manjot4 / NIPS-Reproducibility-Challenge

Shape and Time Distortion Loss forTraining Deep Time Series Forecasting Models

NIPS-Reproducibility-Challenge

Paper: Shape and Time Distortion Loss for Training Deep Time Series Forecasting Models

Prerequisites for Getting Started

• Code in the github repository.

• Data sources.

  • ECG Data
  • Wafer Data
  • Traffic Data

• Related papers.

  • Shape and Time Distortion Loss for Training Deep Time Series Forecasting Models
  • Soft-DTW: a Differentiable Loss Function for Time-Series
  • Introducing the Temporal Distortion Index to perform a bidimensional analysis of renewable energy forecast

Dependencies

The required dependencies are:

numPy 
pandas
matplotlib
sklearn
pytorch
numba
tslearn

Content

Project folder

• In the models folder, you can find four machine learning architectures. (We use conv_lstm, fnn, seq2seq these three models to do the reproduction work.)
• In the loss folder, there are custom loss function dilate loss and it's back propagation implementation.
• alpha_test and gamma_test folders contain our experiments on parameter α and γ, and all the files can be executed on Jupyter notebook.
• data folder includes all the data loaders which would be used when testing on different dataset.
• diff_test folder has 12 Jupyter notebooks which consists of 4 different dataset runing on 3 models.
• run_on_cnn_lstm_model.py, run_on_fcnn_model.py and run_on_seq2seq_model.py are three python files that could be run on Python 3. These experiments only run on synthetic data. If you want to try on different dataset, you need to download other datasets from the above links and put data into corresponding dataloader.