Time2Vec 的理解与简单实现

前言

Time2Vec 从其名字就可以看出其功能,将时间进行 Embedding,并且能够应用于不同的模型。

2019 年的一篇论文:Time2Vec: Learning a Vector Representation of Time

Time2Vec

Time2Vec 的设计主要基于以下几个方面:

  1. 捕获周期性和非周期性模式
  2. 对时间缩放不变
  3. 易于与其他模型融合

Time2Vec 的公式并不复杂:

t2v(τ)[i]={ωiτ+φi,if i=0.F(ωiτ+φi),if 1ik.\mathbf{t2v}(\tau)[i]=\begin{cases}\omega_i\tau+\varphi_i, &\text{if }i=0. \\ \mathcal{F}(\omega_i\tau+\varphi_i), &\text{if }1\leq i\leq k. \end{cases}

其中kk为 time2vec 的维度,F\mathcal{F}为周期激活函数,ωi,φi\omega_i,\varphi_i为可学习参数。为了使算法可以捕获周期性,所以F\mathcal{F}选用sin\sin函数(cos\cos函数同样效果)捕获周期性。

PyTorch 实现

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
def t2v(tau, f, out_features, w, b, w0, b0, arg=None):
if arg:
v1 = f(torch.matmul(tau, w) + b, arg)
else:
v1 = f(torch.matmul(tau, w) + b)
v2 = torch.matmul(tau, w0) + b0
return torch.cat([v1, v2], 1)


class SineActivation(nn.Module):
def __init__(self, in_features, out_features):
super(SineActivation, self).__init__()
self.out_features = out_features
self.w0 = nn.parameter.Parameter(torch.randn(in_features, 1))
self.b0 = nn.parameter.Parameter(torch.randn(in_features, 1))
self.w = nn.parameter.Parameter(torch.randn(in_features, out_features - 1))
self.b = nn.parameter.Parameter(torch.randn(in_features, out_features - 1))
self.f = torch.sin

def forward(self, tau):
return t2v(tau, self.f, self.out_features, self.w, self.b, self.w0, self.b0)


class CosineActivation(nn.Module):
def __init__(self, in_features, out_features):
super(CosineActivation, self).__init__()
self.out_features = out_features
self.w0 = nn.parameter.Parameter(torch.randn(in_features, 1))
self.b0 = nn.parameter.Parameter(torch.randn(in_features, 1))
self.w = nn.parameter.Parameter(torch.randn(in_features, out_features - 1))
self.b = nn.parameter.Parameter(torch.randn(in_features, out_features - 1))
self.f = torch.cos

def forward(self, tau):
return t2v(tau, self.f, self.out_features, self.w, self.b, self.w0, self.b0)


class Time2Vec(nn.Module):
def __init__(self, activation, hiddem_dim):
super(Time2Vec, self).__init__()
if activation == "sin":
self.l1 = SineActivation(1, hiddem_dim)
elif activation == "cos":
self.l1 = CosineActivation(1, hiddem_dim)

self.fc1 = nn.Linear(hiddem_dim, 2)

def forward(self, x):
x = self.l1(x)
x = self.fc1(x)
return x

总结

对于时间的 Embedding 怎么说呢,个人感觉其实有必要又没必要,可有可无,当然不是说时间信息不重要。论文没有仔细看,当然主要是内容也比较少,感觉对于时间、位置这些东西的处理,到底还是 sin、cos 效果会好一点?看代码的时候又看见了作者的Date2Vec,模型没怎么看懂,具体也没解释原理,有兴趣的可以看看。

参考资料