 Common Deformations. |
 InvConvNet (pool of different types of convolutions with concatenation). |
This repository contains the PyTorch code for the proposed invariant convolutions presented in the paper "Time Series Representations with Hard-Coded Invariances", which was accepted at ICML 2025. Experiments include Classification and Anomaly detection settings for univariate/multivariate time series datasets.
Implementations are provided for the proposed InvConvNet (with standard and invariant convolutions) and vanilla CNN baselines (with standard convolutions), including Inception, ResNet taken from their relevant public repositories.
Models can be found in the ./models folder and layers in ./layers. The scripts with the train loops for classification and anomaly detection can be found in ./exp folder. You can test additional models and layers by following the standard format of the example code and adding the names of the models in ./exp/exp_basic.py and their implementations in ./models.
Similarly, the implementations of the rest baselines used in this work can be found in Time-Series-Library official github repository.
-
In python3.8, you first need to install the required libraries:
pip install -r requirements.txt
You can download the datasets from the following public sources and add them in the ./dataset folder:
- UCR and UEA classification datasets can be found at https://www.timeseriesclassification.com.
- UCIHAR, Sleep-EDF, Epilepsy and Fault-Diagnosis datasets are available at the official github repo of TS-TCC method.
- Anomaly detection datasets are obtained from the Time-Series-Library.
- Classification experiments (InvConvNet):
# UEA datasets and InvConvNet architecture
bash ./scripts/classification/train_uea_invconv.sh
# Additional class. datasets and InvConvNet architecture
bash ./scripts/classification/train_restclass_invconv.sh
- Classification experiments (Vanilla CNN baselines):
# UEA datasets and vanilla Inception architecture
bash ./scripts/classification/train_uea_inception.sh
# UEA datasets and vanilla ResNet architecture
bash ./scripts/classification/train_uea_resnet.sh
# UEA datasets and vanilla CNN architecture
bash ./scripts/classification/train_uea_conv.sh
# Additional class. datasets and all CNN baselines
bash ./scripts/classification/train_restclass_cnnbaselines.sh
- Anomaly Detection experiments (InvConvNet):
# AD datasets and vanilla InvConvNet architecture
bash ./scripts/anomaly_detection/train_anodet_invconv.sh
- Anomaly Detection experiments (Vanilla CNN baselines):
# AD datasets and vanilla Inception architecture
bash ./scripts/anomaly_detection/train_anodet_inception.sh
# AD datasets and vanilla ResNet architecture
bash ./scripts/anomaly_detection/train_anodet_resnet.sh
# AD datasets and vanilla CNN architecture
bash ./scripts/anomaly_detection/train_anodet_conv.sh
- Robustness experiments (for Normal Conv, InvConv (offset) and InvConv (trend)):
# Training for different types of convolutions
# (--inv_ablation parameter set to 1,2,3 respectively)
bash ./scripts/robustness/train_ucr_invconv.sh
# Testing for different types of convolutions
# and augmentation type for test set: None, Random_Off, Random_LT, Random_Off_LT, Random_Walk
bash ./scripts/robustness/test_ucr_invconv.sh
- Transfer learning experiments (InvConvNet):
# Training for different source datasets (--tl_source parameter set to A,B,C, or D)
# test for all remaining subdatasets
bash ./scripts/transfer_learning/train_faultd_invconv.sh
- Chrysoula Kosma: chrysoula.kosma@ens-paris-saclay.fr
- Thibaut Germain: thibaut.germain@ens-paris-saclay.fr
- Laurent Oudre: laurent.oudre@ens-paris-saclay.fr
The project is licensed under https://mit-license.org.
If you use our method, please cite our paper. Here is the BibTeX file:
@inproceedings{tsinvconv2025,
title = {Time Series Representations with Hard-Coded Invariances},
author = {Germain, Thibaut and Kosma, Chrysoula and Oudre, Laurent},
booktitle = {International Conference on Machine Learning},
year = {2025}
}