PyTorch 深度学习实践

 1 import numpy as np
 2 import torch
 3 import matplotlib.pyplot as plt
 4 import os
 5 os.environ['KMP_DUPLICATE_LIB_OK']='True'
 6 
 7 #1 prepare dataset
 8 xy = np.loadtxt('diabetes.csv', delimiter=',', dtype=np.float32)
 9 x_data = torch.from_numpy(xy[:, :-1])  # 第一个‘：’是指读取所有行，第二个‘：’是指从第一列开始，最后一列不要
10 y_data = torch.from_numpy(xy[:, [-1]])  # [-1] 最后得到的是个矩阵
11 
12 
13 #2 design model using class
14 class Model(torch.nn.Module):
15     def __init__(self):
16         super(Model, self).__init__()
17         self.linear1 = torch.nn.Linear(8, 6)  # 输入数据x的特征是8维，x有8个特征
18         self.linear2 = torch.nn.Linear(6, 4)
19         self.linear3 = torch.nn.Linear(4, 1)
20         self.activate = torch.nn.ReLU()
21         self.sigmoid = torch.nn.Sigmoid()  # 将其看作是网络的一层，而不是简单的函数使用
22 
23     def forward(self, x):
24         x = self.activate(self.linear1(x))
25         x = self.activate(self.linear2(x))
26         x = self.sigmoid(self.linear3(x))  # y hat
27         return x
28 model = Model()
29 
30 #3 construct loss and optimizer
31 # criterion = torch.nn.BCELoss(size_average = True)
32 criterion = torch.nn.BCELoss(reduction='mean')
33 optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
34 
35 epoch_list = []
36 loss_list = []
37 #4 training cycle forward, backward, update
38 for epoch in range(1000):
39     y_pred = model(x_data)
40     loss = criterion(y_pred, y_data)
41     print(epoch, loss.item())
42     epoch_list.append(epoch)
43     loss_list.append(loss.item())
44 
45     optimizer.zero_grad()
46     loss.backward()
47 
48     optimizer.step()
49 
50 #test
51 # print('w = ', model.linear3.weight.data)
52 # print('b = ', model.linear3.bias.data)
53 # x_test = torch.Tensor([0.176471,0.256281,0.147541,-0.474747,-0.728132,-0.0730253,-0.891546,-0.333333])#0
54 # x_test = torch.Tensor([-0.0588235,-0.00502513,0.377049,0,0,0.0551417,-0.735269,-0.0333333])#1
55 # y_test = model(x_test)
56 # print('y_pred = ', y_test.data)
57 
58 plt.plot(epoch_list, loss_list)
59 plt.ylabel('loss')
60 plt.xlabel('epoch')
61 plt.show()

996 0.4599120318889618
997 0.4599059224128723
998 0.45989990234375
999 0.45989376306533813

训练1000次loss在0.45的样子，预测也不准，单独使用sigmoid函数收敛还慢点，再增加epoch次数也很难达到理想的效果，我达到的最好结果是loss在0.2几，应该是学习函数太简单了