机器学习的“hello world”：波士顿房价预测
波士顿房价预测的背景不用提了，简单了解一下数据集的结构。
波士顿房价的数据集，共有506组数据，每组数据共14项，前13项是影响房价的各种因素，比如：城镇人均犯罪率、一氧化氮浓度之类的，最后一项是基于前面的因素的房价。
一、逻辑框架
二、程序框架
python">import torch # 导入pytorch# data 解析数据# net # 搭建网络# loss # 计算损失# optimizer # 优化器# train # 训练# test # 测试1、训练的初步的脚本
新建demo_net.py，输入代码：
python">import torch# data 解析数据
import numpy as np
import reff = open('housing.data', 'r').readlines() # 读取数据
data = [] # 存储数据
for item in ff:out = re.sub(r'\s{2,}', ' ', item).strip() # 去除多余空格（将多个空格替换为一个空格）data.append(out.split(' ')) # 按照空格分割数据并添加到data中data = np.array(data, dtype=np.float32) # 转换为numpy数组
# print(data.shape) # 打印数据形状 (506, 14)
Y = data[:, -1] # 获取标签(所有行的最后一列)
X = data[:, :-1] # 获取特征(所有行的除最后一列的所有列)Y_train = Y[0:496, ...] # 训练集标签
X_train = X[0:496, ...] # 训练集特征
Y_test = Y[496:, ...] # 测试集标签
X_test = X[496:, ...] # 测试集特征# print(Y_train.shape) # 打印训练集标签形状 (496,)
# print(X_train.shape) # 打印训练集特征形状 (496, 13)
# print(Y_test.shape) # 打印测试集标签形状 (10,)
# print(X_test.shape) # 打印测试集特征形状 (10, 13)# net # 搭建网络
class Net(torch.nn.Module): # 定义网络def __init__(self, n_feature, n_output): # n_feature:输入特征数，n_output:输出特征数super(Net, self).__init__()self.predict = torch.nn.Linear(n_feature, n_output) # 定义网络结构(线性回归）def forward(self, x): # 前向传播y = self.predict(x)return ynet = Net(13, 1) # 实例化网络# loss # 计算损失
loss_func = torch.nn.MSELoss() # 定义损失函数(均方误差）# optimizer # 优化器
optimizer = torch.optim.Adam(net.parameters(), lr=0.0001) # 定义优化器# train # 训练
x_data = torch.tensor(X_train, dtype=torch.float32) # 将训练集特征转换为tensor
y_data = torch.tensor(Y_train, dtype=torch.float32) # 将训练集标签转换为tensor
for i in range(10000): # 训练10000次pred = net.forward(x_data) # 前向传播，返回的是一个二维tensor# print(pred.shape) # 打印预测结果形状 (496, 1)pred = torch.squeeze(pred) # 由于返回的是二维的tensor，所以需要去掉维度为1的维度loss = loss_func(pred, y_data) # 计算损失optimizer.zero_grad() # 清空上一次的梯度loss.backward() # 反向传播optimizer.step() # 优化器更新参数if (i+1) % 1000 == 0: # 每1000次打印一次损失print('第{}次训练，损失为{}'.format(i+1, loss))print('预测结果为{}'.format(pred[:10]))print('真实结果为{}'.format(y_data[:10]))# test # 测试运行结果：
python">第1000次训练，损失为213.7091064453125
预测结果为tensor([30.3859, 35.9642, 30.5334, 27.4028, 29.8967, 30.7044, 31.1800, 38.9477,41.2886, 35.3080], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第2000次训练，损失为115.01570129394531
预测结果为tensor([25.6781, 28.8944, 24.1358, 20.5663, 22.7628, 23.5741, 26.7624, 33.8008,35.9616, 30.7862], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第3000次训练，损失为92.31402587890625
预测结果为tensor([24.9008, 25.9834, 22.5986, 19.7416, 21.3243, 21.9060, 25.3798, 30.4637,31.5297, 28.2170], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第4000次训练，损失为78.46049499511719
预测结果为tensor([24.7822, 24.7402, 22.6722, 20.6557, 21.6591, 21.9934, 24.7024, 27.8985,27.8058, 26.3184], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第5000次训练，损失为69.00062561035156
预测结果为tensor([24.8879, 24.1986, 23.2193, 21.9481, 22.4841, 22.6186, 24.2389, 25.9268,24.7516, 24.8738], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第6000次训练，损失为62.37424850463867
