Our take
- Activity
- 19
- Community
- 4
- Issues
- 21
- Pull requests
- 20
+0 stars in the last 30 days
- Fewer contributors than 91% of the projects we track
Project history
Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.
Stars
2Forks
1Change since Nov 16, 2026: 0 stars·0 forks
Overview
This Python project implements a PyTorch transformer for long-term time-series forecasting. It processes the input at multiple temporal scales using stride-based subsampling, then aligns and fuses the resulting representations to produce forecasts.
It is aimed at developers and researchers exploring multi-resolution forecasting. The implementation is a practical adaptation of the cited MTST paper, not a reproduction of its exact design: it uses sinusoidal positional encoding and interpolation-based fusion, among other differences.
Key Features
- Separate transformer branches process high-, mid-, and low-resolution views of the input.
- Stride-based subsampling provides the different temporal resolutions.
- Interpolation aligns branch outputs before concatenation and linear fusion.
- Configurable input size, embedding dimension, attention heads, dropout, and transformer depth.
- Includes a training script for preprocessing, training, loss tracking, and test inference.
- Uses PyTorch and can select CUDA when available.
Use Cases
- Researchers can prototype multi-scale transformer approaches on forecasting datasets.
- Developers can explore whether combining fine-grained observations and slower trends helps a forecasting task.
- Learners can study a compact end-to-end PyTorch forecasting pipeline, from window creation to predictions.
- Users with OHLCV or similar tabular time-series data can adapt the included training workflow to their own dataset format.
Getting Started
Clone the repository and install its dependencies:
git clone https://github.com/VenkatachalamSubramanianPeriyaSubbu/multiresolution-time-series-transformer
cd mtst
pip install -r requirements.txt
See the README for model usage and training details.
Considerations
- The repository describes its model as an adaptation of the paper. Its stride-based sampling, sinusoidal encoding, and interpolation fusion differ from the paper's described approach.
- The README's performance expectations refer to results reported by the paper, not benchmarks demonstrated for this implementation.
- The training example uses cryptocurrency price data; other datasets may need different preprocessing.
- The README describes the project as MIT-licensed, but the repository metadata supplied here has no license value. Verify the repository's license before reuse.
Found this useful?
Share it with someone who would like multiresolution-time-series-transformer.