PyDyNet/examples/pytorch/ts_prediction.py at main · WeltXing/PyDyNet · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
import numpy as np
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from tqdm import tqdm

import torch
import pydynet.nn as nn
from pydynet.optim import Adam


def windowize(y, input_len, horizon=1, stride=1, step=1):

    y = np.asarray(y)
    max_i = len(y) - (input_len + horizon) * step + step
    idx_inputs = []
    idx_targets = []
    for i in range(0, max_i, stride):
        inp_idx = i + np.arange(0, input_len * step, step)
        tgt_idx = i + input_len * step + np.arange(0, horizon * step, step)
        idx_inputs.append(inp_idx)
        idx_targets.append(tgt_idx)
    X = y[np.array(idx_inputs)]
    Y = y[np.array(idx_targets)]
    return (
        torch.tensor(X[..., np.newaxis]),
        torch.tensor(Y),
    )


TIME_STEP = 40  # rnn 时序步长数
INPUT_SIZE = 1  # rnn 的输入维度
H_SIZE = 32  # rnn 隐藏单元个数
EPOCHS = 50  # 总共训练次数
h_state = None  # 隐藏层状态


def f(t):
    return np.sin(np.pi * t) + 0.5 * np.cos(2 * np.pi * t)


steps = np.arange(0, 100, .05)
X, Y = windowize(f(steps), input_len=TIME_STEP, horizon=1, stride=1, step=1)

X_train, X_test, Y_train, Y_test = train_test_split(
    X,
    Y,
    test_size=0.2,
    random_state=42,
)


class RNN(nn.Module):

    def __init__(self):
        super(RNN, self).__init__()
        self.rnn = nn.GRU(
            input_size=INPUT_SIZE,
            hidden_size=H_SIZE,
            num_layers=1,
            batch_first=True,
            dtype=np.float32,
        )
        self.out = nn.Linear(H_SIZE, 1)

    def forward(self, x, h_state):
        _, h_state = self.rnn(x, h_state)
        out = self.out(h_state[:, self.rnn.num_layers - 1, :])
        return out


rnn = RNN()
optimizer = Adam(rnn.parameters(), lr=0.01)
criterion = nn.MSELoss()

loss_list = []

plt.rcParams['font.sans-serif'] = ['Times New Roman']
plt.rcParams['mathtext.fontset'] = 'stix'
plt.rcParams['xtick.direction'] = 'in'
plt.rcParams['ytick.direction'] = 'in'
plt.rcParams['axes.linewidth'] = 0.5

bar = tqdm(range(EPOCHS))
visual_steps = np.arange(0, 10, .05)
visual_X, visual_Y = windowize(f(visual_steps),
                               TIME_STEP,
                               horizon=1,
                               stride=1,
                               step=1)

for step in bar:

    rnn.train()
    prediction = rnn(X_train, h_state)
    train_loss = criterion(prediction, Y_train)

    optimizer.zero_grad()
    train_loss.backward()
    optimizer.step()

    plt.figure(figsize=(5, 3))
    plt.grid()

    rnn.eval()
    with torch.no_grad():
        test_loss = criterion(rnn(X_test, h_state), Y_test)

        plt.plot(visual_steps[TIME_STEP:],
                 visual_Y.numpy(),
                 'r-',
                 lw=0.7,
                 label=r'$f(x)=\sin(\pi x)+\cos(2\pi x)/2$')
        plt.plot(
            visual_steps[TIME_STEP:],
            rnn(visual_X, h_state).numpy(),
            'b-.',
            lw=0.7,
            label='Prediction',
        )

    plt.xticks([4, 6, 8, 10])
    plt.yticks([-1.6, -.8, 0, .8])

    plt.legend(loc=1)
    plt.ylim(-1.6, 0.8)
    plt.xlim(visual_steps[TIME_STEP], 10)
    plt.title('Prediction with GRU')
    plt.tight_layout()
    plt.savefig("imgs/rnn.png")
    plt.close()

    bar.set_postfix(
        train_loss="{:.5f}".format(train_loss.item()),
        test_loss="{:.5f}".format(test_loss.item()),
    )