代码笔记

源代码 

import torch
import matplotlib.pyplot as plt
import random
from utils import *
%matplotlib inline

def synthtic_data(w,b,num_examples):
    X = torch.normal(0,0.5,(num_examples,len(w)))
    y = torch.matmul(X,w) + b
    return X,y.reshape((-1,1))

true_w = torch.tensor([-3.,-4.2])
true_b = 3.4
features,labels = synthtic_data(true_w,true_b,1000)
print(f"features: {features[0]},\nlabels: {labels[0]}")
plt.scatter(features[:, (1)].detach().numpy(), labels.detach().numpy(), 1)

def data_iter(batch_size,features,labels):
    num_examples = len(features)
    indices = list(range(num_examples))
    random.shuffle(indices)
    for i in range(0,num_examples,batch_size):
        batch_indices = torch.tensor(indices[i:min(i+batch_size,num_examples)])
        yield features[batch_indices], labels[batch_indices]

batch_size = 10
for X,y in data_iter(batch_size,features,labels):
    print(X,"\n",y)
    break

w = torch.normal(0,0.01,size = (2,1),requires_grad=True)
b = torch.zeros(1,requires_grad=True)

def linreg(X,w,b):
    return torch.matmul(X,w) + b
def squared_loss(y_hat,y):
    return (y-y_hat)**2 / 2
def sgd(params,lr,batch_size):
    with torch.no_grad():
        for param in params:
            param -= lr * param.grad/batch_size
            param.grad.zero_()

lr  = 0.01
num_epochs = 10
net = linreg
loss = squared_loss
for epoch in range(num_epochs):
    for X,y in data_iter(batch_size,features,labels):
        l  = loss(net(X,w,b),y)
        l.sum().backward()
        sgd([w,b],lr,batch_size)
    with torch.no_grad():
        train_l = loss(net(features,w,b),labels)
        print(f"epoch: {epoch + 1},train_l: {train_l.mean()}")

具体展开

概述

自己选取真值，基于其生成了一群数据，并用一个简单且直感的回归模型，以验证线性回归模型有效性

对数据的准备 

def synthtic_data(w,b,num_examples):
    X = torch.normal(0,0.5,(num_examples,len(w)))
    y = torch.matmul(X,w) + b
    return X,y.reshape((-1,1))
  • 对尺寸的选择：torch与numpy里一般将矩阵形状定义为(样本数量,特征数量)（每行为一个样本）（X[i]）来获得第i个样本）
  • 有专门的矩阵乘法计算函数torch.matnul()

人造数据 

true_w = torch.tensor([-3.,-4.2])
true_b = 3.4
features,labels = synthtic_data(true_w,true_b,1000)
  • torch.tensor([-3,-4.2])这里造出的是一维张量，可以被按需求解释为行或列向量。

准备dataloader 

def data_iter(batch_size,features,labels):
    num_examples = len(features)
    indices = list(range(num_examples))
    random.shuffle(indices)
    for i in range(0,num_examples,batch_size):
        batch_indices = torch.tensor(indices[i:min(i+batch_size,num_examples)])
        yield features[batch_indices], labels[batch_indices]

batch_size = 10
  • 青春版dataloader的严重问题：它要求我们将所有数据加载到内存中，并执行大量的随机内存访问
用torch提供的dataloader 
from torch.utils import data
def load_array(data_arrays,batch_size,is_train = True):
    dataset = data.TensorDataset(*data_arrays)
    return data.DataLoader(dataset,batch_size,shuffle = is_train)

batch_size = 10
data_iter = load_array((features,labels),batch_size)

（手动封装一个更自定义化的dataloader而已）

大概顺序：原始的样本矩阵需要先转化为torch的dataset对象（如初始化__init__等） 然后基于dataset制作dataloader（如加入shuffle设置，batch_size）设置等 最后对接模型的是dataloader

一个对其效果的简单验证

type(data_iter)查return得到的数据类型，是torch.utils.data.dataloader.DataLoader

next(iter(data_iter))，得到的是形如的存在：

[tensor([[ 0.5695, -0.1118],
         [ 0.1460, -0.3379],
         [-0.8529,  0.1165]]),
 tensor([[ 2.1610e+00],
         [ 4.3810e+00],
         [ 5.4694e+00]])]

可见与之前的青春版dataloader的效果是一样的

另： iter()将iteratable转换为iterator next()提取iterator的当前位置数据，并将其移动到下一个位置

准备参数与模型 

w = torch.normal(0, 0.01, size=(2,1), requires_grad=True)
b = torch.zeros(1, requires_grad=True)

用torch建立参数形状与性质，requires_grad是在这里声明的

处理优化函数 

def sgd(params, lr, batch_size):  
    with torch.no_grad():
        for param in params:
            param -= lr * param.grad / batch_size
            param.grad.zero_()

梯度更新的跟踪：需要跟踪的，如正向传播，损失计算

不需要跟踪的，如参数更新，模型评估

训练 

lr = 0.03
num_epochs = 3
net = linreg
loss = squared_loss

for epoch in range(num_epochs):
    for X, y in data_iter(batch_size, features, labels):
        l = loss(net(X, w, b), y)  
        l.sum().backward()
        sgd([w, b], lr, batch_size) 
    with torch.no_grad():
        train_l = loss(net(features, w, b), labels)
        print(f'epoch {epoch + 1}, loss {float(train_l.mean()):f}')

说些什么吧！

waline