Published Code for Few-Shot Time Series Embedding

This directory contains the published code for our paper on few-shot learning approaches for time series classification.

Structure

The published code includes the following experiments:

1. Augmentations (tsembedding/experiments/augmentations/)

Data augmentation experiments for time series classification.

2. Contrastive TimesNet (tsembedding/experiments/contrastive_timesnet/)

Contrastive learning experiments using the TimesNet architecture. This is a dependency for the supervised fine-tuning experiments.

3. Cross-Dataset Transfer (tsembedding/experiments/cross_dataset_transfer/)

Experiments evaluating transfer learning across different datasets.

4. Inter-subject Representation Generality (tsembedding/experiments/representation_generality/)

Experiments assessing the generality of learned representations across different architectures and datasets. We evaluate the ability of learned features to transfer to the same type of movements, but from different subjects.

5. Supervised Fine-tuning (tsembedding/experiments/supervised_finetune/)

Supervised fine-tuning experiments using TimesNet models.

Installation

Create / activate your python environment of choice, then run the following commands:

1. Install the main package:

pip install -e .

2. Install dependencies:

pip install -r requirements.txt

3. Install local libraries:

cd libs && ./install_all.sh

Core Dependencies

• tsembedding: Main package containing datasets, models, and utilities
• libs/timeserieslibrary: Fork of Time-Series-Library with SOTA models including TimesNet
• libs/micos: Additional time series models

Running Experiments

Each experiment directory contains its own configuration files and training scripts. See the individual README files in each experiment directory for specific instructions.