tonibofarull / ts-autoencoder

Dilated Convolutional Autoencoder for univariate Time Series

Autoencoder

Implementation of a Dilated Convolutional Autoencoder for univariate Time Series.

Requirements

Developed for Python 3.8.

Install dependencies using pip,

pip install -r requirements.txt

Getting Started

Structure of the project,

.
├── data     
│   └── ElectricDevices
├── src
│   ├── configs
│   │   ├── arma.yaml
│   │   ├── arma2357.yaml
│   │   └── config.yaml
│   ├── dataloader.py
│   ├── experiments
│   │   ├── exp1-shapley_value.ipynb
│   │   ├── exp2-acc_cor.py
│   │   └── exp3-plane_representation.py
│   ├── interpretability.py
│   ├── main.ipynb
│   ├── models
│   │   ├── CAE.py
│   │   └── losses.py
│   ├── train.py
│   ├── tuning.py
│   └── utils.py
├── utils
└── weights

Execute the jupyter notebook main.ipynb to load the data, train the model and obtain the evaluation and interpretation.

Pretrained Models

In weights there are the pretrained models with the optimal hyperparameter settings.

The models are:

• ARMA with dilation 2, 3, 5, or 7 in separate datasets. Each dilation is trained separately in armaX.pth with X the corresponding dilation.
• ARMA with dilation 2, 3, 5, or 7 in the same dataset. The pretrained model is arma2357.pth.
• Electric Devices dataset. The pretrained model is mod.pth.

The configuration files are placed in src/configs. Use arma.yaml for the first case, arma2357.yaml for the second and config.yaml for Electric Devices. In the first case, there was no hyperparameter tuning and we use standard values for the hyperparameters.

By default main.ipynb uses the Electric Devices dataset.