pytorch—搭建简单神经网络（回归）

红太狼 2021-05-12 11:49 671阅读 0赞

*  用**pytorch**搭建神经网络，拟合曲线：  y = x 2 y=x^2 y=x2
 *  环境 pytoch1.1.0

### 1. 生成训练数据 ###

# 准备数据
    x_data = torch.unsqueeze(torch.linspace(-1, 1, 100), dim=-1)
    y_data = x_data.pow(2)
    
    # 通过matplotlib可视化生成的数据
    # plt.scatter(x_data.numpy(), y_data.numpy())
    # plt.show()

### 2. 搭建神经网络 ###

class Neuro_net(torch.nn.Module):
        """搭建神经网络"""
        def __init__(self, n_feature, n_hidden_layer, n_output):
            super(Neuro_net, self).__init__()   # 继承__init__功能
            # 隐藏层
            self.hidden_layer = torch.nn.Linear(n_feature, n_hidden_layer) 
            # 输出层
            self.output_layer = torch.nn.Linear(n_hidden_layer, n_output)
    	# 前向传播
        def forward(self, x_data):
            # 隐藏层出来后使用激励函数激活，最后返回预测值
            hidden_layer_x = torch.relu(self.hidden_layer(x_data))
            pridect_y = self.output_layer(hidden_layer_x)
            return pridect_y

### 3. 训练神经网络 ###

*  我这边选择训练500步，拟合效果还行，步数在往上增加你会发现loss还会减小，拟合效果还会更好，但也有可能遇到振荡。

num_feature = 1
    num_hidden_layer = 10
    num_output = 1
    epoch = 500
    # 实例化神经网络
    net = Neuro_net(num_feature, num_hidden_layer, num_output)
    # print(net) # 可查看网络结构
    
    # optimizer 优化 学习率在这里设置
    optimizer = torch.optim.SGD(net.parameters(), lr=0.2)
    # loss funaction
    loss_funaction = torch.nn.MSELoss()
    
    # train
    for step in range(epoch):
        pridect_y = net(x_data)  # 喂入训练数据 得到预测的y值
        loss = loss_funaction(pridect_y, y_data)  # 计算损失
    
        optimizer.zero_grad()    # 为下一次训练清除上一步残余更新参数
        loss.backward()          # 误差反向传播，计算梯度
        optimizer.step()         # 将参数更新值施加到 net 的 parameters 上

### 4. 为了更好理解训练过程，加入可视化 ###

import matplotlib.pyplot as plt
    plt.ion()   # 画图
    
    for step in range(epoch):
        ...
        loss.backward()
        optimizer.step()
        
        if step % 5 == 0:
            print("已训练{}步 | loss：{}。".format(step, loss))
            plt.cla()
            plt.scatter(x_data.numpy(), y_data.numpy())
            plt.plot(x_data.numpy(), pridect_y.data.numpy(), color="green", marker="o", linewidth=6, label="predict_line")
            plt.text(0.5, 0, 'Loss=%.4f' % loss.data.numpy(), fontdict={ 'size': 15, 'color': 'red'})
            plt.pause(0.1)
            
    plt.ioff()
    plt.show()

### 整体代码： ###

import torch
    import matplotlib.pyplot as plt
    
    
    class Neuro_net(torch.nn.Module):
        """搭建神经网络"""
        def __init__(self, n_feature, n_hidden_layer, n_output):
            super(Neuro_net, self).__init__()   # 继承__init__功能
            self.hidden_layer = torch.nn.Linear(n_feature, n_hidden_layer)
            self.output_layer = torch.nn.Linear(n_hidden_layer, n_output)
    
        def forward(self, x_data):
            hidden_layer_x = torch.relu(self.hidden_layer(x_data))
            pridect_y = self.output_layer(hidden_layer_x)
            return pridect_y
    
    # 准备数据
    x_data = torch.unsqueeze(torch.linspace(-1, 1, 100), dim=-1)
    y_data = x_data.pow(2)
    
    # 通过matplotlib可视化生成的数据
    # plt.scatter(x_data.numpy(), y_data.numpy())
    # plt.show()
    
    num_feature = 1
    num_hidden_layer = 10
    num_output = 1
    epoch = 500
    # 实例化神经网络
    net = Neuro_net(num_feature, num_hidden_layer, num_output)
    # print(net)
    
    # optimizer 优化
    optimizer = torch.optim.SGD(net.parameters(), lr=0.2)
    # loss funaction
    loss_funaction = torch.nn.MSELoss()
    
    plt.ion()
    # train
    for step in range(epoch):
        pridect_y = net(x_data)  # 喂入训练数据 得到预测的y值
        loss = loss_funaction(pridect_y, y_data)  # 计算损失
    
        optimizer.zero_grad()    # 为下一次训练清除上一步残余更新参数
        loss.backward()          # 误差反向传播，计算梯度
        optimizer.step()         # 将参数更新值施加到 net 的 parameters 上
    
        if step % 5 == 0:
            print("已训练{}步 | loss：{}。".format(step, loss))
            plt.cla()
            plt.scatter(x_data.numpy(), y_data.numpy())
            plt.plot(x_data.numpy(), pridect_y.data.numpy(), color="green", marker="o", linewidth=6, label="predict_line")
            plt.text(0.5, 0, 'Loss=%.4f' % loss.data.numpy(), fontdict={ 'size': 13, 'color': 'red'})
            plt.pause(0.1)
    
    plt.ioff()
    plt.show()

### 可视化动态拟合图 ###

*  训练开始  
    ![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80NDkxMjE1OQ_size_16_color_FFFFFF_t_70]
 *  训练结束  
    ![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80NDkxMjE1OQ_size_16_color_FFFFFF_t_70 1]

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80NDkxMjE1OQ_size_16_color_FFFFFF_t_70]: /images/20210511/a4a565af114144b2b03e28923988ecf3.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80NDkxMjE1OQ_size_16_color_FFFFFF_t_70 1]: /images/20210511/11d9eeff54314b49bf2cee3a24dccd50.png