Multivariate Time Series Transformer, public version

Multivariate Time Series Transformer Framework

This code corresponds to the paper: George Zerveas et al. A Transformer-based Framework for Multivariate Time Series Representation Learning, in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '21), August 14-18, 2021. ArXiV version:

If you find this code or any of the ideas in the paper useful, please consider citing:

author = {Zerveas, George and Jayaraman, Srideepika and Patel, Dhaval and Bhamidipaty, Anuradha and Eickhoff, Carsten},
title = {A Transformer-Based Framework for Multivariate Time Series Representation Learning},
year = {2021},
isbn = {9781450383325},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {},
doi = {10.1145/3447548.3467401},
booktitle = {Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining},
pages = {2114–2124},
numpages = {11},
keywords = {regression, framework, multivariate time series, classification, transformer, deep learning, self-supervised learning, unsupervised learning, imputation},
location = {Virtual Event, Singapore},
series = {KDD '21}


Instructions refer to Unix-based systems (e.g. Linux, MacOS).

cd mvts_transformer/

Inside an already existing root directory, each experiment will create a time-stamped output directory, which contains model checkpoints, performance metrics per epoch, predictions per sample, the experiment configuration, log files etc. The following commands assume that you have created a new root directory inside the project directory like this: mkdir experiments.

This code has been tested with Python 3.7 and 3.8.

[We recommend creating and activating a conda or other Python virtual environment (e.g. virtualenv) to install packages and avoid conficting package requirements; otherwise, to run pip, the flag --user or sudo privileges will be necessary.]

pip install -r requirements.txt

[Note: Because sometimes newer versions of packages (e.g. sktime) break backward compatibility with previous versions or other packages, if you are encountering issues, you can instead use failsafe_requirements.txt, which contains specific versions of packages tested to work with this codebase.]

Get data from TS Archive

Download dataset files and place them in separate directories, one for regression and one for classification.



Adding your own datasets

To train and evaluate on your own data, you have to add a new data class in datasets/ You can see other examples for data classes in that file, or the template in

The data class sets up one or more pandas DataFrame(s) containing all data, indexed by example IDs. Depending on the task, these dataframes are accessed by the Pytorch Dataset subclasses in

For example, autoregressive tasks (e.g. imputation, transduction) require a member dataframe self.feature_df, while regression and classification (implemented through ClassiregressionDataset) additionally require a self.labels_df member variable to be defined inside the data class in

Once you write your data class, you must add a string identifier for it in the data_factory dictionary inside

data_factory = {'weld': WeldData,
                'tsra': TSRegressionArchive,
                'pmu': PMUData,
                'mydataset': MyNewDataClass}

You can now train and evaluate using your own dataset through the option --data_class mydataset.

Example commands

To see all command options with explanations, run: python src/ --help

You should replace $1 below with the name of the desired dataset. The commands shown here specify configurations intended for BeijingPM25Quality for regression and SpokenArabicDigits for classification.

[To obtain best performance for other datasets, use the hyperparameters as given in the Supplementary Material of the paper. For example, for self-supervised pretraining of BeijingPM25Quality, the correct batch size is 128. Appropriate downsampling with the option --subsample_factor can be often used on datasets with longer time series to speedup training, without significant performance degradation.]

The configurations as shown below will evaluate the model on the TEST set periodically during training, and at the end of training.

Besides the console output and the logfile output.log, you can monitor the evolution of performance (after installing tensorboard: pip install tensorboard) with:

tensorboard dev upload --name my_exp --logdir path/to/output_dir

Train models from scratch


(Note: the loss reported for regression is the Mean Square Error, i.e. without the Root)

python src/ --output_dir path/to/experiments --comment "regression from Scratch" --name $1_fromScratch_Regression --records_file Regression_records.xls --data_dir path/to/Datasets/Regression/$1/ --data_class tsra --pattern TRAIN --val_pattern TEST --epochs 100 --lr 0.001 --optimizer RAdam  --pos_encoding learnable --task regression


python src/ --output_dir experiments --comment "classification from Scratch" --name $1_fromScratch --records_file Classification_records.xls --data_dir path/to/Datasets/Classification/$1/ --data_class tsra --pattern TRAIN --val_pattern TEST --epochs 400 --lr 0.001 --optimizer RAdam  --pos_encoding learnable  --task classification  --key_metric accuracy

Pre-train models (unsupervised learning through input masking)

Can be used for any downstream task, e.g. regression, classification, imputation.

Make sure that the network architecture parameters of the pretrained model match the parameters of the desired fine-tuned model (e.g. use --d_model 64 for SpokenArabicDigits).

python src/ --output_dir experiments --comment "pretraining through imputation" --name $1_pretrained --records_file Imputation_records.xls --data_dir /path/to/$1/ --data_class tsra --pattern TRAIN --val_ratio 0.2 --epochs 700 --lr 0.001 --optimizer RAdam --batch_size 32 --pos_encoding learnable --d_model 128

As noted above, please check the paper for the optimal hyperparameter values for each dataset. E.g. for pretraining on BeijingPM25Quality, one should use --batch_size 128.

Fine-tune pretrained models

Make sure that network architecture parameters (e.g. d_model) used to fine-tune a model match the pretrained model.


python src/ --output_dir experiments --comment "finetune for regression" --name BeijingPM25Quality_finetuned --records_file Regression_records.xls --data_dir /path/to/Datasets/Regression/BeijingPM25Quality/ --data_class tsra --pattern TRAIN --val_pattern TEST  --epochs 200 --lr 0.001 --optimizer RAdam --pos_encoding learnable --d_model 128 --load_model path/to/BeijingPM25Quality_pretrained/checkpoints/model_best.pth --task regression --change_output --batch_size 128


python src/ --output_dir experiments --comment "finetune for classification" --name SpokenArabicDigits_finetuned --records_file Classification_records.xls --data_dir /path/to/Datasets/Classification/SpokenArabicDigits/ --data_class tsra --pattern TRAIN --val_pattern TEST --epochs 100 --lr 0.001 --optimizer RAdam --batch_size 128 --pos_encoding learnable --d_model 64 --load_model path/to/SpokenArabicDigits_pretrained/checkpoints/model_best.pth --task classification --change_output --key_metric accuracy

Get A Weekly Email With Trending Projects For These Categories
No Spam. Unsubscribe easily at any time.