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 的设计主要基于以下几个方面:
Time2Vec 的公式并不复杂:
t 2 v ( τ ) [ i ] = { ω i τ + φ i , if i = 0. F ( ω i τ + φ i ) , if 1 ≤ i ≤ k . \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} t2v(τ)[i]={ωiτ+φi,F(ωiτ+φi),if i=0.if 1≤i≤k.
其中 k k k为 time2vec 的维度, F \mathcal{F} F为周期激活函数, ω i , φ i \omega_i,\varphi_i ωi,φi为可学习参数。为了使算法可以捕获周期性,所以 F \mathcal{F} F选用 sin \sin sin函数( cos \cos cos函数同样效果)捕获周期性。
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
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。