任务6 Pytorch理解更多神经网络优化方法

于 2019-08-18 15:56:57 发布
阅读量356 收藏 1
点赞数
分类专栏： 机器学习
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.youkuaiyun.com/qinhanmin2010/article/details/99705456
版权
本文深入探讨了Pytorch中常用的神经网络优化算法，包括SGD的梯度下降法，Momentum的动量优化，RMSprop的均方根propagation，以及Adam的自适应矩估计优化。这些优化技术对于提升模型训练效率和性能至关重要。

摘要生成于 C知道 ，由 DeepSeek-R1 满血版支持， 前往体验 >

  • SGD
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms

# Hyper-parameters 
input_size = 784
hidden_size = 500
num_classes = 10
num_epochs = 5
batch_size = 100
learning_rate = 0.001
SGD_loss = []

# MNIST dataset 
train_dataset = torchvision.datasets.MNIST(root='data/', 
                                           train=True, 
                                           transform=transforms.ToTensor(),  
                                           download=True)

test_dataset = torchvision.datasets.MNIST(root='data/', 
                                          train=False, 
                                          transform=transforms.ToTensor())

# Data loader
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, 
                                           batch_size=batch_size, 
                                           shuffle=True)

test_loader = torch.utils.data.DataLoader(dataset=test_dataset, 
                                          batch_size=batch_size, 
                                          shuffle=False)

# Fully connected neural network with one hidden layer
class NeuralNet(nn.Module):
    def __init__(self, input_size, hidden_size, num_classes):
        super(NeuralNet, self).__init__()
        self.fc1 = nn.Sequential(nn.Linear(input_size, hidden_size),
                                 nn.ReLU())
        self.fc2 = nn.Linear(hidden_size, num_classes)  
    
    def forward(self, x):
        out = self.fc1(x)
        out = self.fc2(out)
        return out

model = NeuralNet(input_size, hidden_size, num_classes)

# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)  

# Train the model
total_step = len(train_loader)
for epoch in range(num_epochs):
    for i, (images, labels) in enumerate(train_loader):  
        images = images.reshape(-1, 28 * 28)
        
        # Forward pass
        outputs = model(images)
        loss = criterion(outputs, labels)
        SGD_loss.append(loss)
        
        # Backward and optimize
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if (i + 1) % 100 == 0:
            print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}' 
                   .format(epoch+1, num_epochs, i + 1, total_step, loss.item()))

# Test the model
# In test phase, we don't need to compute gradients (for memory efficiency)
with torch.no_grad():
    correct = 0
    total = 0
    for images, labels in test_loader:
        images = images.reshape(-1, 28 * 28)
        outputs = model(images)
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

    print('Accuracy of the network on the 10000 test images: {} %'.format(100 * correct / total))

Epoch [1/5], Step [100/600], Loss: 2.2769
Epoch [1/5], Step [200/600], Loss: 2.2770
Epoch [1/5], Step [300/600], Loss: 2.2553
Epoch [1/5], Step [400/600], Loss: 2.2327
Epoch [1/5], Step [500/600], Loss: 2.2199
Epoch [1/5], Step [600/600], Loss: 2.1841
Epoch [2/5], Step [100/600], Loss: 2.1550
Epoch [2/5], Step [200/600], Loss: 2.1324
Epoch [2/5], Step [300/600], Loss: 2.0896
Epoch [2/5], Step [400/600], Loss: 2.0818
Epoch [2/5], Step [500/600], Loss: 2.0478
Epoch [2/5], Step [600/600], Loss: 2.0630
Epoch [3/5], Step [100/600], Loss: 2.0151
Epoch [3/5], Step [200/600], Loss: 2.0137
Epoch [3/5], Step [300/600], Loss: 1.9326
Epoch [3/5], Step [400/600], Loss: 1.8756
Epoch [3/5], Step [500/600], Loss: 1.8812
Epoch [3/5], Step [600/600], Loss: 1.8286
Epoch [4/5], Step [100/600], Loss: 1.7895
Epoch [4/5], Step [200/600], Loss: 1.7809
Epoch [4/5], Step [300/600], Loss: 1.7720
Epoch [4/5], Step [400/600], Loss: 1.7273
Epoch [4/5], Step [500/600], Loss: 1.6363
Epoch [4/5], Step [600/600], Loss: 1.7014
Epoch [5/5], Step [100/600], Loss: 1.6038
Epoch [5/5], Step [200/600], Loss: 1.5516
Epoch [5/5], Step [300/600], Loss: 1.5422
Epoch [5/5], Step [400/600], Loss: 1.5537
Epoch [5/5], Step [500/600], Loss: 1.5299
Epoch [5/5], Step [600/600], Loss: 1.4731
Accuracy of the network on the 10000 test images: 76.93 %
  • Momentum
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms

