BalazsHoranyi / TimeSeries-D3M-Wrappers

Repository of D3M wrappers for time series classification and forecasting

TimeSeriesD3MWrappers/primitives:

D3M primitives

1. classification_knn.py: wrapper for tslearn's KNeighborsTimeSeriesClassifier algorithm

2. classification_lstm.py: wrapper for LSTM Fully Convolutional Networks for Time Series Classification paper, original repo (https://github.com/titu1994/MLSTM-FCN), paper (https://arxiv.org/abs/1801.04503)

3. forecasting_deepar.py: wrapper for DeepAR recurrent, autoregressive Time Series Forecasting algorithm (https://arxiv.org/abs/1704.04110). Custom implementation repo (git+https://github.com/NewKnowledge/deepar#egg=deepar-0.0.1)

4. forecasting_var.py: wrapper for statsmodels' implementation of vector autoregression for multivariate time series

TimeSeriesD3MWrappers/pipelines:

Example pipelines for primitives. Latest are:

1. forecasting_pipeline_imputer.py: pipeline for DeepAR primitive on all forecasting datasets

2. forecasting_pipeline_var.py: pipeline for VAR primitive on all forecasting datasets (except terra datasets)

3. Kanine_pipeline.py: pipeline for Kanine primitive on all classification datasets

4. LSTM_FCN_pipeline.py: pipeline for LSTM_FCN primitive on all classification datasets

Model utils

1. layer_utils.py: implementation of AttentionLSTM in tensorflow (compatible with 2), originally from https://github.com/houshd/LSTM-FCN

2. lstm_model_utils.py: functions to generate LSTM_FCN model architecture and data generators

3. var_model_utils.py: wrapper of the auto_arima method from pmdarima.arima with some specific parameters fixed