在多分类任务实验中手动实现dropout

9、在多分类任务实验中手动实现dropout

import torch
import torch.nn as nn
import numpy as np
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt

#读取数据
mnist_train = torchvision.datasets.MNIST(root='./data', train=True, download=False, transform=transforms.ToTensor())  
mnist_test = torchvision.datasets.MNIST(root='./data', train=False,download=False, transform=transforms.ToTensor())  

  
batch_size = 256 
train_iter = torch.utils.data.DataLoader(mnist_train, batch_size=batch_size, shuffle=True,num_workers=0)  
test_iter = torch.utils.data.DataLoader(mnist_test, batch_size=batch_size, shuffle=False,num_workers=0)  

#初始化参数  
num_inputs,num_hiddens,num_outputs =784, 256,10
num_epochs=30
lr = 0.001
def init_param():
    W1 = torch.tensor(np.random.normal(0, 0.01, (num_hiddens,num_inputs)), dtype=torch.float32)  
    b1 = torch.zeros(1, dtype=torch.float32)  
    W2 = torch.tensor(np.random.normal(0, 0.01, (num_outputs,num_hiddens)), dtype=torch.float32)  
    b2 = torch.zeros(1, dtype=torch.float32)  
    params =[W1,b1,W2,b2]
    for param in params:  
        param.requires_grad_(requires_grad=True)  
    return W1,b1,W2,b2

#手动定义dropout函数
def dropout(X, drop_prob): #drop_porb是一个概率值，介于0-1，代表要丢弃多少比例的神经元
    X = X.float()
    assert 0 <= drop_prob <= 1
    keep_prob = 1 - drop_prob
    if keep_prob == 0:
        return torch.zeros_like(X)
    mask = (torch.rand(X.shape) < keep_prob).float()
    return mask * X / keep_prob

#定义模型
def net(X, is_training=True):
    X = X.view(-1, num_inputs)
    H1 = (torch.matmul(X, W1.t()) + b1).relu()
    if is_training:      #只有在训练模式下才需要dropout，测试情况下不需要
        H1 = dropout(H1, drop_prob1)
    return (torch.matmul(H1,W2.t()) + b2).relu()

#定义训练函数
def train(net,train_iter,test_iter,loss,num_epochs,batch_size,lr=None,optimizer=None):
    train_ls, test_ls = [], []
    for epoch in range(num_epochs):
        ls, count = 0, 0
        for X,y in train_iter:
            l=loss(net(X),y)
            optimizer.zero_grad()
            l.backward()
            optimizer.step()
            ls += l.item()
            count += y.shape[0]
        train_ls.append(ls)
        ls, count = 0, 0
        for X,y in test_iter:
            l=loss(net(X,is_training=False),y)
            ls += l.item()
            count += y.shape[0]
        test_ls.append(ls)
        if(epoch+1)%10==0:
            print('epoch: %d, train loss: %f, test loss: %f'%(epoch+1,train_ls[-1],test_ls[-1]))
    return train_ls,test_ls

#定义drop从0到1，训练十次，观察不同drop对结果的影响
drop_probs = np.arange(0,1.1,0.1)
Train_ls, Test_ls = [], []
#开始训练
for drop_prob in drop_probs:
    drop_prob1 = drop_prob
    W1,b1,W2,b2 = init_param()
    loss = nn.CrossEntropyLoss()
    optimizer = torch.optim.SGD([W1,b1,W2,b2],lr = 0.001)
    train_ls, test_ls =  train(net,train_iter,test_iter,loss,num_epochs,batch_size,lr,optimizer)   
    Train_ls.append(train_ls)
    Test_ls.append(test_ls)

#结果可视化
x = np.linspace(0,len(train_ls),len(train_ls))
plt.figure(figsize=(10,8))
for i in range(0,len(drop_probs)):
    plt.plot(x,Train_ls[i],label= 'drop_prob=%.1f'%(drop_probs[i]),linewidth=1.5)
    plt.xlabel('epoch')
    plt.ylabel('loss')
# plt.legend()
plt.legend(loc=2, bbox_to_anchor=(1.05,1.0),borderaxespad = 0.)
plt.title('train loss with dropout')
plt.show()