Time2Vec TensorFlow

An intuitive, custom TensorFlow.Keras Layer for Time2Vec Transformation.

Keras implementation of Non-local blocks from [1].

Concept

Time2Vec offers a versatile representation of time with three fundamental properties. It encapsulates scalar notion of time $\tau$, in $\mathbf{t2v}(\tau)$, a vector of size k + 1. This transformation, for an $i^{th}$ element of $\mathbf{t2v}$, is defined as follows:

$$\mathbf{t2v}(\tau)[i] = \begin{cases} \omega_i \tau + \phi_i, & \mathrm{if} & i = 0.\\\ \mathcal{F}(\omega_i \tau + \phi_i), & \mathrm{if} & 1 \leq i \leq k. \end{cases}$$

The above incorporates a periodic activation function denoted as $\mathcal{F}$, and involves learnable parameters $\omega_i$ and $\phi_i$ [1].

Usage Templates

The script time2vec.py contains the Time2Vec instance which takes a single or a group of time series and concatenates the above $\mathbf{t2v}$ tensor to it.

from time2vec import Time2Vec

k = 20
time_series = np.linspace(0, 100, 1000).reshape(-1,1)
time_series = Time2Vec(num_frequency=k)(time_series)
print(time_series)
...

For a more interactive experience check out this Google Colab

1. Seyed Mehran Kazemi, Rishab Goel, Sepehr Eghbali, Janahan Ramanan, Jaspreet Sahota, Sanjay Thakur, Stella Wu, Cathal Smyth, Pascal Poupart, Marcus Brubaker. "Time2Vec: Learning a Vector Representation of Time." arXiv:1907.05321 [cs.LG], 11 Jul 2019. Link