预测结果为tensor([25.2287, 24.1047, 23.8811, 23.1481, 23.3747, 23.3950, 23.9017, 24.6387,22.4893, 23.9274], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第7000次训练，损失为57.1735954284668
预测结果为tensor([25.7705, 24.3499, 24.5864, 24.1829, 24.2396, 24.2240, 23.6420, 23.9154,20.8723, 23.3721], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第8000次训练，损失为52.881893157958984
预测结果为tensor([26.3800, 24.6829, 25.2719, 25.0930, 25.0307, 25.0057, 23.3551, 23.3264,19.4090, 22.8640], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第9000次训练，损失为49.33217239379883
预测结果为tensor([26.9553, 24.9474, 25.9011, 25.8923, 25.7059, 25.6638, 22.9824, 22.6079,17.8117, 22.2006], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第10000次训练，损失为46.42286682128906
预测结果为tensor([27.4543, 25.1425, 26.4893, 26.5836, 26.2601, 26.1888, 22.5538, 21.7727,16.1070, 21.4219], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])看得出，当前的预测结果还是有很大的偏差的。
2、改进后的训练脚本
.提高学习率，将学习率设为0.01
python">第10000次训练，损失为22.985254287719727
预测结果为tensor([29.2464, 24.2876, 30.5644, 28.9724, 28.4927, 25.0255, 21.5255, 18.3571,10.0867, 17.6953], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])提高学习率之后，收敛更快，所以损失从46.4降到了22.9。
.增加训练次数，训练100000次
python">第100000次训练，损失为21.93065071105957
预测结果为tensor([30.0387, 25.0656, 30.6645, 28.7074, 28.0154, 25.3406, 22.7764, 19.2288,11.0369, 18.6134], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])增加训练次数后，损失略有下降，但不明显。
.加入隐藏层
将网络的定义更换为下面的代码：
python"># net # 搭建网络
class Net(torch.nn.Module): # 定义网络def __init__(self, n_feature, n_output): # n_feature:输入特征数，n_output:输出特征数super(Net, self).__init__()self.hidden = torch.nn.Linear(n_feature, 100) # 定义隐藏层self.predict = torch.nn.Linear(100, n_output) # 定义网络结构(线性回归）def forward(self, x): # 前向传播y = self.hidden(x) # 隐藏层y = torch.relu(y) # 激活函数y = self.predict(y)return y运行结果：
python">第1000次训练，损失为16.16332244873047
预测结果为tensor([31.5807, 24.7012, 31.1625, 33.0276, 29.9564, 28.7967, 21.1699, 17.8390,11.0052, 17.8434], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第2000次训练，损失为10.760713577270508
预测结果为tensor([31.2304, 21.0812, 30.4874, 34.0178, 30.2638, 26.8140, 21.4654, 19.2940,15.2969, 18.9148], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第3000次训练，损失为9.041060447692871
预测结果为tensor([30.0590, 22.6623, 31.4998, 34.3279, 34.0962, 28.8110, 22.4006, 20.0406,17.4557, 19.8644], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第4000次训练，损失为7.654926776885986
预测结果为tensor([28.6500, 22.9828, 31.4478, 34.5795, 35.0444, 28.6649, 22.2122, 20.1728,17.4889, 19.9103], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第5000次训练，损失为7.204400539398193
预测结果为tensor([25.6355, 21.1540, 29.2047, 32.6562, 33.3360, 26.2489, 20.5876, 18.6758,15.7114, 18.5272], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第6000次训练，损失为4.584250450134277
预测结果为tensor([25.8020, 21.5723, 30.2412, 33.8467, 34.3686, 27.0365, 21.3067, 19.1034,16.5482, 19.0247], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第7000次训练，损失为3.7115399837493896
