Deep Time Series is a library to help you quickly build complicated time-series deep learning models such as RNN2Dense, Seq2Seq, Attention-Based, etc. This library is based on Python and the famous deep learning package Keras. All the APIs are made to as close to Keras as possible. Therefore, if you are familar with Keras, you should be able to hands-on in seconds.
Note: Time Series data can be various. Any series data that can be vectorized can be considered as inputs to this library. Therefore, multi-variables time series data, time-dependent images, speeches, text translation, etc., should be all compatible to this library.
Deep Time Series has built-in functions that helps you convert your standard time series dataframe to supervised learning format. For example, we have the following time series dataframe in Pandas:
, where the n_memory_step is the number of previous time steps you want to use for each time step and n_forcast_step is the number of future time steps you want to forcast. split = 0.8 will split the first 80% time steps as train set and the rest 20% time steps as test set.
In Deep Time Series, all models are wrapped into a single objects. To build a model, e.g. sequence-to-sequence model, you can just type:
As you may immediately notice that the commands here are almost identical to Keras. Yes, this is the purpose of this library, i.e. a time-series model-based library that helps you build complicated time-series models in just a few lines.
Once your model is trained, you can save and reload the model for future inference. Also, the syntaxs are almost identical to Keras:
A complete example can be found here
Currently Deep Time Series supports three major frameworks as shown below. Each framework supports simple RNN, LSTM, GRU as their RNN cell. Therefore there are 9 popular models. Other models such as teacher-forcing Seq2Seq and Attention-based will be included soon.
