PyTorch笔记 优化器的比较

相关视频：
PyTorch 动态神经网络 (莫烦 Python 教学)

一、导入库、设置超参数、创造数据

二、批数据训练用的TensorDataset和DataLoader

三、创建网络

创建4个网络，分别由优化器SGD、Momentum、RMSprop和Adam训练

四、损失函数

五、训练网络

六、画出loss的图像，比较不同的优化器

可以发现，效果从低到高大致为：SGD < Momentum < RMSprop < Adam

七、代码

import torch
import torch.utils.data as Data
import torch.nn.functional as F
from torch.autograd import Variable
import matplotlib.pyplot as plt

# 自己创建的网路
class Net(torch.nn.Module):
    def __init__(self, n_input, n_hidden, n_output):
        super(Net, self).__init__()
        self.hidden_layer = torch.nn.Linear(n_input, n_hidden)
        self.output_layer = torch.nn.Linear(n_hidden, n_output)
    
    def forward(self, x):
        x = F.relu(self.hidden_layer(x))
        x = self.output_layer(x)
        return x

# 超参数
LR = 0.01
BATCH_SIZE = 32
EPOCH = 12

# 创造数据
x = torch.unsqueeze(torch.linspace(-1,1,1000), dim=1)
print(x.shape)

y = torch.pow(x, 2) + 0.1 * torch.normal(torch.zeros(*x.size()))
print(y.shape)

plt.scatter(x.data.numpy(), y.data.numpy())
plt.show()

# TensorDataset
dataset = Data.TensorDataset(x, y)

# DataLoader
loader = Data.DataLoader(
    dataset=dataset,
    batch_size=BATCH_SIZE,
    shuffle=True,
    num_workers=2 # 非jupyter notebook的环境下删除num_workers
)

# 用SGD优化器训练的网络
net_SGD = Net(1,20,1)

# 用Momentum优化器训练的网络
net_Momentum = Net(1,20,1)

# 用RMSprop优化器训练的网络
net_RMSprop = Net(1,20,1)

# 用Adam优化器训练的网络
net_Adam = Net(1,20,1)

nets = [net_SGD, net_Momentum, net_RMSprop, net_Adam]

# SGD优化器
opt_SGD = torch.optim.SGD(
    net_SGD.parameters(),
    lr=LR
)

# Momentum优化器
opt_Momentum = torch.optim.SGD(
    net_Momentum.parameters(),
    lr=LR,
    momentum=0.8
)

# RMSprop优化器
opt_RMSprop = torch.optim.RMSprop(
    net_RMSprop.parameters(),
    lr=LR,
    alpha=0.9
)

# Adam优化器
opt_Adam = torch.optim.Adam(
    net_Adam.parameters(),
    lr=LR,
    betas=(0.9,0.99)
)

opts = [opt_SGD, opt_Momentum, opt_RMSprop, opt_Adam]

# 损失函数
loss_func = torch.nn.MSELoss()
losses_his = [[],[],[],[]]

for epoch in range(EPOCH):
    print('Epoch',epoch)
    for step, (batch_x, batch_y) in enumerate(loader):
        for net, opt, loss_his in zip(nets, opts, losses_his):
            # 计算loss
            output = net(batch_x)
            loss = loss_func(output, batch_y)
            # 训练优化
            opt.zero_grad()
            loss.backward()
            opt.step()
            # 记录loss
            loss_his.append(loss.data.item())
            
# 画出loss的图像
labels = ['SGD', 'Momentum', 'RMSprop', 'Adam']
for i, loss_his in enumerate(losses_his):
    plt.plot(loss_his, label=labels[i])
plt.legend(loc='best')
plt.xlabel('Steps')
plt.ylabel('Loss')
plt.show()

# 画出loss的图像
labels = ['SGD', 'Momentum', 'RMSprop', 'Adam']
for i, loss_his in enumerate(losses_his):
    plt.plot(loss_his, label=labels[i])
plt.legend(loc='best')
plt.xlabel('Steps')
plt.xlim(0, 100)
plt.ylabel('Loss')
plt.show()

本文地址：https://blog.csdn.net/weixin_43826681/article/details/109633501