预测结果为tensor([26.3502, 21.7220, 30.8541, 34.7135, 35.1162, 27.7273, 22.0456, 19.7786,17.4401, 19.6452], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第8000次训练，损失为4.195467472076416
预测结果为tensor([25.8955, 21.2075, 30.6258, 34.2735, 34.5896, 27.4262, 21.6917, 19.4653,17.4221, 19.2694], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第9000次训练，损失为3.0317327976226807
预测结果为tensor([26.5321, 21.7933, 32.1591, 34.9432, 34.8340, 28.1285, 22.0055, 20.0637,18.4400, 19.3234], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第10000次训练，损失为2.727705478668213
预测结果为tensor([26.0240, 21.3191, 32.9476, 35.2049, 34.8349, 27.9297, 22.2083, 20.0651,18.3588, 18.8366], grad_fn=<SliceBackward0>)
真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
看得出，10000次训练后，损失已经降到了2.7，比起之前已经大有提高。
3、测试
加入测试代码后：
python">import torch# data 解析数据
import numpy as np
import reff = open('housing.data', 'r').readlines() # 读取数据
data = [] # 存储数据
for item in ff:out = re.sub(r'\s{2,}', ' ', item).strip() # 去除多余空格（将多个空格替换为一个空格）data.append(out.split(' ')) # 按照空格分割数据并添加到data中data = np.array(data, dtype=np.float32) # 转换为numpy数组
# print(data.shape) # 打印数据形状 (506, 14)
Y = data[:, -1] # 获取标签(所有行的最后一列)
X = data[:, :-1] # 获取特征(所有行的除最后一列的所有列)Y_train = Y[0:496, ...] # 训练集标签
X_train = X[0:496, ...] # 训练集特征
Y_test = Y[496:, ...] # 测试集标签
X_test = X[496:, ...] # 测试集特征# print(Y_train.shape) # 打印训练集标签形状 (496,)
# print(X_train.shape) # 打印训练集特征形状 (496, 13)
# print(Y_test.shape) # 打印测试集标签形状 (10,)
# print(X_test.shape) # 打印测试集特征形状 (10, 13)# net # 搭建网络
class Net(torch.nn.Module): # 定义网络def __init__(self, n_feature, n_output): # n_feature:输入特征数，n_output:输出特征数super(Net, self).__init__()self.hidden = torch.nn.Linear(n_feature, 100) # 定义隐藏层self.predict = torch.nn.Linear(100, n_output) # 定义网络结构(线性回归）def forward(self, x): # 前向传播y = self.hidden(x) # 隐藏层y = torch.relu(y) # 激活函数y = self.predict(y)return ynet = Net(13, 1) # 实例化网络# loss # 计算损失
loss_func = torch.nn.MSELoss() # 定义损失函数(均方误差）# optimizer # 优化器
optimizer = torch.optim.Adam(net.parameters(), lr=0.01) # 定义优化器# train # 训练
train_x_data = torch.tensor(X_train, dtype=torch.float32) # 将训练集特征转换为tensor
train_y_data = torch.tensor(Y_train, dtype=torch.float32) # 将训练集标签转换为tensor
test_x_data = torch.tensor(X_test, dtype=torch.float32) # 将测试集特征转换为tensor
test_y_data = torch.tensor(Y_test, dtype=torch.float32) # 将测试集标签转换为tensor
for i in range(10000): # 训练10000次pred_train = net.forward(train_x_data) # 前向传播，返回的是一个二维tensor# print(pred.shape) # 打印预测结果形状 (496, 1)pred_train = torch.squeeze(pred_train) # 由于返回的是二维的tensor，所以需要去掉维度为1的维度loss_train = loss_func(pred_train, train_y_data) # 计算损失optimizer.zero_grad() # 清空上一次的梯度loss_train.backward() # 反向传播optimizer.step() # 优化器更新参数# test # 测试pred_test = net.forward(test_x_data)pred_test = torch.squeeze(pred_test)loss_test = loss_func(pred_test, test_y_data) # 计算损失if (i+1) % 1000 == 0: # 每1000次打印一次损失print('第{}次训练，训练损失为{}'.format(i + 1, loss_train))print('训练预测结果为{}'.format(pred_train[:10]))print('训练真实结果为{}'.format(train_y_data[:10]))print('第{}次训练，测试损失为{}'.format(i + 1, loss_test))print('测试预测结果为{}'.format(pred_test[:10]))print('测试真实结果为{}'.format(test_y_data[:10]))运行结果：
