Regularization-free Diffeomorphic Temporal Alignment Nets

Official PyTorch implementation for our ICML 2023 paper, Regularization-free Diffeomorphic Temporal Alignment Nets.

The Inverse Consistency Averaging Error

Forward and Inverse Alignment

Time Series Joint Alignment

Installation

1. Clone the repository:

   git clone https://github.com/BGU-CS-VIL/RF-DTAN.git

2. Create a new conda environment:

   conda create --name rfdtan python=3.9

3. Activate the conda environment:

   conda activate rfdtan

4. Install the required dependencies:

   pip install -r requirements.txt

Usage

To run the training process, execute the following command:

python train_model.py --dataset ECGFiveDays --ICAE_loss

Replace ECGFiveDays with the desired dataset name and add any additional arguments as needed.

We support the following losses:

• ICAE_loss
• ICAE_triplet_loss
• WCSS_loss
• WCSS_triplet_loss
• smoothness_prior

ICAE - Inverse Consistecny Averaging Error

WCSS - Within-Class Sum of Squares

Requirements

difw==0.0.29
matplotlib==3.5.1
numpy==1.20.3
scikit_learn==1.0.2
torch==1.10.1
tqdm==4.62.3
tsai==0.2.24
tslearn==0.5.2