Practical overview
When it fits
Use Crossformer when you need a documented Python/PyTorch research implementation for multivariate forecasting and can prepare data in the repository’s expected format.
Who it is for
- Researchers running multivariate time-series forecasting experiments
- Practitioners prototyping forecasts that model relationships across variables
- Teams comparing transformer-based forecasting approaches on benchmark datasets
Documented uses
- Train and evaluate forecasts using historical observations from multiple variables
- Adapt a documented CSV-based workflow to a custom multivariate dataset
- Save predictions and ground truth for evaluation
Requirements & limits
The public GitHub repository identifies the project as Apache-2.0 licensed.
- The documented environment pins Python 3.7.10 and specific NumPy, pandas, PyTorch and einops versions.
- Custom data must be formatted as a date or blank first column followed by the multivariate series.
- The architecture uses Dimension-Segment-Wise embedding, two-stage attention with routers, and a hierarchical encoder-decoder.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Train and evaluate with a custom CSV dataset
Use the documented AirQuality example to train Crossformer and save forecasts alongside ground-truth series.
Before you start
- Python 3.7.10
- numpy==1.20.3, pandas==1.3.2, torch==1.8.1 and einops==0.4.1
- A CSV formatted with a date column (or a blank first column) followed by 13 time-series columns
- Place the formatted CSV in the datasets/ folder
Steps
Format AirQualityUCI.csv with the first column as the date or blank and the remaining 13 columns as series; place it in datasets/.
Run the command below.
python main_crossformer.py --data AirQuality --data_path AirQualityUCI.csv --data_dim 13 --in_len 168 --out_len 24 --seg_len 6Run the command below.
python eval_crossformer.py --setting_name Crossformer_AirQuality_il168_ol24_sl6_win2_fa10_dm256_nh4_el3_itr0 --save_pred
Expected outcome
The evaluation produces predictions and ground-truth series in the documented AirQuality results directory.
Guide limitations
- The example specifies hourly data, a 168-hour history, a 24-hour forecast and segment length 6.
- The repository instructions do not establish that other data layouts work without adapting the dataset.
Reproduce the documented ETTh1 experiment
Train and test Crossformer on ETTh1 using the repository’s specified history, forecast and segment lengths.
Before you start
- Python 3.7.10
- numpy==1.20.3, pandas==1.3.2, torch==1.8.1 and einops==0.4.1
- The ETTh1 dataset in the datasets/ folder
Steps
Place the dataset in datasets/.
Run the command below.
python main_crossformer.py --data ETTh1 --in_len 168 --out_len 24 --seg_len 6 --itr 1
Expected outcome
The documented workflow trains and tests the model, saves the trained model under checkpoints/ and writes evaluation metrics under results/.
Guide limitations
- This command documents the ETTh1 configuration with a 168-step input, 24-step output and segment length 6.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- Thinklab-SJTU/Crossformer · Official homepage title · retained in its source language