# Hyper-parameters 
input_size = 784
hidden_size = 500
num_classes = 10
num_epochs = 5
batch_size = 100
learning_rate = 0.001
momentum_loss = []

# MNIST dataset 
train_dataset = torchvision.datasets.MNIST(root='data/', 
                                           train=True, 
                                           transform=transforms.ToTensor(),  
                                           download=True)

test_dataset = torchvision.datasets.MNIST(root='data/', 
                                          train=False, 
                                          transform=transforms.ToTensor())

# Data loader
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, 
                                           batch_size=batch_size, 
                                           shuffle=True)

test_loader = torch.utils.data.DataLoader(dataset=test_dataset, 
                                          batch_size=batch_size, 
                                          shuffle=False)

# Fully connected neural network with one hidden layer
class NeuralNet(nn.Module):
    def __init__(self, input_size, hidden_size, num_classes):
        super(NeuralNet, self).__init__()
        self.fc1 = nn.Sequential(nn.Linear(input_size, hidden_size),
                                 nn.ReLU())
        self.fc2 = nn.Linear(hidden_size, num_classes)  
    
    def forward(self, x):
        out = self.fc1(x)
        out = self.fc2(out)
        return out

model = NeuralNet(input_size, hidden_size, num_classes)

# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, momentum=0.8)  

# Train the model
total_step = len(train_loader)
for epoch in range(num_epochs):
    for i, (images, labels) in enumerate(train_loader):  
        images = images.reshape(-1, 28 * 28)
        
        # Forward pass
        outputs = model(images)
        loss = criterion(outputs, labels)
        momentum_loss.append(loss)
        
        # Backward and optimize
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if (i + 1) % 100 == 0:
            print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}' 
                   .format(epoch+1, num_epochs, i + 1, total_step, loss.item()))

# Test the model
# In test phase, we don't need to compute gradients (for memory efficiency)
with torch.no_grad():
    correct = 0
    total = 0
    for images, labels in test_loader:
        images = images.reshape(-1, 28 * 28)
        outputs = model(images)
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

    print('Accuracy of the network on the 10000 test images: {} %'.format(100 * correct / total))

Epoch [1/5], Step [100/600], Loss: 2.2002
Epoch [1/5], Step [200/600], Loss: 2.0610
Epoch [1/5], Step [300/600], Loss: 1.9739
Epoch [1/5], Step [400/600], Loss: 1.6769
Epoch [1/5], Step [500/600], Loss: 1.6539
Epoch [1/5], Step [600/600], Loss: 1.4437
Epoch [2/5], Step [100/600], Loss: 1.2269
Epoch [2/5], Step [200/600], Loss: 1.1834
Epoch [2/5], Step [300/600], Loss: 0.9331
Epoch [2/5], Step [400/600], Loss: 0.9166
Epoch [2/5], Step [500/600], Loss: 0.8548
Epoch [2/5], Step [600/600], Loss: 0.8767
Epoch [3/5], Step [100/600], Loss: 0.7520
Epoch [3/5], Step [200/600], Loss: 0.7892
Epoch [3/5], Step [300/600], Loss: 0.7434
Epoch [3/5], Step [400/600], Loss: 0.6134
Epoch [3/5], Step [500/600], Loss: 0.6039
Epoch [3/5], Step [600/600], Loss: 0.7010
Epoch [4/5], Step [100/600], Loss: 0.5687
Epoch [4/5], Step [200/600], Loss: 0.5281
Epoch [4/5], Step [300/600], Loss: 0.5487
Epoch [4/5], Step [400/600], Loss: 0.4700
Epoch [4/5], Step [500/600], Loss: 0.4120
Epoch [4/5], Step [600/600], Loss: 0.4535
Epoch [5/5], Step [100/600], Loss: 0.4345
Epoch [5/5], Step [200/600], Loss: 0.5986
Epoch [5/5], Step [300/600], Loss: 0.4147
Epoch [5/5], Step [400/600], Loss: 0.4893
Epoch [5/5], Step [500/600], Loss: 0.3821
Epoch [5/5], Step [600/600], Loss: 0.5339
Accuracy of the network on the 10000 test images: 88.88 %
  • RMSprop
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms

