LiTE (Lightweight Inception Time for Time Series)

Reference: M. Devanne, M. Ismail-Fawaz et al., “LiTE: Lightweight Inception Time for Time Series,” DSAA 2023. Paper: https://maxime-devanne.com/delegation/publis/ismail-fawaz_dsaa2023.pdf

This repo provides a PyTorch implementation of LiTE and simple training/inference scripts on UCR datasets (example: PhalangesOutlinesCorrect in data/).

Quickstart

1. Create/activate venv

   cd /Users/quentinpour/python/LiTE_pytorch
   python3 -m venv .venv
   source .venv/bin/activate
   pip install -e .          # or: pip install -r requirements.txt

2. Train on the provided dataset

   python train.py --dataset PhalangesOutlinesCorrect --epochs 20 --batch-size 32

3. Inspect quick predictions (printed after training) and optional model save (--save-path model.pth).

Key Files

• src/lite.py: LiTE model.
• src/layers/*: Inception, hybrid, and FCN modules.
• train.py: Data loading, training loop, and simple inference.
• utils/utils.py, utils/trainer.py: datasets, normalization, training helpers.

Notes

• train.py expects UCR-format files under data/<dataset>/<dataset>_TRAIN.txt and _TEST.txt.
• Loss is cross-entropy on the model’s softmax output (CrossEntropyFromProbs wrapper). Adjust flags (filters, dilation) via CLI args.***