gobert / deepar Goto Github PK

Tensorflow 2 implementation of Amazon DeepAR Time Series Forecasting algorithm (https://arxiv.org/abs/1704.04110).

Influenced by these two open-source implementations: https://github.com/arrigonialberto86/deepar and https://github.com/zhykoties/TimeSeries (pytorch).

deepar/dataset:

1. time_series.py: contains TimeSeries and TimeSeriesTest objects that perform covariate augmentation, grouping, scaling, and standardization according to Salinas et al. The objects are also easy to integrate with the D3M AutoML DARPA primitive and piepline infrastructure (https://docs.datadrivendiscovery.org/).

deepar/model:

1. learner.py: contains a DeepARLearner class, which creates the model structure and implements a custom training loop. The model learns a categorical embedding for each unique time series group. It also performs ancestral sampling during inference (for arbitrary horizons into the future) and generates n samples at each timestep. Ancestral sampling can be conditioned with the whole training time series or just the final window.

2. layers.py: contains custom LSTMResetStateful layer and GaussianLayer layer (the latter is from https://github.com/arrigonialberto86/deepar and unused in current codebase)

3. loss.py: contains custom GaussianLogLikelihood loss for real data and NegativeBinomialLogLikelihood loss for positive count data. Both losses support masking and inverse scaling per Salinas et al.