python">C:\Users\DY\.conda\envs\torch\python.exe E:\AI_tset\boston\main.py
第1000次训练，训练损失为13.062897682189941
训练预测结果为tensor([32.4538, 22.4545, 31.6084, 34.6321, 30.7499, 28.7717, 21.2537, 18.4388,14.3816, 17.8076], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第1000次训练，测试损失为15.603384017944336
测试预测结果为tensor([17.2584, 20.4116, 21.6079, 19.7471, 20.3963, 21.4410, 21.3918, 28.3847,25.5809, 21.5427], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第2000次训练，训练损失为9.312514305114746
训练预测结果为tensor([28.3546, 20.2573, 28.6708, 33.5919, 31.0789, 25.5264, 20.3330, 18.5948,15.6572, 17.6856], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第2000次训练，测试损失为14.083860397338867
测试预测结果为tensor([18.4503, 20.9926, 22.3660, 20.1845, 21.4613, 22.1625, 20.6208, 28.3789,25.4795, 20.2372], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第3000次训练，训练损失为6.208388328552246
训练预测结果为tensor([28.0445, 21.0570, 29.5777, 34.5575, 34.3055, 28.3384, 20.9634, 19.6007,16.1345, 18.6369], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第3000次训练，测试损失为6.905886650085449
测试预测结果为tensor([16.9234, 19.6245, 21.3344, 18.5896, 20.3094, 21.4729, 19.0355, 24.3601,22.2127, 18.4200], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第4000次训练，训练损失为5.641458511352539
训练预测结果为tensor([25.9988, 21.0524, 29.1003, 33.8196, 34.5036, 27.8455, 20.7901, 19.6908,15.6711, 18.6173], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第4000次训练，测试损失为8.056257247924805
测试预测结果为tensor([17.5898, 20.1642, 21.8730, 19.0862, 21.0310, 22.3734, 19.6710, 24.7076,22.8086, 18.9423], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第5000次训练，训练损失为5.656107425689697
训练预测结果为tensor([25.2531, 20.6933, 28.9231, 32.7898, 34.1414, 27.3489, 20.4208, 19.7374,15.6711, 18.3137], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第5000次训练，测试损失为10.021121978759766
测试预测结果为tensor([19.1231, 20.8267, 22.5154, 19.6338, 21.5662, 23.3536, 20.3911, 25.4182,23.7853, 19.5190], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第6000次训练，训练损失为4.3244428634643555
训练预测结果为tensor([26.9243, 21.3463, 30.2934, 34.4237, 35.6593, 29.0201, 21.4505, 20.5173,17.6037, 18.9231], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第6000次训练，测试损失为6.1388702392578125
测试预测结果为tensor([19.1049, 20.0117, 21.6194, 18.9644, 20.5173, 22.2967, 19.1247, 23.9584,22.5197, 18.1853], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第7000次训练，训练损失为4.283013343811035
训练预测结果为tensor([27.0127, 22.1808, 30.9352, 35.3205, 36.2137, 29.5154, 21.7429, 20.6820,18.5019, 19.3641], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第7000次训练，测试损失为4.814659118652344
测试预测结果为tensor([18.5595, 19.3107, 21.0298, 18.4600, 19.9520, 21.3664, 17.8218, 23.2565,21.3452, 16.9308], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第8000次训练，训练损失为4.039644718170166
