
  @size: torch.Size([5, 3])
  @ele_sum: 15
  @dtype: torch.float32
  @data:
tensor([[-1.2649, -1.2649,  1.0000],
        [-0.6325,  0.0000,  1.0000],
        [ 0.0000,  1.2649,  1.0000],
        [ 0.6325, -0.6325,  1.0000],
        [ 1.2649,  0.6325,  1.0000]])
#############################

我们对之前的线性回归进行归一化，可以缩小迭代次数。之前我们都是用30001次，现在我们只需要10001次。

import torch
import torch.nn
import torch.optim

x = torch.tensor([[1000000, 0.0001], [2000000, 0.0003],
        [3000000, 0.0005], [4000000, 0.0002], [5000000, 0.0004]])
y = torch.tensor([-1000., 1200., 1400., 1600., 1800.]).reshape(-1, 1)

x_mean, x_std = torch.mean(x,dim=0), torch.std(x,dim=0)
x_norm = (x-x_mean)/x_std

y_mean, y_std = torch.mean(y,dim=0), torch.std(y,dim=0)
y_norm=(y-y_mean)/y_std

fc=torch.nn.Linear(2,1)
criterion=torch.nn.MSELoss()
optimizer=torch.optim.Adam(fc.parameters())


for step in range(10001):
    if step:
        optimizer.zero_grad()
        loss_norm.backward()
        optimizer.step()
    
    pred_norm=fc(x_norm)
    loss_norm=criterion(pred_norm,y_norm)
    pred=pred_norm * y_std + y_mean
    loss = criterion (pred,y)
    if step % 1000 ==0:
        print('step = {}, loss = {:g}'.format(step, loss))

输出：

step = 0, loss = 2.25879e+06
step = 1000, loss = 292322
step = 2000, loss = 216877
step = 3000, loss = 213382
step = 4000, loss = 213333
step = 5000, loss = 213333
step = 6000, loss = 213333
step = 7000, loss = 213333
step = 8000, loss = 213333
step = 9000, loss = 213333
step = 10000, loss = 213333

Demo: 世界人口数据线性回归实验

（1）从CSV文件获取数据

url='C:/Users/HUAWEI/Desktop/深度学习/Blog附带代码/population.csv'
df = pd.read_csv(url, index_col=0)
#print(df) #数据预览

years=torch.tensor(df.iloc[:,0],dtype=torch.float32) # 取出第0列数据作为Years
populations=torch.tensor(df.iloc[:,1],dtype=torch.float32) #取出第1列数据作为人口数据

（2）Reshape

x=years.reshape(-1,1) # reshpae(-1,1)等价于reshape(n,1),因为大多时候我们都不知道所有元素的总数
ptf_tensor(x,'years reshape',)
y=populations
'''
这是我们需要的数据形式:
    X=[ [,,] , [,,] , [,,] ] # X=m*n m是样本数量，n是特征数
    y=[,,,] # Y=m*1
    W=[,,] # W=n*1
Y=XW
'''

（3）归一化

def get_mean_std_norm(x):
    x_mean,x_std=torch.mean(x),torch.std(x)
    x_norm=(x-x_mean)/x_std
    return x_mean,x_std,x_norm

x_mean,x_std,x_norm=get_mean_std_norm(x)
y_mean,y_std,y_norm=get_mean_std_norm(y)

（4）建造四大组件：线性容器，标准容器，梯度容器，权值浏览器

Linear，criterion，optimizer，weights和bias
分别用于：

• 计算函数
• 计算当前Loss
• 更新梯度
• 浏览计算过程中的中间结果

fc=torch.nn.Linear(1,1) #Y=XW式子中y的数量和x的数量
criterion=torch.nn.MSELoss()
optimizer=torch.optim.Adam(fc.parameters())

weights_norm,bias_norm=fc.parameters() #每次结果生成器

（5）迭代

for step in range(20001):
    if step:
        optimizer.zero_grad()
        loss_norm.backward()
        optimizer.step()
    
    output_norm=fc(x_norm)
    pred_norm=output_norm.squeeze()
    loss_norm=criterion(pred_norm,y_norm) # 统一size才能计算loss

    weights=y_std/x_std *weights_norm
    bias=(weights_norm*(0-x_mean)/x_std+bias_norm)*y_std+y_mean
    if step % 5000==0:
        print('step={}, weights={}, loss={}'.format(step,weights.item(),loss_norm))

