###仅为自己练习，没有其他用途
1 import torch
2 import torch.nn as nn
3 import torch.utils.data as Data
4 import torchvision
5 import matplotlib.pyplot as plt
6 from mpl_toolkits.mplot3d import Axes3D
7 from matplotlib import cm
8 import numpy as np
9
10
11 # torch.manual_seed(1) # reproducible
12
13 # Hyper Parameters
14 EPOCH = 10
15 BATCH_SIZE = 64
16 LR = 0.005 # learning rate
17 DOWNLOAD_MNIST = False
18 N_TEST_IMG = 5
19
20 # Mnist digits dataset
21 train_data = torchvision.datasets.MNIST(
22 root='./mnist/',
23 train=True, # this is training data
24 transform=torchvision.transforms.ToTensor(), # Converts a PIL.Image or numpy.ndarray to
25 # torch.FloatTensor of shape (C x H x W) and normalize in the range [0.0, 1.0]
26 download=DOWNLOAD_MNIST, # download it if you don't have it
27 )
28
29 # plot one example
30 print(train_data.train_data.size()) # (60000, 28, 28)
31 print(train_data.train_labels.size()) # (60000)
32 plt.imshow(train_data.train_data[2].numpy(), cmap='gray')
33 plt.title('%i' % train_data.train_labels[2])
34 plt.show()
35
36 # Data Loader for easy mini-batch return in training, the image batch shape will be (50, 1, 28, 28)
37 train_loader = Data.DataLoader(dataset=train_data, batch_size=BATCH_SIZE, shuffle=True)
38
39
40 class AutoEncoder(nn.Module):
41 def __init__(self):
42 super(AutoEncoder, self).__init__()
43
44 self.encoder = nn.Sequential(
45 nn.Linear(28*28, 128),
46 nn.Tanh(),
47 nn.Linear(128, 64),
48 nn.Tanh(),
49 nn.Linear(64, 12),
50 nn.Tanh(),
51 nn.Linear(12, 3), # compress to 3 features which can be visualized in plt
52 )
53 self.decoder = nn.Sequential(
54 nn.Linear(3, 12),
55 nn.Tanh(),
56 nn.Linear(12, 64),
57 nn.Tanh(),
58 nn.Linear(64, 128),
59 nn.Tanh(),
60 nn.Linear(128, 28*28),
61 nn.Sigmoid(), # compress to a range (0, 1)
62 )
63
64 def forward(self, x):
65 encoded = self.encoder(x)
66 decoded = self.decoder(encoded)
67 return encoded, decoded
68
69
70 autoencoder = AutoEncoder()
71
72 optimizer = torch.optim.Adam(autoencoder.parameters(), lr=LR)
73 loss_func = nn.MSELoss()
74
75 # initialize figure
76 f, a = plt.subplots(2, N_TEST_IMG, figsize=(5, 2))
77 plt.ion() # continuously plot
78
79 # original data (first row) for viewing
80 view_data = train_data.train_data[:N_TEST_IMG].view(-1, 28*28).type(torch.FloatTensor)/255.
81 for i in range(N_TEST_IMG):
82 a[0][i].imshow(np.reshape(view_data.data.numpy()[i], (28, 28)), cmap='gray'); a[0][i].set_xticks(()); a[0][i].set_yticks(())
83
84 for epoch in range(EPOCH):
85 for step, (x, b_label) in enumerate(train_loader):
86 b_x = x.view(-1, 28*28) # batch x, shape (batch, 28*28)
87 b_y = x.view(-1, 28*28) # batch y, shape (batch, 28*28)
88
89 encoded, decoded = autoencoder(b_x)
90
91 loss = loss_func(decoded, b_y) # mean square error
92 optimizer.zero_grad() # clear gradients for this training step
93 loss.backward() # backpropagation, compute gradients
94 optimizer.step() # apply gradients
95
96 if step % 100 == 0:
97 print('Epoch: ', epoch, '| train loss: %.4f' % loss.data.numpy())
98
99 # plotting decoded image (second row)
100 _, decoded_data = autoencoder(view_data)
101 for i in range(N_TEST_IMG):
102 a[1][i].clear()
103 a[1][i].imshow(np.reshape(decoded_data.data.numpy()[i], (28, 28)), cmap='gray')
104 a[1][i].set_xticks(()); a[1][i].set_yticks(())
105 plt.draw(); plt.pause(0.05)
106
107 plt.ioff()
108 plt.show()
109
110 # visualize in 3D plot
111 view_data = train_data.train_data[:200].view(-1, 28*28).type(torch.FloatTensor)/255.
112 encoded_data, _ = autoencoder(view_data)
113 fig = plt.figure(2); ax = Axes3D(fig)
114 X, Y, Z = encoded_data.data[:, 0].numpy(), encoded_data.data[:, 1].numpy(), encoded_data.data[:, 2].numpy()
115 values = train_data.train_labels[:200].numpy()
116 for x, y, z, s in zip(X, Y, Z, values):
117 c = cm.rainbow(int(255*s/9)); ax.text(x, y, z, s, backgroundcolor=c)
118 ax.set_xlim(X.min(), X.max()); ax.set_ylim(Y.min(), Y.max()); ax.set_zlim(Z.min(), Z.max())
119 plt.show()
