Multi-Scale Representation Networks for Time Series Classification
Official PyTorch implementation of MRMS-Net and LMRMS-Net, two convolutional architectures designed for efficient and robust time series classification (TSC).
The models exploit multi-scale temporal representations using parallel convolutional branches while maintaining strong computational efficiency.
This repository provides:
- Reference implementations of MRMS-Net and LMRMS-Net
- A minimal training pipeline
- A one-command demo using a sample dataset
- Architecture diagrams
The full experimental framework used in the paper (large-scale evaluation across 142 datasets) is not included in this repository.
Time series classification often requires capturing patterns occurring at multiple temporal scales. MRMS-Net and LMRMS-Net address this by combining parallel convolutional filters with different receptive fields.
A multi-scale architecture with three convolutional branches capturing short-, medium-, and long-term temporal patterns.
Key characteristics:
- Parallel convolutions: k = 3, 5, 7
- Feature fusion block
- Global average pooling
- Lightweight yet expressive representation learning
A computationally efficient variant designed for fast inference.
Key characteristics:
- Two multi-scale branches
- Optional early-exit inference
- Reduced parameter count
- Designed for low-latency scenarios
Left: MRMS-Net architecture with three parallel convolutional branches. Right: LMRMS-Net lightweight architecture with optional early exit.
Clone the repository:
git clone https://github.com/alagoz/mrmsnet-tsc
cd mrmsnet-tsc
Install dependencies:
pip install -r requirements.txt
Run the demo training script:
python demo/run_demo.py
This script will:
- Load a small example dataset
- Initialize MRMS-Net
- Train the model for a few epochs
- Print training progress
The demo runs in under a minute on CPU.
MRMS-Net-lsnet-tsc
│
├── models
│ ├── msnet.py
│ └── lsnet.py
│
├── demo
│ └── run_demo.py
│
├── data
│ └── ucr_loader.py
│
├── figures
│ └── architecture.png
│
├── requirements.txt
└── README.md
| Model | Description |
|---|---|
| MRMS-Net | Multi-scale CNN with 3 convolution branches |
| LMRMS-Net | Lightweight multi-scale CNN with optional early exit |
from models.msnet import MSNet
model = MSNet(
in_channels=1,
n_classes=5
)
If you use MRMS-Net or LMRMS-Net in your research, please cite:
@article{alagoz2026msnet,
title={Multi-Scale Representation Networks for Time Series Classification},
author={Alagoz, Celal},
journal={},
year={2026}
}
MIT License
For questions regarding the models or implementation, please open a GitHub issue.