# Hyper-parameters 
input_size = 784
hidden_size = 500
num_classes = 10
num_epochs = 5
batch_size = 100
learning_rate = 0.001
rmsprop_loss = []

# MNIST dataset 
train_dataset = torchvision.datasets.MNIST(root='data/', 
                                           train=True, 
                                           transform=transforms.ToTensor(),  
                                           download=True)

test_dataset = torchvision.datasets.MNIST(root='data/', 
                                          train=False, 
                                          transform=transforms.ToTensor())

# Data loader
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, 
                                           batch_size=batch_size, 
                                           shuffle=True)

test_loader = torch.utils.data.DataLoader(dataset=test_dataset, 
                                          batch_size=batch_size, 
                                          shuffle=False)

# Fully connected neural network with one hidden layer
class NeuralNet(nn.Module):
    def __init__(self, input_size, hidden_size, num_classes):
        super(NeuralNet, self).__init__()
        self.fc1 = nn.Sequential(nn.Linear(input_size, hidden_size),
                                 nn.ReLU())
        self.fc2 = nn.Linear(hidden_size, num_classes)  
    
    def forward(self, x):
        out = self.fc1(x)
        out = self.fc2(out)
        return out

model = NeuralNet(input_size, hidden_size, num_classes)

# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.RMSprop(model.parameters(), lr=learning_rate)  

# Train the model
total_step = len(train_loader)
for epoch in range(num_epochs):
    for i, (images, labels) in enumerate(train_loader):  
        images = images.reshape(-1, 28 * 28)
        
        # Forward pass
        outputs = model(images)
        loss = criterion(outputs, labels)
        rmsprop_loss.append(loss)
        
        # Backward and optimize
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if (i + 1) % 100 == 0:
            print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}' 
                   .format(epoch+1, num_epochs, i + 1, total_step, loss.item()))

# Test the model
# In test phase, we don't need to compute gradients (for memory efficiency)
with torch.no_grad():
    correct = 0
    total = 0
    for images, labels in test_loader:
        images = images.reshape(-1, 28 * 28)
        outputs = model(images)
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

    print('Accuracy of the network on the 10000 test images: {} %'.format(100 * correct / total))

Epoch [1/5], Step [100/600], Loss: 0.3109
Epoch [1/5], Step [200/600], Loss: 0.3325
Epoch [1/5], Step [300/600], Loss: 0.2390
Epoch [1/5], Step [400/600], Loss: 0.1352
Epoch [1/5], Step [500/600], Loss: 0.2187
Epoch [1/5], Step [600/600], Loss: 0.1740
Epoch [2/5], Step [100/600], Loss: 0.1470
Epoch [2/5], Step [200/600], Loss: 0.1407
Epoch [2/5], Step [300/600], Loss: 0.1190
Epoch [2/5], Step [400/600], Loss: 0.0942
Epoch [2/5], Step [500/600], Loss: 0.0472
Epoch [2/5], Step [600/600], Loss: 0.1146
Epoch [3/5], Step [100/600], Loss: 0.0343
Epoch [3/5], Step [200/600], Loss: 0.0184
Epoch [3/5], Step [300/600], Loss: 0.0288
Epoch [3/5], Step [400/600], Loss: 0.0271
Epoch [3/5], Step [500/600], Loss: 0.0554
Epoch [3/5], Step [600/600], Loss: 0.0332
Epoch [4/5], Step [100/600], Loss: 0.0931
Epoch [4/5], Step [200/600], Loss: 0.1027
Epoch [4/5], Step [300/600], Loss: 0.0247
Epoch [4/5], Step [400/600], Loss: 0.0367
Epoch [4/5], Step [500/600], Loss: 0.0299
Epoch [4/5], Step [600/600], Loss: 0.1616
Epoch [5/5], Step [100/600], Loss: 0.0055
Epoch [5/5], Step [200/600], Loss: 0.0111
Epoch [5/5], Step [300/600], Loss: 0.0779
Epoch [5/5], Step [400/600], Loss: 0.0161
Epoch [5/5], Step [500/600], Loss: 0.0395
Epoch [5/5], Step [600/600], Loss: 0.0185
Accuracy of the network on the 10000 test images: 98.14 %
  • Adam
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms

# Hyper-parameters 
input_size = 784
hidden_size = 500
num_classes = 10
num_epochs = 5
batch_size = 100
learning_rate = 0.001
adam_loss = []

