搭建神经网络基础综合

搭建神经网络基础综合

以下内容是根据torch官网和莫烦python学习所得

基本步骤

1. 载入数据，训练集，预测集，标注集
2. 搭建网络，即 class Net
3. 实例化网络 net
4. 创建 optimizer
5. 确定损失函数 loss_func
6. 开始训练
   • 计算预测值 predict
   • 计算损失函数值 loss
   • 优化器 zerograd()
   • 损失反馈 loss.backward()
   • 优化器步进 optimizer.step()

import torch
from torch.autograd import Variable
import torch.nn.functional as F
import matplotlib.pyplot as plt
import torch.utils.data as Data
torch.manual_seed(1)

# fake data. y = 2.863* x^2 + 5.652 * x + 3.423
x = torch.unsqueeze(torch.linspace(-5, 5, 5000), dim=1)
y = 2.863 * x.pow(2) + 5.652 * x + 3.423 * (torch.rand(x.size()) - 0.5)

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

# 参见批训练
torch_dataset = Data.TensorDataset(x, y)
loader = Data.DataLoader(
    dataset=torch_dataset,
    batch_size=100,
    shuffle=True,
    num_workers=2,
)


# create Net
class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.hidden = torch.nn.Linear(1, 20)
        self.predict = torch.nn.Linear(20, 1)

    def forward(self, x):
        x = F.relu(self.hidden(x))
        x = self.predict(x)
        return x


if __name__ == '__main__':
    # 实例化网络对象
    net_SGD = Net()
    net_Momentum = Net()
    net_RMSProp = Net()
    net_Adam = Net()
    nets = [net_SGD, net_Momentum, net_RMSProp, net_Adam]

    # 创建 optimizer
    opt_SGD = torch.optim.SGD(net_SGD.parameters(), lr=0.01)
    opt_Momentum = torch.optim.SGD(net_Momentum.parameters(), lr=0.01, momentum=0.8)
    opt_RMSProp = torch.optim.RMSprop(net_RMSProp.parameters(), lr=0.01, alpha=0.9)
    opt_Adam = torch.optim.Adam(net_Adam.parameters(), lr=0.01, betas=(0.9, 0.99))
    optimizer = [opt_SGD, opt_Momentum, opt_RMSProp, opt_Adam]

    # 创建损失函数
    loss_func = torch.nn.MSELoss()
    loss_net = [[], [], [], []]

    for epoch in range(100): 
        # 整套数据训练100次
        for step, (batch_x, batch_y) in enumerate(loader):
            # 每次取100个样本训练
            for net, opt, loss_opt in zip(nets, optimizer, loss_net):
                # 分别用四个网络训练
                predict = net(batch_x)
                loss = loss_func(predict, batch_y)
                opt.zero_grad()
                loss.backward()
                opt.step()
                loss_opt.append(loss.data.numpy()) # 将损失函数的值储存起来

    labels = ['SGD', 'Momentum', 'RMSprop', 'Adam']
    for i, l_his in enumerate(loss_net):
        plt.plot(l_his, label=labels[i])
    plt.legend(loc='best')
    plt.xlabel('Steps')
    plt.ylabel('Loss')
    plt.ylim((0, 1100))
    plt.show()


# 以下是不采用 batch 的方法
    # for t in range(1000):
    #     for net, opt, loss_opt in zip(nets, optimizer, loss_net):
    #         predict = net(x)
    #         loss = loss_func(predict, y)
    #         opt.zero_grad()
    #         loss.backward()
    #         opt.step()
    #         loss_opt.append(loss.data.numpy())

    # labels = ['SGD', 'Momentum', 'RMSprop', 'Adam']
    # for i, l_his in enumerate(loss_net):
    #     plt.plot(l_his, label=labels[i])
    # plt.legend(loc='best')
    # plt.xlabel('Steps')
    # plt.ylabel('Loss')
    # plt.ylim((0, 1100))
    # plt.show()