###仅为自己练习，没有其他用途
1 # library
2 # standard library
3 import os
4
5 # third-party library
6 import torch
7 import torch.nn as nn
8 import torch.utils.data as Data
9 import torchvision
10 import matplotlib.pyplot as plt
11
12 # torch.manual_seed(1) # reproducible
13
14 # Hyper Parameters
15 EPOCH = 1 # train the training data n times, to save time, we just train 1 epoch
16 BATCH_SIZE = 50
17 LR = 0.001 # learning rate
18 DOWNLOAD_MNIST = False
19
20
21 # Mnist digits dataset
22 if not(os.path.exists('./mnist/')) or not os.listdir('./mnist/'):
23 # not mnist dir or mnist is empyt dir
24 DOWNLOAD_MNIST = True
25
26 train_data = torchvision.datasets.MNIST(
27 root='./mnist/',
28 train=True, # this is training data
29 transform=torchvision.transforms.ToTensor(), # Converts a PIL.Image or numpy.ndarray to
30 # torch.FloatTensor of shape (C x H x W) and normalize in the range [0.0, 1.0]
31 download=DOWNLOAD_MNIST,
32 )
33
34 # # plot one example
35 # print(train_data.train_data.size()) # (60000, 28, 28)
36 # print(train_data.train_labels.size()) # (60000)
37 # plt.imshow(train_data.train_data[0].numpy(), cmap='gray')
38 # plt.title('%i' % train_data.train_labels[0])
39 # plt.show()
40
41 # Data Loader for easy mini-batch return in training, the image batch shape will be (50, 1, 28, 28)
42 train_loader = Data.DataLoader(dataset=train_data, batch_size=BATCH_SIZE, shuffle=True)
43 # pick 2000 samples to speed up testing
44 test_data = torchvision.datasets.MNIST(root='./mnist/', train=False)
45 test_x = torch.unsqueeze(test_data.test_data, dim=1).type(torch.FloatTensor)[:2000]/255. # shape from (2000, 28, 28) to (2000, 1, 28, 28), value in range(0,1)
46 test_y = test_data.test_labels[:2000]
47
48
49 class CNN(nn.Module):
50 def __init__(self):
51 super(CNN, self).__init__()
52 self.conv1 = nn.Sequential( # input shape (1, 28, 28)
53 nn.Conv2d(
54 in_channels=1, # input height
55 out_channels=16, # n_filters
56 kernel_size=5, # filter size
57 stride=1, # filter movement/step
58 padding=2, # if want same width and length of this image after Conv2d, padding=(kernel_size-1)/2 if stride=1
59 ), # output shape (16, 28, 28)
60 nn.ReLU(), # activation
61 nn.MaxPool2d(kernel_size=2), # choose max value in 2x2 area, output shape (16, 14, 14)
62 )
63 self.conv2 = nn.Sequential( # input shape (16, 14, 14)
64 nn.Conv2d(16, 32, 5, 1, 2), # output shape (32, 14, 14)
65 nn.ReLU(), # activation
66 nn.MaxPool2d(2), # output shape (32, 7, 7)
67 )
68 self.out = nn.Linear(32 * 7 * 7, 10) # fully connected layer, output 10 classes
69
70 def forward(self, x):
71 x = self.conv1(x)
72 x = self.conv2(x)
73 x = x.view(x.size(0), -1) # flatten the output of conv2 to (batch_size, 32 * 7 * 7)
74 output = self.out(x)
75 return output, x # return x for visualization
76
77
78 cnn = CNN()
79 print(cnn) # net architecture
80
81 optimizer = torch.optim.Adam(cnn.parameters(), lr=LR) # optimize all cnn parameters
82 loss_func = nn.CrossEntropyLoss() # the target label is not one-hotted
83
84 # following function (plot_with_labels) is for visualization, can be ignored if not interested
85 from matplotlib import cm
86 try: from sklearn.manifold import TSNE; HAS_SK = True
87 except: HAS_SK = False; print('Please install sklearn for layer visualization')
88 def plot_with_labels(lowDWeights, labels):
89 plt.cla()
90 X, Y = lowDWeights[:, 0], lowDWeights[:, 1]
91 for x, y, s in zip(X, Y, labels):
92 c = cm.rainbow(int(255 * s / 9)); plt.text(x, y, s, backgroundcolor=c, fontsize=9)
93 plt.xlim(X.min(), X.max()); plt.ylim(Y.min(), Y.max()); plt.title('Visualize last layer'); plt.show(); plt.pause(0.01)
94
95 plt.ion()
96 # training and testing
97 for epoch in range(EPOCH):
98 for step, (b_x, b_y) in enumerate(train_loader): # gives batch data, normalize x when iterate train_loader
99
100 output = cnn(b_x)[0] # cnn output
101 loss = loss_func(output, b_y) # cross entropy loss
102 optimizer.zero_grad() # clear gradients for this training step
103 loss.backward() # backpropagation, compute gradients
104 optimizer.step() # apply gradients
105
106 if step % 50 == 0:
107 test_output, last_layer = cnn(test_x)
108 pred_y = torch.max(test_output, 1)[1].data.numpy()
109 accuracy = float((pred_y == test_y.data.numpy()).astype(int).sum()) / float(test_y.size(0))
110 print('Epoch: ', epoch, '| train loss: %.4f' % loss.data.numpy(), '| test accuracy: %.2f' % accuracy)
111 if HAS_SK:
112 # Visualization of trained flatten layer (T-SNE)
113 tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=5000)
114 plot_only = 500
115 low_dim_embs = tsne.fit_transform(last_layer.data.numpy()[:plot_only, :])
116 labels = test_y.numpy()[:plot_only]
117 plot_with_labels(low_dim_embs, labels)
118 plt.ioff()
119
120 # print 10 predictions from test data
121 test_output, _ = cnn(test_x[:10])
122 pred_y = torch.max(test_output, 1)[1].data.numpy()
123 print(pred_y, 'prediction number')
124 print(test_y[:10].numpy(), 'real number')
