#

固定位置编码

位置编码将序列中的位置编码成一个大小为d_model 的向量。

P E_{p, 2 i} P E_{p, 2 i + 1} = s in (\frac{p}{1000 0 ^{\frac{2 i}{d _{m o d e l}}}}) = cos (\frac{p}{1000 0 ^{\frac{2 i}{d _{m o d e l}}}})

其中 $1 \leq 2 i, 2 i + 1 \leq d_{m o d e l}$ ，这些索引是编码中的特征索引， $p$ 是位置。

23import math
24
25import numpy as np
26import torch
27import torch.nn as nn

#

30class PositionalEncoding(nn.Module):

#

31    def __init__(self, d_model: int, dropout_prob: float, max_len: int = 5000):
32        super().__init__()
33        self.dropout = nn.Dropout(dropout_prob)
34
35        self.register_buffer('positional_encodings', get_positional_encoding(d_model, max_len), False)

#

37    def forward(self, x: torch.Tensor):
38        pe = self.positional_encodings[:x.shape[0]].detach().requires_grad_(False)
39        x = x + pe
40        x = self.dropout(x)
41        return x

#

44def get_positional_encoding(d_model: int, max_len: int = 5000):

#

空编码向量

46    encodings = torch.zeros(max_len, d_model)

#

位置索引

48    position = torch.arange(0, max_len, dtype=torch.float32).unsqueeze(1)

#

$2 * i$

50    two_i = torch.arange(0, d_model, 2, dtype=torch.float32)

#

$1000 0^{\frac{2 i}{d _{m o d e l}}}$

52    div_term = torch.exp(two_i * -(math.log(10000.0) / d_model))

#

$P E_{p, 2 i} = s in (\frac{p}{1000 0 ^{\frac{2 i}{d _{m o d e l}}}})$

54    encodings[:, 0::2] = torch.sin(position * div_term)

#

$P E_{p, 2 i + 1} = cos (\frac{p}{1000 0 ^{\frac{2 i}{d _{m o d e l}}}})$

56    encodings[:, 1::2] = torch.cos(position * div_term)

#

增加批处理维度

59    encodings = encodings.unsqueeze(1).requires_grad_(False)
60
61    return encodings

#

64def _test_positional_encoding():
65    import matplotlib.pyplot as plt
66
67    plt.figure(figsize=(15, 5))
68    pe = get_positional_encoding(20, 100)
69    plt.plot(np.arange(100), pe[:, 0, 4:8].numpy())
70    plt.legend(["dim %d" % p for p in [4, 5, 6, 7]])
71    plt.title("Positional encoding")
72    plt.show()
73
74
75if __name__ == '__main__':
76    _test_positional_encoding()