###仅为自己练习,没有其他用途
1 import torch
2 from torch import nn
3 import torchvision.datasets as dsets
4 import torchvision.transforms as transforms
5 import matplotlib.pyplot as plt
6
7
8 # torch.manual_seed(1) # reproducible
9
10 # Hyper Parameters
11 EPOCH = 1 # train the training data n times, to save time, we just train 1 epoch
12 BATCH_SIZE = 64
13 TIME_STEP = 28 # rnn time step / image height
14 INPUT_SIZE = 28 # rnn input size / image width
15 LR = 0.01 # learning rate
16 DOWNLOAD_MNIST = True # set to True if haven't download the data
17
18
19 # Mnist digital dataset
20 train_data = dsets.MNIST(
21 root='./mnist/',
22 train=True, # this is training data
23 transform=transforms.ToTensor(), # Converts a PIL.Image or numpy.ndarray to
24 # torch.FloatTensor of shape (C x H x W) and normalize in the range [0.0, 1.0]
25 download=DOWNLOAD_MNIST, # download it if you don't have it
26 )
27
28 # # plot one example
29 # print(train_data.train_data.size()) # (60000, 28, 28)
30 # print(train_data.train_labels.size()) # (60000)
31 # plt.imshow(train_data.train_data[0].numpy(), cmap='gray')
32 # plt.title('%i' % train_data.train_labels[0])
33 # plt.show()
34
35 # Data Loader for easy mini-batch return in training
36 train_loader = torch.utils.data.DataLoader(dataset=train_data, batch_size=BATCH_SIZE, shuffle=True)
37
38 # convert test data into Variable, pick 2000 samples to speed up testing
39 test_data = dsets.MNIST(root='./mnist/', train=False, transform=transforms.ToTensor())
40 test_x = test_data.test_data.type(torch.FloatTensor)[:2000]/255. # shape (2000, 28, 28) value in range(0,1)
41 test_y = test_data.test_labels.numpy()[:2000] # covert to numpy array
42
43
44 class RNN(nn.Module):
45 def __init__(self):
46 super(RNN, self).__init__()
47
48 self.rnn = nn.LSTM( # if use nn.RNN(), it hardly learns
49 input_size=INPUT_SIZE,
50 hidden_size=64, # rnn hidden unit
51 num_layers=1, # number of rnn layer
52 batch_first=True, # input & output will has batch size as 1s dimension. e.g. (batch, time_step, input_size)
53 )
54
55 self.out = nn.Linear(64, 10)
56
57 def forward(self, x):
58 # x shape (batch, time_step, input_size)
59 # r_out shape (batch, time_step, output_size)
60 # h_n shape (n_layers, batch, hidden_size)
61 # h_c shape (n_layers, batch, hidden_size)
62 r_out, (h_n, h_c) = self.rnn(x, None) # None represents zero initial hidden state
63
64 # choose r_out at the last time step
65 out = self.out(r_out[:, -1, :])
66 return out
67
68
69 rnn = RNN()
70 print(rnn)
71
72 optimizer = torch.optim.Adam(rnn.parameters(), lr=LR) # optimize all cnn parameters
73 loss_func = nn.CrossEntropyLoss() # the target label is not one-hotted
74
75 # training and testing
76 for epoch in range(EPOCH):
77 for step, (b_x, b_y) in enumerate(train_loader): # gives batch data
78 b_x = b_x.view(-1, 28, 28) # reshape x to (batch, time_step, input_size)
79
80 output = rnn(b_x) # rnn output
81 loss = loss_func(output, b_y) # cross entropy loss
82 optimizer.zero_grad() # clear gradients for this training step
83 loss.backward() # backpropagation, compute gradients
84 optimizer.step() # apply gradients
85
86 if step % 50 == 0:
87 test_output = rnn(test_x) # (samples, time_step, input_size)
88 pred_y = torch.max(test_output, 1)[1].data.numpy()
89 accuracy = float((pred_y == test_y).astype(int).sum()) / float(test_y.size)
90 print('Epoch: ', epoch, '| train loss: %.4f' % loss.data.numpy(), '| test accuracy: %.2f' % accuracy)
91
92 # print 10 predictions from test data
93 test_output = rnn(test_x[:10].view(-1, 28, 28))
94 pred_y = torch.max(test_output, 1)[1].data.numpy()
95 print(pred_y, 'prediction number')
96 print(test_y[:10], 'real number')