TSI-Bench

The code scripts, configurations, and logs here are for TSI-Bench, the first comprehensive benchmark for time series imputation.

❖ Python Environment Creation

A proper Python environment is necessary to reproduce the results. Please ensure that all the below library requirements are satisfied.

tsdb ==0.4
pygrinder ==0.6
benchpots ==0.1.1
pypots ==0.6

For Linux OS, it is able to create the environment with Conda by running conda create -f conda_env.yml. For other OS, library version requirements can also be checked out in conda_env.yml.

❖ Datasets Generation

Please refer to data/README.md.

❖ Results Reproduction

Neural network training

For example, to reproduce the results of SAITS on the dataset Pedestrian, please execute the following command.

nohup python train_model.py \
  --model SAITS \
  --dataset Pedestrian \
  --dataset_fold_path data/melbourne_pedestrian_rate01_step24_point \
  --saving_path results_point_rate01 \
  --device cuda:2 \
  > results_point_rate01/SAITS_pedestrian.log &

After the execution finished, please check out the logging information in the according .log file.

Additionally, as claimed in the paper, hyperparameters of all models get optimized by the tuning functionality in PyPOTS. Hence, tuning configurations are available in the directory PyPOTS_tuning_configs. If you'd like to explore this feature, please check out the details there.

Naive methods

To obtain the results of the naive methods, check out the commands in the shell script naive_imputation.sh.

❖ Downstream Tasks

We're cleaning up the code for the downstream tasks. Will release the scripts soon.