（6）绘图

plt.scatter(years, populations, s=0.1, label='actual', color='k')
plt.plot(years.tolist(), (years*weights + bias).squeeze(dim=0).tolist(), label='result', color='k') 
#注意：这里years.size=[67],result.size=[1,67],所以要对result进行squeeze

plt.xlabel('Year')
plt.ylabel('Population')
plt.legend()
plt.show()

Demo 完整代码：

import pandas as pd
import matplotlib.pyplot as plt

import torch
import torch.nn
import torch.optim

from debug import ptf_tensor

url='C:/Users/HUAWEI/Desktop/深度学习/Blog附带代码/population.csv'
df = pd.read_csv(url, index_col=0)
#print(df) #数据预览

years=torch.tensor(df.iloc[:,0],dtype=torch.float32) # 取出第0列数据作为Years
populations=torch.tensor(df.iloc[:,1],dtype=torch.float32) #取出第1列数据作为人口数据

ptf_tensor(years,'years',data_only=True) #预览数据
ptf_tensor(populations,'populations',data_only=True) #预览数据

x=years.reshape(-1,1) # reshpae(-1,1)等价于reshape(n,1),因为大多时候我们都不知道所有元素的总数
ptf_tensor(x,'years reshape',)
y=populations
'''
这是我们需要的数据形式:
    X=[ [,,] , [,,] , [,,] ] # X=m*n m是样本数量，n是特征数
    y=[,,,] # Y=m*1
    W=[,,] # W=n*1
Y=XW
'''
def get_mean_std_norm(x):
    x_mean,x_std=torch.mean(x),torch.std(x)
    x_norm=(x-x_mean)/x_std
    return x_mean,x_std,x_norm

x_mean,x_std,x_norm=get_mean_std_norm(x)
y_mean,y_std,y_norm=get_mean_std_norm(y)

fc=torch.nn.Linear(1,1) #Y=XW式子中y的数量和x的数量
criterion=torch.nn.MSELoss()
optimizer=torch.optim.Adam(fc.parameters())

weights_norm,bias_norm=fc.parameters() #每次结果生成器

for step in range(20001):
    if step:
        optimizer.zero_grad()
        loss_norm.backward()
        optimizer.step()
    
    output_norm=fc(x_norm)
    pred_norm=output_norm.squeeze()
    loss_norm=criterion(pred_norm,y_norm) # 统一size才能计算loss

    weights=y_std/x_std *weights_norm
    bias=(weights_norm*(0-x_mean)/x_std+bias_norm)*y_std+y_mean
    if step % 5000==0:
        print('step={}, weights={}, loss={}'.format(step,weights.item(),loss_norm))


plt.scatter(years, populations, s=0.1, label='actual', color='k')
plt.plot(years.tolist(), (years*weights + bias).squeeze(dim=0).tolist(), label='result', color='k') 
#注意：这里years.size=[67],result.size=[1,67],所以要对result进行squeeze

plt.xlabel('Year')
plt.ylabel('Population')
plt.legend()
plt.show()

输出：

step=0, weights=-3741121.0, loss=1.0858874320983887
step=5000, weights=75291096.0, loss=0.0031296168453991413
step=10000, weights=75291232.0, loss=0.0031296154484152794
step=15000, weights=75291248.0, loss=0.0031296159140765667
step=20000, weights=75291248.0, loss=0.0031296173110604286