The model is described in the associated Medium post: https://medium.com/p/168b47e54d54
The problem is divided into:
-
Generate 2 datasets: one that will serve to train the model (100K time series), the other as unseen data (2.5M time series) on which to serve the model.
-
Train a 1-D CNN model on
train
dataset -
Serve the model on
unseen
dataset (when serving mode) or on test set (when evaluation mode)
-
notebooks/data.json
: configuration file for the data generation (temporarily included in the training and scoring codes) -
notebooks/config.json
: model configuration file (temporarily included in the training and scoring codes) -
notebooks/synthetic_data_generation.py
: the code for synthetic data generation -
notebooks/training.py
: the code for model training -
notebooks/scoring.py
: the code for model serving / evaluation -
notebooks/utils.py
: the python module containing helper functions
old-autoencoder folder (old): This folder contains examples of how to perform time series forecast using LSTM autoencoders and 1-d convolutional neural networks in Keras