训练预测结果为tensor([26.9686, 22.5262, 31.3496, 35.2846, 36.3410, 29.3871, 21.7442, 20.7460,18.5969, 19.5158], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第8000次训练，测试损失为4.698439598083496
测试预测结果为tensor([18.2167, 19.3340, 21.3022, 18.2839, 20.1922, 21.6993, 17.1337, 23.3157,21.3075, 16.1727], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第9000次训练，训练损失为4.54683256149292
训练预测结果为tensor([23.9196, 20.3218, 29.5554, 32.8275, 34.2579, 27.2489, 20.4540, 18.9822,17.2778, 18.3638], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第9000次训练，测试损失为8.086548805236816
测试预测结果为tensor([19.2664, 20.9391, 22.9442, 19.6506, 21.8333, 23.6144, 18.8776, 24.7309,22.6941, 17.8243], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])
第10000次训练，训练损失为3.7810146808624268
训练预测结果为tensor([23.9642, 20.4657, 29.9442, 32.6707, 34.3400, 27.2285, 20.6211, 19.3677,17.7232, 18.5607], grad_fn=<SliceBackward0>)
训练真实结果为tensor([24.0000, 21.6000, 34.7000, 33.4000, 36.2000, 28.7000, 22.9000, 27.1000,16.5000, 18.9000])
第10000次训练，测试损失为7.744905948638916
测试预测结果为tensor([19.1582, 20.7719, 22.6360, 19.5414, 21.7677, 23.4768, 19.0810, 24.2579,22.2269, 17.9408], grad_fn=<SliceBackward0>)
测试真实结果为tensor([19.7000, 18.3000, 21.2000, 17.5000, 16.8000, 22.4000, 20.6000, 23.9000,22.0000, 11.9000])进程已结束,退出代码0
4、保存和加载
python"># 第一种保存和加载方法
torch.save(net.state_dict(), 'net_dict.pkl') # 保存模型参数
net.load_state_dict(torch.load('net_dict.pkl')) # 加载模型参数# 第二种保存和加载方法
torch.save(net, 'net.pkl') # 保存模型
torch.load('net.pkl') # 加载模型加入保存代码后的完整代码：
python">import torch# data 解析数据
import numpy as np
import reff = open('housing.data', 'r').readlines() # 读取数据
print(ff)
data = [] # 存储数据
for item in ff:out = re.sub(r'\s{2,}', ' ', item).strip() # 去除多余空格（将多个空格替换为一个空格）data.append(out.split(' ')) # 按照空格分割数据并添加到data中data = np.array(data, dtype=np.float32) # 转换为numpy数组
# print(data.shape) # 打印数据形状 (506, 14)
Y = data[:, -1] # 获取标签(所有行的最后一列)
X = data[:, 0:-1] # 获取特征(所有行的除最后一列的所有列)Y_train = Y[0:496, ...] # 训练集标签
X_train = X[0:496, ...] # 训练集特征
Y_test = Y[496:, ...] # 测试集标签
X_test = X[496:, ...] # 测试集特征# print(Y_train.shape) # 打印训练集标签形状 (496,)
# print(X_train.shape) # 打印训练集特征形状 (496, 13)
# print(Y_test.shape) # 打印测试集标签形状 (10,)
# print(X_test.shape) # 打印测试集特征形状 (10, 13)# net # 搭建网络
class Net(torch.nn.Module): # 定义网络def __init__(self, n_feature, n_output): # n_feature:输入特征数，n_output:输出特征数super(Net, self).__init__()self.hidden = torch.nn.Linear(n_feature, 100) # 定义隐藏层self.predict = torch.nn.Linear(100, n_output) # 定义网络结构(线性回归）def forward(self, x): # 前向传播y = self.hidden(x) # 隐藏层y = torch.relu(y) # 激活函数y = self.predict(y)return ynet = Net(13, 1) # 实例化网络# loss # 计算损失
loss_func = torch.nn.MSELoss() # 定义损失函数(均方误差）# optimizer # 优化器
optimizer = torch.optim.Adam(net.parameters(), lr=0.01) # 定义优化器# train # 训练
train_x_data = torch.tensor(X_train, dtype=torch.float32) # 将训练集特征转换为tensor
train_y_data = torch.tensor(Y_train, dtype=torch.float32) # 将训练集标签转换为tensor
