关于RNN模型参数的解释，可以参看RNN参数解释
###仅为自己练习，没有其他用途
1 import torch
2 from torch import nn
3 import numpy as np
4 import matplotlib.pyplot as plt
5
6 # torch.manual_seed(1) # reproducible
7
8 # Hyper Parameters
9 TIME_STEP = 10 # rnn time step
10 INPUT_SIZE = 1 # rnn input size
11 LR = 0.02 # learning rate
12
13 # show data
14 steps = np.linspace(0, np.pi*2, 100, dtype=np.float32) # float32 for converting torch FloatTensor
15 x_np = np.sin(steps)
16 y_np = np.cos(steps)
17 plt.plot(steps, y_np, 'r-', label='target (cos)')
18 plt.plot(steps, x_np, 'b-', label='input (sin)')
19 plt.legend(loc='best')
20 plt.show()
21
22
23 class RNN(nn.Module):
24 def __init__(self):
25 super(RNN, self).__init__()
26
27 self.rnn = nn.RNN(
28 input_size=INPUT_SIZE,
29 hidden_size=32, # rnn hidden unit
30 num_layers=1, # number of rnn layer
31 batch_first=True, # input & output will has batch size as 1s dimension. e.g. (batch, time_step, input_size)
32 )
33 self.out = nn.Linear(32, 1)
34
35 def forward(self, x, h_state):
36 # x (batch, time_step, input_size)
37 # h_state (n_layers, batch, hidden_size)
38 # r_out (batch, time_step, hidden_size)
39 r_out, h_state = self.rnn(x, h_state)
40
41 outs = [] # save all predictions
42 for time_step in range(r_out.size(1)): # calculate output for each time step
43 outs.append(self.out(r_out[:, time_step, :]))
44 return torch.stack(outs, dim=1), h_state
45
46 # instead, for simplicity, you can replace above codes by follows
47 # r_out = r_out.view(-1, 32)
48 # outs = self.out(r_out)
49 # outs = outs.view(-1, TIME_STEP, 1)
50 # return outs, h_state
51
52 # or even simpler, since nn.Linear can accept inputs of any dimension
53 # and returns outputs with same dimension except for the last
54 # outs = self.out(r_out)
55 # return outs
56
57 rnn = RNN()
58 print(rnn)
59
60 optimizer = torch.optim.Adam(rnn.parameters(), lr=LR) # optimize all cnn parameters
61 loss_func = nn.MSELoss()
62
63 h_state = None # for initial hidden state
64
65 plt.figure(1, figsize=(12, 5))
66 plt.ion() # continuously plot
67
68 for step in range(100):
69 start, end = step * np.pi, (step+1)*np.pi # time range
70 # use sin predicts cos
71 steps = np.linspace(start, end, TIME_STEP, dtype=np.float32, endpoint=False) # float32 for converting torch FloatTensor
72 x_np = np.sin(steps)
73 y_np = np.cos(steps)
74
75 x = torch.from_numpy(x_np[np.newaxis, :, np.newaxis]) # shape (batch, time_step, input_size)
76 y = torch.from_numpy(y_np[np.newaxis, :, np.newaxis])
77
78 prediction, h_state = rnn(x, h_state) # rnn output
79 # !! next step is important !!
80 h_state = h_state.data # repack the hidden state, break the connection from last iteration
81
82 loss = loss_func(prediction, y) # calculate loss
83 optimizer.zero_grad() # clear gradients for this training step
84 loss.backward() # backpropagation, compute gradients
85 optimizer.step() # apply gradients
86
87 # plotting
88 plt.plot(steps, y_np.flatten(), 'r-')
89 plt.plot(steps, prediction.data.numpy().flatten(), 'b-')
90 plt.draw(); plt.pause(0.05)
91
92 plt.ioff()
93 plt.show()