bluemk00/brits

Code of NIPS18 Paper: BRITS: Bidirectional Recurrent Imputation for Time Series

README

It is a pytorch implemention of paper "BRITS: Bidirectional Recurrent Imputation for Time Series, Wei Cao, Dong Wang, Jian Li, Hao Zhou, Lei Li Yitan Li. (NerIPS 2018)". The paper can be found here. http://papers.nips.cc/paper/7911-brits-bidirectional-recurrent-imputation-for-time-series

To train the BRIST model, first please unzip the PhysioNet data into raw folder, including the label file Outcomes-a.txt.

To run the model:

• make a empty folder named json, and run inpute_process.py.
• run different models:
  • e.g., RITS_I: python main.py --model rits_i --epochs 1000 --batch_size 64 --impute_weight 0.3 --label_weight 1.0 --hid_size 108
  • for most cases, using impute_weight=0.3 and label_weight=1.0 lead to a good performance. Also adjust hid_size to control the number of parameters