# MNIST dataset 
train_dataset = torchvision.datasets.MNIST(root='data/', 
                                           train=True, 
                                           transform=transforms.ToTensor(),  
                                           download=True)

test_dataset = torchvision.datasets.MNIST(root='data/', 
                                          train=False, 
                                          transform=transforms.ToTensor())

# Data loader
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, 
                                           batch_size=batch_size, 
                                           shuffle=True)

test_loader = torch.utils.data.DataLoader(dataset=test_dataset, 
                                          batch_size=batch_size, 
                                          shuffle=False)

# Fully connected neural network with one hidden layer
class NeuralNet(nn.Module):
    def __init__(self, input_size, hidden_size, num_classes):
        super(NeuralNet, self).__init__()
        self.fc1 = nn.Sequential(nn.Linear(input_size, hidden_size),
                                 nn.ReLU())
        self.fc2 = nn.Linear(hidden_size, num_classes)  
    
    def forward(self, x):
        out = self.fc1(x)
        out = self.fc2(out)
        return out

model = NeuralNet(input_size, hidden_size, num_classes)

# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)  

# Train the model
total_step = len(train_loader)
for epoch in range(num_epochs):
    for i, (images, labels) in enumerate(train_loader):  
        images = images.reshape(-1, 28 * 28)
        
        # Forward pass
        outputs = model(images)
        loss = criterion(outputs, labels)
        adam_loss.append(loss)
        
        # Backward and optimize
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if (i + 1) % 100 == 0:
            print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}' 
                   .format(epoch+1, num_epochs, i + 1, total_step, loss.item()))

# Test the model
# In test phase, we don't need to compute gradients (for memory efficiency)
with torch.no_grad():
    correct = 0
    total = 0
    for images, labels in test_loader:
        images = images.reshape(-1, 28 * 28)
        outputs = model(images)
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

    print('Accuracy of the network on the 10000 test images: {} %'.format(100 * correct / total))

Epoch [1/5], Step [100/600], Loss: 0.2601
Epoch [1/5], Step [200/600], Loss: 0.1910
Epoch [1/5], Step [300/600], Loss: 0.1297
Epoch [1/5], Step [400/600], Loss: 0.1297
Epoch [1/5], Step [500/600], Loss: 0.1411
Epoch [1/5], Step [600/600], Loss: 0.1438
Epoch [2/5], Step [100/600], Loss: 0.1741
Epoch [2/5], Step [200/600], Loss: 0.1034
Epoch [2/5], Step [300/600], Loss: 0.1199
Epoch [2/5], Step [400/600], Loss: 0.0795
Epoch [2/5], Step [500/600], Loss: 0.1422
Epoch [2/5], Step [600/600], Loss: 0.1893
Epoch [3/5], Step [100/600], Loss: 0.1063
Epoch [3/5], Step [200/600], Loss: 0.1087
Epoch [3/5], Step [300/600], Loss: 0.1863
Epoch [3/5], Step [400/600], Loss: 0.0304
Epoch [3/5], Step [500/600], Loss: 0.0621
Epoch [3/5], Step [600/600], Loss: 0.0527
Epoch [4/5], Step [100/600], Loss: 0.0198
Epoch [4/5], Step [200/600], Loss: 0.0367
Epoch [4/5], Step [300/600], Loss: 0.0670
Epoch [4/5], Step [400/600], Loss: 0.0470
Epoch [4/5], Step [500/600], Loss: 0.0559
Epoch [4/5], Step [600/600], Loss: 0.0343
Epoch [5/5], Step [100/600], Loss: 0.0762
Epoch [5/5], Step [200/600], Loss: 0.0617
Epoch [5/5], Step [300/600], Loss: 0.0472
Epoch [5/5], Step [400/600], Loss: 0.0191
Epoch [5/5], Step [500/600], Loss: 0.0089
Epoch [5/5], Step [600/600], Loss: 0.0869
Accuracy of the network on the 10000 test images: 97.59 %

plt.figure(figsize=(10, 8))
plt.plot(range(len(SGD_loss)), SGD_loss, label="sgd")
plt.plot(range(len(momentum_loss)), momentum_loss, label="momentum")
plt.plot(range(len(rmsprop_loss)), rmsprop_loss, label="rmsprop")
plt.plot(range(len(adam_loss)), adam_loss, label="adam")
plt.legend()
plt.show()

在这里插入图片描述

评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
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
实付元
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