test_x_data = torch.tensor(X_test, dtype=torch.float32) # 将测试集特征转换为tensor
test_y_data = torch.tensor(Y_test, dtype=torch.float32) # 将测试集标签转换为tensor
for i in range(10000): # 训练10000次pred_train = net.forward(train_x_data) # 前向传播，返回的是一个二维tensor# print(pred.shape) # 打印预测结果形状 (496, 1)pred_train = torch.squeeze(pred_train) # 由于返回的是二维的tensor，所以需要去掉维度为1的维度loss_train = loss_func(pred_train, train_y_data) # 计算损失optimizer.zero_grad() # 清空上一次的梯度loss_train.backward() # 反向传播optimizer.step() # 优化器更新参数# test # 测试pred_test = net.forward(test_x_data)pred_test = torch.squeeze(pred_test)loss_test = loss_func(pred_test, test_y_data) # 计算损失if (i+1) % 1000 == 0: # 每1000次打印一次损失print('第{}次训练，训练损失为{}'.format(i + 1, loss_train))print('训练预测结果为{}'.format(pred_train[:10]))print('训练真实结果为{}'.format(train_y_data[:10]))print('第{}次测试，测试损失为{}'.format(i + 1, loss_test))print('测试预测结果为{}'.format(pred_test[:10]))print('测试真实结果为{}'.format(test_y_data[:10]))# # 第一种保存和加载方法
# torch.save(net.state_dict(), 'net_dict.pkl') # 保存模型参数
# net.load_state_dict(torch.load('net_dict.pkl')) # 加载模型参数# 第二种保存和加载方法
torch.save(net, 'net.pkl') # 保存模型
# torch.load('net.pkl') # 加载模型
print('已保存')5、推理
新建inference.py，输入：
python">import torch
# from demo_net import Net # 使用这种方法会重复训练
# data 解析数据
import numpy as np
import reclass Net(torch.nn.Module): # 定义网络 # 将之前定义的网络复制过来def __init__(self, n_feature, n_output): # n_feature:输入特征数，n_output:输出特征数super(Net, self).__init__()self.hidden = torch.nn.Linear(n_feature, 100) # 定义隐藏层self.predict = torch.nn.Linear(100, n_output) # 定义网络结构(线性回归）def forward(self, x): # 前向传播y = self.hidden(x) # 隐藏层y = torch.relu(y) # 激活函数y = self.predict(y)return yff = open('housing.data', 'r').readlines() # 读取数据
data = [] # 存储数据
for item in ff:out = re.sub(r'\s{2,}', ' ', item).strip() # 去除多余空格（将多个空格替换为一个空格）data.append(out.split(' ')) # 按照空格分割数据并添加到data中data = np.array(data).astype(np.float32) # 转换为numpy数组
# print(data.shape) # 打印数据形状 (506, 14)
Y = data[:, -1] # 获取标签(所有行的最后一列)
X = data[:, :-1] # 获取特征(所有行的除最后一列的所有列)Y_train = Y[0:496, ...] # 训练集标签
X_train = X[0:496, ...] # 训练集特征
Y_test = Y[496:, ...] # 测试集标签
X_test = X[496:, ...] # 测试集特征net = torch.load('net.pkl')
loss_func = torch.nn.MSELoss() # 定义损失函数(均方误差）test_x_data = torch.tensor(X_test, dtype=torch.float32) # 将测试集特征转换为tensor
test_y_data = torch.tensor(Y_test, dtype=torch.float32) # 将测试集标签转换为tensor
pred_test = net.forward(test_x_data)
pred_test = torch.squeeze(pred_test)
loss_test = loss_func(pred_test, test_y_data) # 计算损失print('测试集预测结果：', pred_test)
print('测试集实际结果:', Y_test)
print('测试集损失：', loss_test) # 打印测试集损失运行后，直接得出了推理结果：
python">测试集预测结果： tensor([17.9976, 20.7273, 22.6870, 19.7387, 22.0204, 23.5350, 19.9008, 23.2567,22.4825, 18.6679], grad_fn=<SqueezeBackward0>)
测试集实际结果: [19.7 18.3 21.2 17.5 16.8 22.4 20.6 23.9 22. 11.9]
测试集损失： tensor(9.1494, grad_fn=<MseLossBackward0>)存疑：在推理的脚本里使用torch.load()调用之前自定义并保存的网络，如使用from demo_net import Net的方法调用，运行推理脚本后，则会又执行一次训练过程。而将之前定义的网络直接复制到推理脚本则不会。
